D2.2 Product Validation & Algorithm Selection Report (PVASR) Sea Ice Concentration Sea Ice Concentration

Transcription

1 Sea Ice Climate Change Initiative: Phase 2 D2.2 Product Validation & Algorithm Selection Report (PVASR) Sea Ice Concentration Sea Ice Concentration Doc Ref: SICCI-P2-PVASR (SIC) Version: 1.0 Date: 25th September

2 Change Record Issue Date Reason for Change Author September First Issue Rasmus Tonboe Authorship Role Name Signature Written by: Checked by: Approved by: Authorised by: Rasmus Tonboe (DMI) G. Timms (CGI) S. Sandven (NERSC) P. Lecomte (ESA) Distribution Organisation Names Contact Details ESA Pascal Lecomte NERSC CGI MET Norway Stein Sandven, Kirill Khvorostovsky Gary Timms, Sabrina Mbajon, Clive Farquhar Thomas Lavergne, Atle Sørensen DMI Rasmus Tonboe DTU Roberto Saldo, Henriette Skourup, Leif Toudal Pedersen FMI Marko Mäkynen, Eero Rinne University of Hamburg University of Bremen Stefan Kern, Lars Kaleschke, Xiangshan Tian-Kunze Georg Heygster MPI-M Dirk Notz Ifremer Fanny Ardhuin AWI Marcel Nicolaus, Stefan Hendricks, Thomas Hollands page 2 of 189

3 Table of Contents 1 Introduction Purpose and Scope Document Structure Document Status Applicable Documents Applicable Standards Reference Documents Acronyms and Abbreviations The algorithm intercomparison and selection procedure Composite algorithms names Evaluation of the algorithms over open water (SIC = 0%) Northern hemisphere open water Southern hemisphere open water SIC = 15% Northern hemisphere Southern hemisphere SIC = 100% Northern hemisphere Southern hemisphere FY vs MY SIC = 85% Northern hemisphere Southern hemisphere Thin ice Collection and visualisation of dataset Results for TB variability Melt ponds Comparison between SIC and open water fraction Simulated data The simulated sea ice concentration The simulated data The microwave emission models The snow and sea ice thermodynamic and mass model Simulation procedure The simulated data (Antarctic cases) The simulated data (Arctic cases) Open Water Weather filters Description of weather filters Results for selected algorithms Simulated data (15, 20, 25, 30% ice and cut-off) Atmospheric correction The effect of atmospheric correction Reduction in variance of TBs page 3 of 189

4 10.3 Reduction in variance in SIC (SIC=0) Reduction in variance in SIC (SIC=1) Examples for SIC=0 (Histograms) Integrated retrieval Discussion on atmospheric correction Algorithms and tie-points Algorithms Tie-points for normal processing Tie-points for atmospheric correction tests Near 90 GHz algorithms Instrument drift Northern Hemisphere Summary and conclusions Acknowledgements References page 4 of 189

5 List of Figures Figure 2-1: Standard deviations, SIC = 0%, Northern hemisphere, summer Figure 2-2: Bias, SIC = 0%, Northern hemisphere, summer Figure 2-3: Standard deviations, SIC = 0%, Northern hemisphere, winter Figure 2-4: Bias, SIC = 0%, Northern hemisphere, winter Figure 2-5: Standard deviations, SIC = 0%, Southern hemisphere, summer Figure 2-6: Bias, SIC = 0%, Southern hemisphere, summer Figure 2-7: Standard deviations, SIC = 0%, Southern hemisphere, winter Figure 2-8: Bias, SIC = 0%, Southern hemisphere, winter Figure 3-1: Standard deviations, SIC = 15%, Northern hemisphere, summer Figure 3-2: Bias, SIC = 15%, Northern hemisphere, summer Figure 3-3: Standard deviations, SIC = 15%, Northern hemisphere, winter Figure 3-4: Bias, SIC = 15%, Northern hemisphere, winter Figure 3-5: Standard deviations, SIC = 15%, Southern hemisphere, summer Figure 3-6: Bias, SIC = 15%, Southern hemisphere, summer Figure 3-7: Standard deviations, SIC = 15%, Southern hemisphere, winter Figure 3-8: Bias, SIC = 15%, Southern hemisphere, winter Figure 4-1: Standard deviations, SIC = 100%, Northern hemisphere, summer Figure 4-2: Bias, SIC = 100%, Northern hemisphere, summer Figure 4-3: Standard deviations, SIC = 100%, Northern hemisphere, winter Figure 4-4: Bias, SIC = 100%, Northern hemisphere, winter Figure 4-5: Standard deviations, SIC = 100%, Southern hemisphere, summer Figure 4-6: Bias, SIC = 100%, Southern hemisphere, summer Figure 4-7: Standard deviations, SIC = 100%, Southern hemisphere, winter Figure 4-8: Bias, SIC = 100%, Southern hemisphere, winter Figure 4-9: Standard deviation MYI vs FYI, SIC = 100%, Northern hemisphere, summer, SSM/I Figure 4-10: Standard deviation MYI vs FYI, SIC = 100%, Northern hemisphere, winter, SSM/I page 5 of 189

6 Figure 4-11: Standard deviation MYI vs FYI, SIC = 100%, Southern hemisphere, summer, SSM/I Figure 4-12: Standard deviation MYI vs FYI, SIC = 100%, Southern hemisphere, winter, SSM/I Figure 5-1: Standard deviations, SIC = 85%, Northern hemisphere, summer Figure 5-2: Bias, SIC = 85%, Northern hemisphere, summer Figure 5-3: Histograms of SIC for SSMI Northern hemisphere SIC85 data (Summer and Winter combined). The orange bar marks SIC=85% Figure 5-4: Standard deviations, SIC = 85%, Northern hemisphere, winter Figure 5-5: Bias, SIC = 85%, Northern hemisphere, winter Figure 5-6: Bias, SIC = 85%, Northern hemisphere, winter. Examples of correlation between SSMI SIC85 derived by different algorithms. Note the saturation of the NT2 algorithm and the ASI algorithm Figure 5-7: Standard deviations, SIC = 85%, Southern hemisphere, winter Figure 5-8: Bias, SIC = 85%, Southern hemisphere, winter Figure 5-9: NT2 atmosphere, SIC = 85%, Northern hemisphere, all year Figure 5-10: NT2 atmosphere, SIC = 100%, Northern hemisphere, all year Figure 6-1: Left: Example map of identified thin ice regions. Right: histogram of thicknesses distribution for the included data points Figure 6-2: Plot including a subset of algorithms Figure 6-3: Selected algorithms from the subset (a) Figure 6-4: Selected algorithms from the subset (b) Figure 6-5: Number of SMOS data points per thickness category (1 cm). Above 35 cm the number of data points is small Figure 6-6: Algorithm estimation of SIC for 100% 5cm thick ice Figure 6-7: Algorithm estimation of SIC for 100% 10 cm thick ice Figure 6-8: Algorithms estimation of SIC for 100% 15 cm thick ice Figure 6-9: Algorithms estimation of SIC for 100% 20 cm thick ice Figure 6-10: Algorithms estimation of SIC for 100% 25 cm thick ice Figure 6-11: Algorithm estimation of SIC for 100% ice at various thicknesses. The 35 cm data should be used with caution due to the limited number of datapoints at this thickness page 6 of 189

7 Figure 6-12: AMSR vertically polarized TBs as a function of SMOS ice thickness for our validation dataset Figure 6-13: AMSR horizontally polarized TBs as a function of SMOS ice thickness for our validation dataset Figure 6-14: AMSR Polarization ratio (PR) as a function of SMOS ice thickness Figure 6-15: NASA-Team multiyear ice concentration as a function of SMOS ice thickness Figure 7-1: Melt pond fraction throughout the dataset. X-axis goes from June 1 to August 31. July 1 is at 5310, August 1 at 8122, so very few points from August. 97 Figure 7-2: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) Figure 7-3: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) Figure 7-4: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) Figure 7-5: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) Figure 8-1: The simulated snow and ice profile in the Ross Sea (75 S, 200 E) Figure 8-2: The simulated Ross Sea ice concentration Figure 8-3: The 9 different ice concentration algorithms sensitivity to the snow ice interface temperature in the Ross Sea profile (see figure 8-1) Figure 8-4: Sensitivity of 9 sea ice concentration algorithms to cloud liquid water at the Ross Sea ice simulation Figure 8-5: The sensitivity of 9 algorithms to snow depth at the Ross Sea ice profile. 111 Figure 8-6: The sensitivity of 9 sea ice concentration algorithms to the snow temperature gradient in the Ross Sea ice profile Figure 8-7: The sensitivity of 9 sea ice concentration algorithms to the snow surface density in the Ross Sea ice profile Figure 8-8: The sensitivity of 9 sea ice concentration algorithms to atmospheric water vapor in the Ross Sea ice profile Figure 8-9: The sensitivity of 9 sea ice concentration algorithms to the average snow correlation length in the Ross Sea ice profile Figure 8-10: The sensitivity of the 9 sea ice concentration algorithms to the snow surface temperature in the Ross Sea ice profile Figure 8-11: Snow and ice profile in the Lincoln Sea multiyear ice Figure 8-12: The sea ice concentration for the Lincoln Sea multiyear ice profile page 7 of 189

8 Figure 8-13: The snow - ice interface temperature in the Lincoln Sea multiyear ice profile Figure 8-14: The sea ice concentration estimate from 9 different algorithms sensitivity to snow surface temperature in the Lincoln Sea multiyear ice profile Figure 8-15: The 9 sea ice concentration algorithms sensitivity to cloud liquid water in the Lincoln Sea multiyear ice profile Figure 8-16: The 9 sea ice concentration algorithms sensitivity to snow depth in the Lincoln Sea multiyear ice profile Figure 8-17: The 9 sea ice concentration algorithms sensitivity to snow temperature gradient in the Lincoln Sea multiyear ice profile Figure 8-18: The 9 sea ice concentration algorithms sensitivity to snow surface density in the Lincoln Sea multiyear ice Figure 8-19: The 9 sea ice concentration algorithms sensitivity to atmospheric water vapor in the Lincoln Sea multiyear ice profile Figure 8-20: The 9 sea ice concentration algorithms sensitivity to average snow correlation length in the Lincoln Sea multiyear ice profile Figure 8-21: The simulated sensitivity of the 9 sea ice concentration algorithms to cloud liquid water at 64 S 280 E over open water Figure 8-22: The simulated sensitivity of the 9 sea ice concentration algorithms to sea surface temperature at 64 S 280 E over open water Figure 8-23: The simulated sensitivity of the 9 sea ice concentration algorithms to atmospheric water vapor at 64 S 280 E over open water Figure 8-24: The simulated sensitivity of the 9 sea ice concentration algorithms to surface wind at 64 S 280 E over open water Figure 8-25: The simulated sensitivity of the 9 sea ice concentration algorithms to cloud liquid water at 70 N 0 E over open water Figure 8-26: The simulated sensitivity of the 9 sea ice concentration algorithms to sea surface temperature at 70 N 0 E over open water Figure 8-27: The simulated sensitivity of the 9 sea ice concentration algorithms to atmospheric water vapor at 70 N 0 E over open water Figure 8-28: The simulated sensitivity of the 9 sea ice concentration algorithms to wind at 70 N 0 E over open water Figure 9-1: Illustration of weather filter performance with AMSR-E data from 2008 Northern hemisphere. X-axis is GR1, and y-axix is GR2 from equation Figure 10-1: Example histogram of WaterVapour (SIC=0, 2008, AMSR) Figure 10-2: Example histogram of wind speed (SIC=0, 2008, AMSR) Figure 10-3: Example histogram of cloud liquid water (SIC=0, 2008, AMSR) page 8 of 189

9 Figure 10-4: Example histogram of T2m (SIC=0, 2008, AMSR) Figure 10-5: Example of scatter plot of T2m vs Ts. Marker size is proportional to WV. (ERA Interim, 2008, SIC=0, AMSR) Figure 10-6: TB36V before atmospheric correction Figure 10-7: TB36V after atmospheric correction Figure 10-8: TB36H before correction Figure 10-9: TB36H after correction Figure 10-10: TB89V before correction Figure 10-11: TB89V after correction Figure 10-12: Histogram SIC=0, Near 90 GHz Lin dyn. Note that the histogram has been truncated at -100 and Figure 10-13: Histogram after RTM SIC=0, Near 90 GHz Lin dyn Figure 10-14: Histogram SIC=0, Bootstrap P Figure 10-15: Histogram after RTM SIC=0, Bootstrap P Figure 10-16: Histogram SIC=0, Bristol Figure 10-17: Histogram after RTM SIC=0, Bristol Figure 10-18: Histogram SIC=0, Bootstrap F Figure 10-19: Histogram after RTM SIC=0, Bootstrap F Figure 10-20: Histogram SIC=0, OSISAF Figure 10-21: Histogram after RTM SIC=0, OSISAF Figure 10-22: Histogram of SIC=0 retrieval using, Bootstrap F Figure 10-23: Histogram integrated retrievalof SIC=0. Same X-axis (-10% to 10%) as figure but given in fractions rather than % Figure 11-1: Code for computing atmospheric corrected tie-points for NORSEX algorithm. The same atmospheric opacity are used for SMMR, SSMI and AMSR Figure 11-2: Relationship between P90 (n90v-n90h) and Sea Ice Concentration Figure 11-3: Relationship between P85 (85V-85H) and Sea Ice Concentration near SIC0 for the 4 near90 algorithms in the form they were tested Figure 11-4: Scatterplot of P90 before atmospheric correction (x-axis) vs P90 after correction (y-axis) Figure 12-1: SIC = 0%, Northern Hemisphere, average winter sea ice concentrations from selected algorithms. SMMR, SSM/I, AMSR page 9 of 189

10 Figure 12-2: SIC = 0%, Northern Hemisphere, average winter brightness temperatures for selected channels. SMMR, SSM/I, AMSR Figure 12-3: SIC = 100%, Northern Hemisphere, average winter sea ice concentrations from selected algorithms. SSM/I, AMSR Figure 12-4: SIC = 100%, Northern Hemisphere, average winter brightness temperatures for selected channels. SSM/I, AMSR Figure 12-5: SIC from 2008 SIC1 SSMI dataset before ATM correction. X-axis is sample number. Most points are in the Winter-spring. July 1 is October 1 is page 10 of 189

11 List of Tables Table 1-1: Applicable Documents Table 1-2: Applicable Standards Table 1-3: Reference Documents Table 1-4: Acronyms Table 1-5: Composite algorithms names Table 2-1: SIC = 0%, Northern Hemisphere, summer Table 2-2: SIC = 0%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm Table 2-3: SIC = 0%, Northern Hemisphere, winter Table 2-4: SIC = 0%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm Table 2-5: SIC = 0%, Southern Hemisphere, summer Table 2-6: SIC = 0%, Southern Hemisphere, summer. Average over all the instruments present for given algorithm Table 2-7: SIC = 0%, Southern Hemisphere, winter Table 2-8: SIC = 0%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm Table 3-1: SIC = 15%, Northern Hemisphere, summer Table 3-2: SIC = 15%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm Table 3-3: SIC = 15%, Northern Hemisphere, winter Table 3-4: SIC = 15%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm Table 3-5: SIC = 15%, Southern Hemisphere, summer Table 3-6: SIC = 15%, Southern Hemisphere, summer. Average over all the instruments present for given algorithm Table 3-7: SIC = 15%, Southern Hemisphere, winter Table 3-8: SIC = 15%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm Table 4-1: SIC = 100%, Northern Hemisphere, summer page 11 of 189

12 Table 4-2: SIC = 100%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm Table 4-3: SIC = 100%, Northern Hemisphere, winter Table 4-4: SIC = 100%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm Table 4-5: SIC = 100%, Southern Hemisphere, summer Table 4-6: SIC = 100%, Southern Hemisphere, summer. Average over all the instruments present for given algorithm Table 4-7: SIC = 100%, Southern Hemisphere, winter Table 4-8: SIC = 100%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm Table 4-9: SIC = 100%, Northern Hemisphere, summer, MYI vs FYI Table 4-10: SIC = 100%, Northern Hemisphere, winter, MYI vs FYI Table 4-11: SIC = 100%, Southern Hemisphere, summer, MYI vs FYI Table 4-12: SIC = 100%, Southern Hemisphere, winter, MYI vs FYI Table 5-1: SIC = 85%, Northern Hemisphere, summer Table 5-2: SIC = 85%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm Table 5-3: SIC = 85%, Northern Hemisphere, winter Table 5-4: SIC = 85%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm Table 5-5: SIC = 85%, Southern Hemisphere, winter Table 5-6: SIC = 85%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm Table 7-1: Average values and stdevs for MODIS derived SIC and MPF and for AMSR TBs Table 7-2: Calculated SIC and their stdevs for the tested algorithms. Upper row in red are MODIS derived reference values. The bottom row is multi-year ice concentration as derived with the Nasa Team algorithm Table 8-1: Categorization of the 9 selected algorithms. The polarization algorithms are using the polarization difference or ratio. The gradient algorithms are using the spectral gradient e.g. at Tb19v and Tb37v. The hybrid refers to a combination of polarization and gradient. Here low frequency is 6 GHz and high frequency is near 90 GHz. ESMR is the single channel (Tb19h) radiometer on NIMBUS Table 8-2: The Ross Sea correlation matrix. The snow surface density: Dens, The average snow correlation length: appc, The snow surface temperature: Ti, The page 12 of 189

13 snow ice interface temperature: ist, The snow depth: St, the snow temperature gradient: snowg, The atmospheric water vapor: Vapor, The cloud liquid water: Liquid Table 8-3: mean ice concentration, the standard deviation with or without atmosphere and the mean atmosphere/ no atmosphere ice concentration ratio for the 9 algorithms in the Ross Sea ice profile Table 8-4: The partial correlation of each of the 6 parameters (the snow ice interface temperature, the snow surface temperature, the snow depth, the snow surface density, atmospheric water vapor, cloud liquid water) and the ice concentration with the effects of the other physical parameters (snow surface density, the average snow correlation length, snow surface temperature, the snow ice interface temperature, snow depth, snow temperature gradient, atmospheric water vapor, cloud liquid water, excluding one of the 6) removed Table 8-5: Shows the mean ice concentration, the standard deviation with or without atmosphere and the mean atmosphere/ no atmosphere ice concentration ratio for the 9 algorithms in the Lincoln Sea ice profile. The no atmosphere case is the surface emission only. The atmosphere included is the atmospheric emission absorption and reflection computed with a modified Wentz model in addition to the surface emission. This includes oxygen absorption Table 8-6: Partial correlations: The partial correlation of each of the 6 parameters (snow ice interface temperature, surface temperature, snow depth, snow surface density, atmospheric water vapor, cloud liquid water) and the ice concentration with the effects of the other physical parameters (snow surface density, snow correlation length, snow surface temperature, snow ice interface temperature, snow depth, snow temperature gradient, atmospheric water vapor, cloud liquid water, excluding one of the 6) removed Table 8-7: The 64 S 280 E sample MEAN and STDEV Table 8-8: The Bellinghausen Sea open water correlation matrix, r Table 8-9: The partial correlation, r, of each of the 4 input parameters to the Wentz model (The surface wind speed, atmospheric water vapor, the cloud liquid water, and the sea surface temperature) and the ice concentration with the effects of the other physical parameters removed (the one in question excluded) Table 8-10: The partial correlation, r, of each of the 4 input parameters to the Wentz model (The surface wind speed, atmospheric water vapor, the cloud liquid water, and the sea surface temperature) and the ice concentration with the effects of the other physical parameters removed (the one in question excluded) Table 8-11: Correlation matrix at 70 N 0 E, r, with the four input parameters to the Wentz model and the ice concentration from selected algorithms Table 8-12: The mean, MEAN, and standard deviation, STDEV, of the simulated data at 70 N 0 E Table 9-1: SIC = 15%, summer. Standard deviations of concentrations with and without weather filters Table 9-2: SIC = 15%, summer. Average concentrations with and without weather filters page 13 of 189

14 Table 9-3: SIC = 15%, winter. Standard deviations of concentrations with and without weather filters Table 9-4: SIC = 15%, winter. Average concentrations with and without weather filters145 Table 10-1: Reduction of TB variance by correction for various atmospheric terms from ERA Interim co-located data. A total of 6621 datapoints from both hemispheres and all seasons were used in the assessment. Only SIC= Table 10-2: Reduction of TB variance by correction for various Teff terms from ERA Interim co-located data. A total of 1220 data-points from both hemispheres and all seasons were used in the assessment Table 10-3: Average SIC and standard deviation of SIC before (orange) and after (green) atmospheric correction of TBs. Open water cases. Results are computed with theoriginal Tie-Points without atm correction for all cases which explains the substantial biases of many algorithms Table 10-4: SIC=0 standard deviations before and after atmospheric correction. Atmospheric corrected TPs used for atmospheric corrected data. AMSR, Northern hemisphere Table 10-5: SIC=0 average and standard deviations before (red) and after (green) atmospheric correction of TBs. Atmospheric corrected Tie-points used for atmospheric corrected analysis. The table summarises results from northern and southern hemispheres and for summer and winter. Near90 results are suspected to be erroneous Table 10-6: Average SIC and standard deviation of SIC before (orange) and after (green) atmospheric correction of TBs. SIC1 cases Table 10-7: Results of atmospheric correction for SIC=1 and split in FY and MY. MY is 15% datapoints with lowest 36H and FY is 15% with highest 36H. Northern hemisphere only Table 10-8: Average and standard deviation of retrieval of SIC and a number of other parameters using an integrated retrieval method Table 11-1: Tie-points for Northern Hemisphere used with non-atmospheric corrected TBs, NX columns are tie-points for the NORSEX algorithm Table 11-2: Tie-points for Southern Hemisphere used with non-atmospheric corrected TBs, NX columns are tie-points for the NORSEX algorithm Table 11-3: Tie-points for Northern Hemisphere used with non-atmospheric corrected TBs, NX columns are tie-points for the NORSEX algorithm. SMMR tie-points for FY and MY ice are set to AMSR tie-points since we do not have RRDP data for SMMR from 100% ice Table 11-4: Tie-points for Southern Hemisphere used with non-atmospheric corrected TBs, NX columns are tie-points for the NORSEX algorithm. SMMR tie-points for FY and MY ice are set to AMSR tie-points since we do not have RRDP data for SMMR from 100% ice Table 11-5: Tie-points for Northern Hemisphere used with non-atmospheric corrected TBs, NX columns are tie-points for the NORSEX algorithm. SMMR tie-points for FY page 14 of 189

15 and MY ice are set to AMSR tie-points since we do not have RRDP data for SMMR from 100% ice Table 11-6: Tie-points for Southern Hemisphere used with non-atmospheric corrected TBs, NX columns are tie-points for the NORSEX algorithm. SMMR tie-points for FY and MY ice are set to AMSR tie-points since we do not have RRDP data for SMMR from 100% ice Table 11-7: Tie-points SIC=0 used for testing algorithm performance with atmospheric corrected TBs. Same TPs are used for N and S page 15 of 189

16 1 Introduction 1.1 Purpose and Scope This document, the product validation and algorithm selection report or the PVASR, is describing the analysis of the different algorithms using the round robin test data. It is describing the criteria s for selection and the results of the evaluation in the algorithm omnium. The algorithm which is then selected is described in the algorithm theoretical basis document: the ATBD. The ATBD also includes a more detailed description of all the algorithms tested, including the python computer code. 1.2 Document Structure After this introduction and the list of references, the document is divided into a number of chapters dealing with each part of the algorithm validation Algorithm evaluation Algorithms will be evaluated independently for each identified source of error/uncertainty. Eventually we intend to fine-tune the selected algorithm(s) with dynamic tie-points in order to minimize the effect of sensor drift and inter sensor differences. In their published form the algorithms were tuned to a specific version of the source microwave radiometer data. We will be using different sources of data demanding a new calibration (tie-point tuning) of the algorithms anyway. As a consequence, in algorithm evaluation we will be less interested in eventual biases relative to the correct ice concentrations, and more interested in RMS errors around the expected values. Algorithms will be tested against the following sources of errors: Sensitivity to atmosphere Open water dataset selected in regional seas around the Arctic and Antarctic. Simulated data using the Wentz forward models Met/ocean data screening Sensitivity to emissivity variations 100% ice dataset Simulated data Snow/ice/atmosphere data screening Summer winter sensitivity differences In addition the algorithms will be compared and evaluated concerning the following characteristics: page 16 of 189

17 Summer performance (melt ponds etc.) SMMR vs. SSMI vs AMSR performance Thin ice performance Potential obtainable spatial resolution vs. uncertainty Potential time period e.g. 10, 20 or 30 years Hemisphere differences (North and South) Summer winter sensitivity differences 1.3 Document Status This is a first issue release to ESA as part of the project s contractual deliverable set for Phase Applicable Documents The following table lists the Applicable Documents that have a direct impact on the contents of this document. Acronym Title Reference Issue AD-1 Sea Ice ECV Project Management Plan ESA-CCI_SICCI_PMP_D6.1_v Table 1-1: Applicable Documents 1.5 Applicable Standards Acronym Title Reference Issue Table 1-2: Applicable Standards 1.6 Reference Documents Acronym Title Reference Issue URD-1 Sea Ice ECV User Requirement Survey ATBD D2.1 SIC Algorithm Theoretical Basis Document (SIC ATBD) 1.0 SICCI2-SIC-ATBD page 17 of 189

18 Acronym Title Reference Issue CECR_01 PVP Sea Ice ECV Comprehensive Error Characterisation Report Sea Ice ECV Product Validation Protocol Table 1-3: Reference Documents Acronyms and Abbreviations Acronym AMSR-E AO ASCII ASI_NWFRAS CM-SAF DMSP DWD ECV Envisat ESA EUMETSAT FCDR FOC FOV FTP GB GCOM H H+V MB MODIS n.a. NetCDF NSIDC OIB OSI-SAF PI PMW POES PRF RADAR SAR Meaning Advanced Microwave Scanning Radiometer aboard EOS Announcement of Opportunity American Standard Code for Information Interchange Airborne Synthetic Aperture and Interferometric Radar Altimeter System Climate Monitoring Satellite Application Facility Defence Meteorological Satellite Program Deutscher Wetterdienst Essential Climate Variable Environmental Satellite European Space Agency European Organisation for the Exploitation of Meteorological Satellites Fundamental Climate Data Record Free of Charge Field-of-View File Transfer Protocol GigaByte Global Change Observation Mission Horizontal polarization Horizontal and vertical polarization MegaByte Moderate Resolution Imaging Spectroradiometer Not applicable Network Common Data Format National Snow and Ice Data Center Operation Ice Bridge Ocean and Sea Ice Satellite Application Facility Principal Investigator Passive Microwave Polar Operational Environmental Satellite Pulse Repetition Frequency Radio Detection and Ranging Synthetic Aperture Radar page 18 of 189

19 Acronym SIC SIRAL SIT SMMR SSM/I SSM/IS TB t.b.d. TM ULS URL V Meaning Sea Ice Concentration SAR/Interferometric Radar Altimeter Sea Ice Thickness Satellite Multichannel Microwave Radiometer Special Sensor Microwave / Imager Special Sensor Microwave / Imager+Sounder TeraByte To be determined Thematic Mapper Upward Looking Sonar Uniform Resource Locator Vertical polarization Table 1-4: Acronyms 1.8 The algorithm intercomparison and selection procedure Intersensor calibration Sea ice concentration observations will be intercalibrated among observations from the Scanning Multichannel Microwave Radiometer (SMMR) ( ), various Special Sensor Microwave/Imager (SSMI) instruments (1987-present) and the Advanced Microwave Scanning Radiometer (AMSR) ( ) (see DARD: IDs 1.01 to 1.03). Tie-points are typical signatures of 100% ice and open water which are used in the ice concentration algorithms as a reference. The tie-points are derived by selecting brightness temperatures from regions of known open water and 100% ice. Usually these tie-points are static in time and space, but they can be adjusted to follow the seasonally changing signatures of ice and open water (see e.g. Kern and Heygster, 2001) as it is currently done, for instance, in the CDR and operational OSISAF ice concentration processing (Tonboe et al., 2016). Static tie-points are prone to be affected by sensor drift, inter sensor calibration differences and climatic trends in surface and atmospheric emission. The data must therefore be carefully calibrated before computing the ice concentrations. Here we will use dynamic tiepoints, a method that minimizes these unwanted effects, with or without prior calibration of the passive microwave data. Such dynamic tie points also facilitate the combination of data from different sensors Validation and error bars The complete transfer of uncertainty will be investigated in this project, from L1b swath-based brightness temperatures to daily gridded composite. The sensitivity from various sea ice concentration algorithms will be evaluated, with respect to various factors such as atmospheric noise, surface emissivity uncertainty, noise in the observed brightness temperatures, etc. A research effort will be conducted on the gridding and projection algorithms, and will result in a quantitative transfer of uncertainty associated to this step, depending on the FoV size and shape of the instruments. Error bars will be assessed both theoretically and empirically as page 19 of 189

20 their consistency is checked against results from an extensive validation exercise Discrepancies at high ice concentrations (~100%) During winter, in consolidated ice, well within the ice edge, the ice concentration is very near 100% [Andersen et al., 2007]. This has been established using high resolution SAR data, ship observations and by comparing the estimates from different ice concentration algorithms. The fluxes between the ocean/ice and atmosphere are sensitive to small variations in these ranges of sea ice concentration and thereby these discrepancies are of large importance for coupled climate models. The apparent fluctuations in the derived ice concentration in the near 100% ice regime are primarily attributed to snow/ice surface emissivity variability around the tie-point signature and only secondarily to actual ice concentration fluctuations [Kwok, 2002]. In the marginal ice zone the atmospheric extinction may be significant. The fluctuations due to atmospheric and surface emission are systematic. In fact, different algorithms with different sensitivity to atmospheric extinction and surface emission compute quite different trends in sea ice area and extent on seasonal and decadal time scales [Andersen et al., 2007]. This means that not only does the sea ice area have a climatic trend, but the atmospheric and surface constituents affecting the microwave emission are also changing. For example, different wind patterns, water vapour and liquid water concentrations in the atmosphere, snow depth, fraction of perennial ice etc. The present project will pay particular attention to retrieval close to 100% sea ice concentration by 1) including 100% ice covered areas in the test/validation datasets, 2) detecting these 100% ice covered situations by SAR and high-resolution optical data, 3) supplementing the detection of these situations using deformation information processed from a SAR sea ice motion dataset, and 4) studying the sensitivity of PMW algorithms to variations surface emissivity (ice-type, snow, etc.) using model simulations Discrepancies in summer conditions: melt ponds, wet snow and ice. Reflectance from MODIS channels 1, 3 and 4 are used to derive the meltpond cover fraction and a summer-time ice concentration estimate. The melt pond cover fraction is determined using a classification which follows a mixed-pixel approach. It is assumed that the reflectance measured over each MODIS 500 m grid cell comprises contributions from three surface types: melt ponds, open water, sea ice/snow [Roesel et al., 2012a]. By using known reflectance values [e.g. Tschudi et al., 2008] a neural network is built, trained, and applied [Roesel et al., 2012a]. The resulting surface type class distributions are saved and made available as 12.5 km grid resolution product (DARD ID: 2.08); melt pond cover fractions are adjusted to the fraction of the sea ice cover per pixel, i.e. the melt pond cover fraction is not given relative to the area of the grid cell but given relative to the sea ice area of that grid cell Weather filters Many algorithms deploy weather filters to avoid detecting ice in open water. The brightness temperature noise is generated by wind roughening of the page 20 of 189

21 ocean surface, by water (vapour or liquid) in the atmosphere or by precipitation. We will as far as possible the algorithms and the weather filters separately and with and without these weather filters, and we may decide to use a set of algorithms with weather filters published with non-selected algorithms. 1.9 Composite algorithms names The OSI-SAF sea ice concentration data set is based on a combination of different algorithm. In addition to original algorithms we will also test performance of this combination as well as other composite algorithms. The short names of these composite algorithms will be used in the figures of the document for convenience. Short name Combo1 Combo2 Combo3 Combo4 Description (NASA Team + Bootstrap_F)/2 (NASA Team + Bootstrap_F + Near90GHz lin dyn)/3 (P37 + Near 90GHz lin dyn)/2 (P37 + Near 90GHz lin dyn + Bootstrap_F)/3 Combo5 (Bootstrap_F +( Bootstrap_F 2 )* Near 90GHz lin dyn)/(1+ Bootstrap_F 2 ) Combo6 (Bootstrap_F +( Bootstrap_F 3 )* Near 90GHz lin dyn)/(1+ Bootstrap_F 3 ) Combo7 Combo8 (Bootstrap_F + Near 90GHz lin dyn)/2 (Bootstrap_F + Bootstrap_F * Near 90GHz lin dyn)/(1+ Bootstrap_F) Table 1-5: Composite algorithms names page 21 of 189

22 2 Evaluation of the algorithms over open water (SIC = 0%) In principle every algorithm should be evaluated over open water, at intermediate concentrations and at near 100% ice cover. In practise it is very difficult to find reference data at intermediate concentrations especially for large areas covering entire satellite footprints (70km) and covering all seasons and ice types. These issues compromise the meaningfulness of the evaluation at intermediate concentrations. Therefore the evaluation is carried out only for the SIC = 0% and SIC = 100% cases. These are the cases where we can get the most high quality reference data. For some algorithms it was necessary to use SIC = 15% and SIC = 85% instead of 0% and 100%. The reason for this will be explained in section 2.1. The input data to the algorithms are sets of brightness temperatures corresponding to sea ice concentrations equal 0%. Ice charts were used in order to identify such zones. The data are provided from the instruments: SMMR, SSM/I and AMSR for Northern and Southern hemispheres. The Northern hemisphere summer is defined as months from June to September, and winter as October May. The Southern hemisphere summer is defined as months from December to March, and winter is April through November. In this and the following 3 sections (SIC = 15%, SIC = 100%, SIC = 85%) the algorithms are run without weather filters applied. The weather filters study is presented in the section 9. The algorithms are characterized by standard deviation and bias from the validation dataset. The bias is defined as difference from the given concentration value of the validation dataset (SIC = 0%, 15%, 85%, 100%). The ASI algorithm is called ASI_NWF to mark the version of the algorithm used here (NWF stands for no weather filter ). 2.1 Northern hemisphere open water The SMMR, SSM/I and AMSR data were collected at open water sites on the Northern Hemisphere at a safe distance but not too far from the ice edge. There are different sites for summer and winter to follow the seasonal variation of the ice edge. The sites are in the Atlantic and in the Pacific. We know that the ice concentration over open water is always zero and all variability and standard deviation in the data produced by the different algorithms is a quantification of the sensitivity to noise. The noise may be from the radiometer instrument, from wind induced surface roughness, from surface and atmospheric temperature variability and from atmospheric water vapour and clouds. Normally a bias in the ice concentration over open water is a question of adjusting the tie-points. We use a standard set of tie-points described in section 11 and a bias is not necessarily an indication of poor algorithm performance. However, some algorithms have a special way of estimating tie-points and a non-linear way of dealing with ice concentrations near 0% page 22 of 189

23 (and at 100%). This makes it impossible to compare these algorithms directly to other algorithms because the standard deviation is affected by the treatment at 0% and 100% reference points. Therefore, we have artificially produced reference datasets at 15% and at 85% for evaluating these special algorithms at these points instead. A potential bias at these intermediate reference points may indicate a bias at intermediate concentrations in general Summer (June- September) Table 2-1 is showing the standard deviation and average ice concentration at the open water reference sites for all 3 sensors separately and all algorithms during summer (June-September). The gaps in the table are when a given algorithm is using near 90GHz channels which is not on SMMR and 6GHz which is not on SSM/I. In the tables throughout the document we will use the full descriptions of the composite algorithms to give better insight, however, in the figures the short names listed in the Table 1-5 will be used for practical reasons. The SMMR has in general the lowest standard deviation. This may be due to the 18GHz channel being further away from the water absorption line at near 22GHz than the other two sensors. The average and standard deviation for all the three sensors are combined in Table 2-2 for clarity. All algorithms using near 90GHz channels have very high standard deviations. The Bootstrap F and other similar channel combinations has the lowest standard deviation. Algorithm Standard deviation Average AMSR SSM/I SMMR AMSR SSM/I SMMR Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ page 23 of 189

24 Algorithm Standard deviation Average AMSR SSM/I SMMR AMSR SSM/I SMMR (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 2-1: SIC = 0%, Northern Hemisphere, summer Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) page 24 of 189

25 Algorithm Standard deviation Average (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 2-2: SIC = 0%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm The bar-chart in Figure 2-1 and Figure 2-2 is showing Table 2-1 graphically. The algorithms using the 19v 37v channel combinations have the lowest standard deviations for all three sensors. Figure 2-1: Standard deviations, SIC = 0%, Northern hemisphere, summer page 25 of 189

26 Figure 2-2: Bias, SIC = 0%, Northern hemisphere, summer Winter (October May) Table 2-3 and Table 2-4 (and Figure 2-3 and Figure 2-4) are showing the average and standard deviation of the ice concentration over the open water sites in winter (October May) on the northern hemisphere. The overall pattern is very similar to the summer situation. The differences are reflecting the different algorithms sensitivity to wind and atmospheric humidity and other seasonally changing quantities [Andersen et al. 2006]. These different sensitivities are identified in section 8 using a simulated dataset. Some of the quantities such as water vapour have climatological trends and a small difference between the summer and winter dataset is an asset for an algorithm. For example the Bootstrap F has a low summer winter difference (<0.3) in standard deviation, the Bristol and the NASA Team have a moderate difference ( ) and the Bootstrap-P and the ASI a high difference ( ). The algorithms with a high summer winter difference are potentially sensitive to noise with a climatological trend. Algorithm Standard deviation Average AMSR SSM/I SMMR AMSR SSM/I SMMR Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P page 26 of 189

27 Algorithm Standard deviation Average AMSR SSM/I SMMR AMSR SSM/I SMMR P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 2-3: SIC = 0%, Northern Hemisphere, winter Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR page 27 of 189

28 Algorithm Standard deviation Average NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 2-4: SIC = 0%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm page 28 of 189

29 Figure 2-3: Standard deviations, SIC = 0%, Northern hemisphere, winter Figure 2-4: Bias, SIC = 0%, Northern hemisphere, winter page 29 of 189

30 2.2 Southern hemisphere open water The standard deviation of the open water sea ice concentration is in general lower on the southern hemisphere than on the northern hemisphere with a few exceptions (e.g. the Bootstrap P). Also the summer winter difference in standard deviation is smaller on the southern hemisphere than on the northern hemisphere. The NASA Team has among the lowest summer winter difference of 0.04 and the ASI has among the highest of Summer (December March) Table 2-5, Table 2-6, Figure 2-5 and Figure 2-6 show the average and standard deviation of the ice concentration over open water for the sites on the southern hemisphere during austral summer. Algorithm Standard deviation Average AMSR SSM/I SMMR AMSR SSM/I SMMR Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) page 30 of 189

31 Algorithm Standard deviation Average AMSR SSM/I SMMR AMSR SSM/I SMMR OSISAF OSISAF OSISAF Nr. of points Table 2-5: SIC = 0%, Southern Hemisphere, summer Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points page 31 of 189

32 Table 2-6: SIC = 0%, Southern Hemisphere, summer. Average over all the instruments present for given algorithm Figure 2-5: Standard deviations, SIC = 0%, Southern hemisphere, summer Figure 2-6: Bias, SIC = 0%, Southern hemisphere, summer page 32 of 189

33 2.2.2 Winter (April November) Table 2-7, Table 2-8, Figure 2-7 and Figure 2-8 are showing the average and the standard deviation of the ice concentration over open water during austral winter. The Bootstrap F has the lowest standard deviation among the algorithms using channels which are present on all three instruments: AMSR, SSM/I and SMMR (19 37GHz). Algorithm Standard deviation Average AMSR SSM/I SMMR AMSR SSM/I SMMR Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points page 33 of 189

34 Table 2-7: SIC = 0%, Southern Hemisphere, winter Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 2-8: SIC = 0%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm page 34 of 189

35 Figure 2-7: Standard deviations, SIC = 0%, Southern hemisphere, winter Figure 2-8: Bias, SIC = 0%, Southern hemisphere, winter page 35 of 189

36 3 SIC = 15% Due to nonlinearity and cut-offs in some of the algorithms it is difficult to compare these to other algorithms at 0% and 100%. Therefore, artificial files that correspond to SIC = 15% and 85% were generated to evaluate the algorithms in a range where cut-offs are not affecting the standard deviation. SMMR is not presented in sections with nonzero ice concentration because of the absence of the validation data from this case: there were no ENVISAT or any other SAR data at the time of SMMR, and these were needed to identify zones of convergence and thereby the regions where the ice concentration is 100%. In terms of centre frequency AMSR is the instrument which is closest to SMMR. The 15% artificial dataset was generated as follows. Sea ice concentration of 15% is typically a mixture of first-year (FY) ice and water, therefore: SIC * SIC0(t) 0.15* SIC100(FY ), (3-1) where SIC0 (open water) is multiplied by 0.85 (85% water) and is varying with time, while added 15% of ice signature is an average value. The variability is that of open water and here biases are important because they potentially indicate offsets at intermediate concentrations or cut-offs working as weather filters. 3.1 Northern hemisphere Summer (June September) The average and the standard deviation of the ice concentration at a nominal concentration of 15% is shown in Table 3-1 and Table 3-2 and illustrated in Figure 3-1 and Figure 3-2 for the AMSR and SSM/I sensors. The ECICE and the NASA Team 2 have been included in the list of algorithms. Many of the algorithms such as the Bootstrap P, the Bristol the NASA Team and the Bootstrap F (and related algorithms) have average ice concentrations close to 15% as expected. All of the algorithms using near 90GHz channels have average ice concentrations greater than 30% including the ECICE and the NASA Team 2. The standard deviation of the ice concentration is low for the Bootstrap F family of algorithms, intermediate for the NASA Team 2, the ECICE and Bootstrap p and high for the algorithms using only 90GHz channels. Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P page 36 of 189

37 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 3-1: SIC = 15%, Northern Hemisphere, summer Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol page 37 of 189

38 Algorithm Standard deviation Average PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 3-2: SIC = 15%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm page 38 of 189

39 Figure 3-1: Standard deviations, SIC = 15%, Northern hemisphere, summer Figure 3-2: Bias, SIC = 15%, Northern hemisphere, summer page 39 of 189

40 3.1.2 Winter (October May) Table 3-3, Table 3-4, Figure 3-3 and Figure 3-4 show the average and standard deviation of the ice concentration for SIC = 15%, winter in the Northern Hemisphere. The overall pattern is similar to the summer situation described above. The summer winter differences in standard deviation of the ice concentration for ECICE and NASA Team 2 is around This is greater than e.g. NASA Team and Bootstrap F at about 0.4 and less than e.g. Bootstrap P at The summer winter difference is an indicator of sensitivity to noise that may have a climatological trend. Clearly atmospheric parameters such as water vapour is different in the summer winter situation and water vapour may have a climatic trend. Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF page 40 of 189

41 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I NASA Team ECICE Nr. of points Table 3-3: SIC = 15%, Northern Hemisphere, winter Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE page 41 of 189

42 Algorithm Standard deviation Average Nr. of points Table 3-4: SIC = 15%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm Figure 3-3: Standard deviations, SIC = 15%, Northern hemisphere, winter page 42 of 189

43 Figure 3-4: Bias, SIC = 15%, Northern hemisphere, winter 3.2 Southern hemisphere The standard deviation of the ice concentration at a nominal ice concentration of 15% (shown together with average values in Table 3-5, Table 3-6, Figure 3-5 and Figure 3-6) is in general lower for the southern hemisphere than the northern hemisphere while the overall pattern is similar. Also the summer and winter differences are smaller on the southern hemisphere than on the northern hemisphere. In terms of summer winter differences in standard deviation of the ice concentration NASA Team 2 is in the same category with Bootstrap F and NASA Team with very small differences (<0.1) while the ECICE with a 0.29 difference is closer to e.g. Bootstrap P at Summer (December March) Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P page 43 of 189

44 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 3-5: SIC = 15%, Southern Hemisphere, summer Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol page 44 of 189

45 Algorithm Standard deviation Average PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 3-6: SIC = 15%, Southern Hemisphere, summer. Average over all the instruments present for given algorithm page 45 of 189

46 Figure 3-5: Standard deviations, SIC = 15%, Southern hemisphere, summer Figure 3-6: Bias, SIC = 15%, Southern hemisphere, summer page 46 of 189

47 3.2.2 Winter (April November) Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 3-7: SIC = 15%, Southern Hemisphere, winter Algorithm Standard deviation Average page 47 of 189

48 Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 3-8: SIC = 15%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm page 48 of 189

49 Figure 3-7: Standard deviations, SIC = 15%, Southern hemisphere, winter Figure 3-8: Bias, SIC = 15%, Southern hemisphere, winter page 49 of 189

50 The Bias figures above show that most algorithms are practically unbiased with the notable exception of the ones with fixed coefficients (ECICE, NT2, ASI, P90 and PR) which could not be tuned to the data correctly. The Standard deviations of similar algorithms indicate that the stdev results for these biased algorithms are not severely impacted by these biases and thus can be considered to be representative for algorithm performance at this 15% ice test. Note also that algorithms such as NT2 which is designed to find and correct for an approximate atmosphere may have difficulties with the mixed atmosphere we have introduced in the mix of OW and FY data to build this artificial 15% ice dataset. However, the results of integrated retrieval presented in section 10.6 indicates that adaptation to the mixed atmosphere is possible. page 50 of 189

51 4 SIC = 100% For the 100% validation dataset identification of 100% ice areas was done from convergence in ENVISAT ASAR derived sea ice drift fields. The necessary sea ice drift fields are available from the PolarView and MyOcean projects. We know that the ice concentration is at 100% and the standard deviation of the ice concentration is an indication of sensitivity to noise. The geophysical noise sources over ice are atmospheric water vapour and cloud liquid water, surface emissivity and temperature variability of the snow and sea ice. The basic assumption for the convergence method to provide 100% ice is that after 24 hours of convergence in an area the open water areas (leads) that were there at the beginning have either closed (by convergence) or refrozen (due to the cold temps). During Summer, melt ponds jeopardizes this assumption and variability since they neither close nor refrezes during this period. Our 100% ice dataset is therefore only valid for accurate tests during winter (~October-May) and results from Summer should be used with great caution. 4.1 Northern hemisphere Standard deviations and average values at 100% sea ice concentration in the Northern Hemisphere are shown in Table 4-1, Table 4-2, Figure 4-1 and Figure 4-2 for summer and in Table 4-3, Table 4-4, Figure 4-3 and Figure 4-4 for winter. Compared to the open water case differences in the standard deviation of ice concentration are smaller and around 9% to 12% for most algorithms in summer and 4% to 7% in winter on the northern hemisphere. The near 90 GHz algorithms have a low standard deviation of the ice concentration in both summer and winter. There are several algorithms having high or intermediate standard deviations. The polarisation algorithms have low average ice concentration in summer while most of the algorithms have an average ice concentration near 100% in winter Summer (June September) Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team page 51 of 189

52 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-1: SIC = 100%, Northern Hemisphere, summer Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal page 52 of 189

53 Algorithm Standard deviation Average UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-2: SIC = 100%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm Figure 4-1: Standard deviations, SIC = 100%, Northern hemisphere, summer page 53 of 189

54 Figure 4-2: Bias, SIC = 100%, Northern hemisphere, summer Most algorithms underestimate SIC in summer. This is ascribed to the fact that microwave penetration into water is very small and that melt-ponds therefore are seen as water. Notable exceptions are the 19V/37V algorithms (NORSEX, Bootstrap_f, CalVal and UMass_AES and TUD) which through the combined effect of overestimation of the SIC of wet snow and seeing melt ponds as water produces fairly small biases. Field studies have shown that there is always snow on sea ice even during late summer. The apparent small bias for ASI and P90 are due to the saturation effect described in section Winter (October May) Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P page 54 of 189

55 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-3: SIC = 100%, Northern Hemisphere, winter Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team page 55 of 189

56 Algorithm Standard deviation Average NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-4: SIC = 100%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm page 56 of 189

57 Figure 4-3: Standard deviations, SIC = 100%, Northern hemisphere, winter Figure 4-4: Bias, SIC = 100%, Northern hemisphere, winter Most algorithms perform well at high concentrations. Standard deviations are in the order of 4-7%. Bristol/OSISAF performs somewhat better than most others, but also some of the Combo algorithms that combine polarisation and spectral gradients produce standard deviations in the 4-5% range. The spectral gradient only algorithms (CF, NORSEX,...) are somewhat worse. Again the N90 nonlinear algorithms benefit from their fading out at high concentrations. Biases are generally small for all algorithms. The small remaining biases seen are due to the test period being different from the period for finding tie-points (2008) 4.2 Southern hemisphere Similarly to the northern hemisphere the algorithms only using near 90GHz channels have a low standard deviation of the ice concentration both summer and winter. Contrary to the northern hemisphere there are certain polarisation algorithms e.g. Bootstrap P that stand out with a high standard deviation of the ice concentration in both summer and winter. Also contrary to the northern hemisphere is the positive bias in the average ice concentration for all algorithms during summer. page 57 of 189

58 4.2.1 Summer (December March) Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-5: SIC = 100%, Southern Hemisphere, summer Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz page 58 of 189

59 Algorithm Standard deviation Average ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-6: SIC = 100%, Southern Hemisphere, summer. Average over all the instruments present for given algorithm page 59 of 189

60 Figure 4-5: Standard deviations, SIC = 100%, Southern hemisphere, summer Figure 4-6: Bias, SIC = 100%, Southern hemisphere, summer page 60 of 189

61 Most algorithms show positive biases during austral summer due to the variability in the presence of wet snow / bare ice during the summer season. Standard deviations are less reliable measure since real ice concentration during Summer may not be 100% in our test dataset where 1-day convergence is used to find SIC=100% so the dataset may include some unknown but real variability in SIC. The basic assumption for the convergence method to provide 100% ice is that after 24 hours of convergence in an area the open water areas (leads) that were there at the beginning have either closed (by convergence) or refrozen (due to the cold temps). During Summer, melt ponds jeopardizes this assumption and variability since they neither close nor refrezes during this period. Our 100% ice dataset is therefore only valid for accurate tests during winter (~October-May) and results from Summer should be used with great caution Winter (April November) Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) page 61 of 189

62 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I OSISAF OSISAF OSISAF Nr. of points Table 4-7: SIC = 100%, Southern Hemisphere, winter Winter results for the Southern hemisphere are quite similar to Northern hemisphere. Bristol/OSISAF again showing the smallest stdev. The Combo algorithms are not as good here as in the Northern Hemisphere, and the 19/37V algorithms perform somewhat better than in the north. Results for AMSR are slightly worse in general than results for SSMI. Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF page 62 of 189

63 Algorithm Standard deviation Average OSISAF OSISAF Nr. of points Table 4-8: SIC = 100%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm Figure 4-7: Standard deviations, SIC = 100%, Southern hemisphere, winter page 63 of 189

64 Figure 4-8: Bias, SIC = 100%, Southern hemisphere, winter 4.3 FY vs MY In order to assess if the performance of the SIC algorithms is different in areas of FY and areas of MY ice, we split the dataset in 3 parts according to the prevalent ice type. Multi-year ice concentration (CMY) was defined from the NASA Team algorithm. Group 1 (represented mostly by MY ice): all the data where CMY >=80% Group 2 (mix): middle part, where 20% < CMY <80% Group 3 (represented mostly by FY ice): all the data where CMY <= 20% The tables below show standard deviation and bias for SSMI and AMSR for groups 1 and 3. Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI page 64 of 189

65 Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-9: SIC = 100%, Northern Hemisphere, summer, MYI vs FYI Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI Near 90GHz lin, dyn Near 90GHz ASI_NWF P page 65 of 189

66 Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-10: SIC = 100%, Northern Hemisphere, winter, MYI vs FYI Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P page 66 of 189

67 Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-11: SIC = 100%, Southern Hemisphere, summer, MYI vs FYI Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team page 67 of 189

68 Standard deviation Average Algorithm AMSR SSM/I AMSR SSM/I MYI FYI MYI FYI MYI FYI MYI FYI NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF Nr. of points Table 4-12: SIC = 100%, Southern Hemisphere, winter, MYI vs FYI Figure 4-9: Standard deviation MYI vs FYI, SIC = 100%, Northern hemisphere, summer, SSM/I page 68 of 189

69 Figure 4-10: Standard deviation MYI vs FYI, SIC = 100%, Northern hemisphere, winter, SSM/I Figure 4-11: Standard deviation MYI vs FYI, SIC = 100%, Southern hemisphere, summer, SSM/I page 69 of 189

70 Figure 4-12: Standard deviation MYI vs FYI, SIC = 100%, Southern hemisphere, winter, SSM/I Results from Winter show (tables 4-10 and 4-12 and figure 4-10 and 4-12) that in the Arctic FY variability during winter is larger than MY, whereas in the Southern Hemisphere it is the other way around. Variability is generally smaller than for the complete dataset indicating that the algorithms have larger difficulties at FY/MY mixtures than for purer surface types. Summer results are not easy to interpret (see beginning of chapter 4). page 70 of 189

71 5 SIC = 85% The 85% artificial dataset was generated similar to the 15% dataset (see section 3), but with full variability of ice and a 15% average open water signature: SIC * SIC100( t) 0.15 * SIC 0( OW ) (5-1) Summer of the Southern Hemisphere is not included here due to shortness of available dataset (only 41 measurements for AMSR and SSM/I in total). Similarly to the 15% comparison the 85% nominal reference is done to be able to compare algorithms and their standard deviation of the ice concentration with cut-offs and non-linearity at 100% ice concentration e.g. the NASA Team 2 and ASI. It turns out that these two algorithms have large positive biases which are affecting the standard deviation even at the 85% reference. Differences in standard deviation and in the bias between summer and winter may indicate sensitivity to noise which may have a climatological trend e.g. the melt season length or atmospheric opacity. All of the algorithms have a higher standard deviation of the ice concentration in summer than in winter. The difference in summer winter standard deviation is 2-3% for e.g. Bootstrap - P, Bristol, NASA Team, ECICE and Bootstrap - F. The ASI has a difference of 3.8% and the NASA Team2 of 4.9%. The standard deviation for these two algorithms is affected by the high winter bias. The summer winter difference in averages is 2-6% for the ASI, Bristol, Bootstrap - F, the NASA Team 2 and ECICE. The difference is 11% for the NASA Team and 14% for the Bootstrap - P. 5.1 Northern hemisphere Summer (June September) Most algorithms have a negative bias in summer except ASI with an average ice concentration at 92.8% and NASA Team 2 at 92.5%. With standard deviations near 7% to 8% the cut-off may have reduced the standard deviations even at the nominal 85% reference. Many algorithms e.g. the ASI, Bootstrap - P, Bristol, NASA Team, Bootstrap - F, ECICE have standard deviation of the ice concentration around 6-8% in summer. NASA Team 2 has a standard deviation in the high end of Many of the algorithms have a negative bias during summer due to melt ponds and other snow cover processes during summer. The Bootstrap - P and the NASA Team have large negative biases with averages for the 85% reference ice concentration of 69% and 71% respectively. The ASI and the NASA Team 2 have large positive biases with averages at 93%. Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn page 71 of 189

72 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 5-1: SIC = 85%, Northern Hemisphere, summer Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz page 72 of 189

73 Algorithm Standard deviation Average ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 5-2: SIC = 85%, Northern Hemisphere, summer. Average over all the instruments present for given algorithm page 73 of 189

74 Figure 5-1: Standard deviations, SIC = 85%, Northern hemisphere, summer Figure 5-2: Bias, SIC = 85%, Northern hemisphere, summer page 74 of 189

75 5.1.2 Winter (October May) In winter nearly all the algorithms have small biases except the ASI and the NASA Team 2 with averages at 97%. At these high averages the cut off is affecting the standard deviation of the ice concentration which is then artificially low. The Bristol and the ECICE have standard deviations near 3.5% while NASA Team, Bootstrap P and F are near 5%. Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points page 75 of 189

76 Table 5-3: SIC = 85%, Northern Hemisphere, winter Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF NASA Team ECICE Nr. of points Table 5-4: SIC = 85%, Northern Hemisphere, winter. Average over all the instruments present for given algorithm page 76 of 189

77 frequency 200 frequency N90Lin ASI frequency 300 frequency Bootstrap_P Bristol frequency frequency NASA TEAM Bootstrap_F frequency frequency NT2 ECICE Figure 5-3: Histograms of SIC for SSMI Northern hemisphere SIC85 data (Summer and Winter combined). The orange bar marks SIC=85%. page 77 of 189

78 Figure 5-4: Standard deviations, SIC = 85%, Northern hemisphere, winter Figure 5-5: Bias, SIC = 85%, Northern hemisphere, winter page 78 of 189

79 Near90GHz 120 ASI OSISAF OSISAF Near90_lin_dyn 120 Bootstrap_p OSISAF OSISAF ECICE 120 Bootstrap_f OSISAF OSISAF NT2 120 NASA_Team OSISAF OSISAF Figure 5-6: Bias, SIC = 85%, Northern hemisphere, winter. Examples of correlation between SSMI SIC85 derived by different algorithms. Note the saturation of the NT2 algorithm and the ASI algorithm. page 79 of 189

80 5.2 Southern hemisphere Winter (April November) The standard deviations of the ice concentration for the different algorithms on the southern hemisphere in winter are quite different than on the northern hemisphere. For example the Bootstrap P has a standard deviation on the northern hemisphere of 5.37% and 10.1% on the southern hemisphere. The Bootstrap F which had a standard deviation of 5.11% on the northern hemisphere was 3.77% on the southern hemisphere. The Bristol has the same on both hemispheres around 3.5%. Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF page 80 of 189

81 Algorithm Standard deviation Average AMSR SSM/I AMSR SSM/I OSISAF NASA Team ECICE Nr. of points Table 5-5: SIC = 85%, Southern Hemisphere, winter Algorithm Standard deviation Average Near 90GHz lin, dyn Near 90GHz ASI_NWF P P Bootstrap P P Bristol PR NASA Team NORSEX Bootstrap F CalVal UMass-AES P One channel (6H) TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF 2 *N90lin_dyn)/(1+BF 2 ) (BF+BF 3 *N90lin_dyn)/(1+BF 3 ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) OSISAF OSISAF OSISAF page 81 of 189

82 Algorithm Standard deviation Average NASA Team ECICE Nr. of points Table 5-6: SIC = 85%, Southern Hemisphere, winter. Average over all the instruments present for given algorithm Figure 5-7: Standard deviations, SIC = 85%, Southern hemisphere, winter page 82 of 189

83 Figure 5-8: Bias, SIC = 85%, Southern hemisphere, winter Histograms in this section show that several algorithms have biases so large that it influences (reduces) their standard deviation and thus make them look better than they really are at intermediate (high) concentration. However, the results are somewhat easier to interpret than the SIC1 results since this effect is much smaller. Note that NT2 has a positive bias of >10% for SSMI and almost 15% for AMSR data. This is most likely due to the fact that the algorithm was originally developed for the original near-real-time SSMI data-stream. It is unclear how much this bias (lack of calibration) influences the performance of its built-in atmospheric correction. Figure 5-9 below shows that most of the 12 possible atmospheres come into play. Most algorithms have very small biases (~1%) and standard deviations in the order of 3-5%. Higher frequency algorithms performs somewhat worse than lower frequency ones. Most algorithms have very similar performance for AMSR and SSMI. page 83 of 189

84 1200 NT2 Atmosphere frequency Figure 5-9: NT2 atmosphere, SIC = 85%, Northern hemisphere, all year NT2 Atmosphere frequency Figure 5-10: NT2 atmosphere, SIC = 100%, Northern hemisphere, all year. page 84 of 189

85 6 Thin ice 6.1 Collection and visualisation of dataset ESA s L-band sensor SMOS is intended to observe soil moisture and ocean salinity (SMOS). Because of its long wavelength the observations are also suitable to retrieve the thickness of thin sea ice data. Two algorithms have been suggested, one based on intensity and incidence angles below 40 [Kaleschke et al. 2012], and one based on simultaneously using both intensity and polarization difference at incidence angles above 40 [Heygster et al., 2012]. Here, the latter procedure is applied to observations of the Arctic in the period of October to December 2010, the first Arctic freezing season, which was observed by SMOS. All processing is based on SMOS L1C version 3.46 data. These data are organized in Discrete Global Grid (DGG) cells of 9 km spacing in an Icosahedral Snyder Equal Area projection with aperture 4 and resolution 9 (ISEA 4H9). The grid cells have a size of about 15x15 km 2. Observations nearer than 50 km from land have been excluded. All SMOS snapshots containing one or more pixels with brightness temperature higher than 300 K are considered RFI contaminated and excluded. The mere retrieval is organized in four steps: 1. convert observations of the four Stokes components from the instrument reference frame to the surface reference frame, 2. averaging of all observations of one day within one DGG and incidence angle range of 40 to retrieval of ice thickness based on intensity of polarization difference values 4. interpolate results to polar stereographic projection of 12.5 km (NSIDC grid). Subsequently, regions of homogeneous sea ice thickness have been identified by selecting regions of limited local standard deviation. For this purpose, the resulting sequence of sea ice maps is considered as a 3-d data cube with two horizontal dimensions and one time dimension. The local standard deviation is calculated using convolution operations of variable size in horizontal direction in units of 12.5x12.5 km 2 pixels and in time direction in units of days. A (7,7,3) window (x,y,t) has been selected in order to guarantee sufficient homogeneity in a region larger than the footprint of the SSM/I channel of the lowest frequency involved in the SIC retrieval (19 GHz, 70x43 km 2 ). In addition, the days before and after are included when calculating the standard deviation in order to exclude regions of strong change from one day to the next which might be an indicator of variable ice concentration. The criteria are set to identify large regions with thin ice dominated by thermodynamical growth and at 100% ice concentration. The assumption here is that if an area is stable over 3 days during winter freeze up it is likely to be totally ice covered. A partial ice cover would show variability over the 3 days and be discarded. We evaluated the validity of this assumption by chacking the SIC with SAR data from the same areas on a few days (where SAR data were available). page 85 of 189

86 As period of largest extent of Arctic thin ice, the 18 days from 2 to 19 November 2010 have been. After a set of tests, the data fulfilling the following conditions were selected: ice thicknesses from 4 to 30 cm to avoid open water and noisy retrievals. We included some results for ice thicknesses up to 50 cm but as can be seen from fig 6-5 we have very little data for those thicknesses. ice thickness standard deviation below 2 cm. With this procedure, in total 4784 pixels were identified, each resulting in one data set for which the following parameters were extracted for the Round Robin exercise: latitude, longitude, x-position, y-position, thickness, standard deviation, date Figure 6-1 shows the map of all extracted positions and the histogram of sea ice thickness. Figure 6-1: Left: Example map of identified thin ice regions. Right: histogram of thicknesses distribution for the included data points 6.2 Results for 2010 The plots below shows Sea Ice Concentration calculated by the implemented algorithms as a function of SMOS ice thickness. Ice thicknesses up to 50 cm are investigated, but very few data points exist above ca. 35 cm, so results for the thicker ice are more uncertain. page 86 of 189

87 Near90_lin_dyn Near90GHz ASI P90 P37 Bootstrap_p P18 Bristol PR NASA_Team NORSEX Bootstrap_f CalVal UMass_AES P10 One_channel TUD Figure 6-2: Plot including a subset of algorithms Note that all algorithms show a severe underestimation of SIC for ice thickness up to cm Near90_lin_dyn Bootstrap_p Bristol NASA_Team Bootstrap_f TUD osisaf Figure 6-3: Selected algorithms from the subset (a) page 87 of 189

88 Performance of selected algorithms is shown in figures above and below Near90GHz ASI Bootstrap_f P10 One_channel combo Figure 6-4: Selected algorithms from the subset (b) Figure 6-5: Number of SMOS data points per thickness category (1 cm). Above 35 cm the number of data points is small page 88 of 189

89 Figure 6-6: Algorithm estimation of SIC for 100% 5cm thick ice Figure 6-7: Algorithm estimation of SIC for 100% 10 cm thick ice page 89 of 189

90 Figure 6-8: Algorithms estimation of SIC for 100% 15 cm thick ice page 90 of 189

91 Figure 6-9: Algorithms estimation of SIC for 100% 20 cm thick ice Figure 6-10: Algorithms estimation of SIC for 100% 25 cm thick ice page 91 of 189

92 P10 PR P18 NASA_Team One_channel P37 Bootstrap_p combo1 combo3 combo2 combo4 Near90_lin_dyn Bristol osisaf osisaf2 osisaf3 combo7 combo8 combo5 combo6 Bootstrap_f CalVal UMass_AES TUD NORSEX Near90GHz P90 ASI Figure 6-11: Algorithm estimation of SIC for 100% ice at various thicknesses. The 35 cm data should be used with caution due to the limited number of datapoints at this thickness. 6.3 TB variability The following plots show AMSR-E brightness temperatures as a function of SMOS ice thickness colocated at the 75 km grid cell size of the SMOS thin ice data page 92 of 189

93 V 10.7V 18.7V 23.8V 36.5V 89.0V Figure 6-12: AMSR vertically polarized TBs as a function of SMOS ice thickness for our validation dataset H 10.7H 18.7H 23.8H 36.5H 89.0H Figure 6-13: AMSR horizontally polarized TBs as a function of SMOS ice thickness for our validation dataset. page 93 of 189

94 It has been suggested to use AMSR PR as a proxy for ice thickness. The plot below therefore shows PR(h) PR6 PR10 PR18 PR23 PR36 PR Figure 6-14: AMSR Polarization ratio (PR) as a function of SMOS ice thickness For completeness we also include a figure showing NASA-Team Multi-year ice concentration as a function if SMOS ice thickness Figure 6-15: NASA-Team multiyear ice concentration as a function of SMOS ice thickness page 94 of 189

95 All algorithms underestimate SIC for thin ice. The near 90 GHz algorithms perform slightly better than the rest, but several others in particular the Bootstrap F family of algorithms using 19v and 37v GHz channels perform almost as good. The polarisation type of algorithms e.g. the NASA Team and the Bootstrap P yield relatively low ice concentrations for thin ice types. There may be a correlation between thin ice and ice concentration on the scale of the footprint and it is not expected that the algorithms yield 100% for very thin ice. It is noted that the magnitude of TB s at vertical polarisation for ice and water are comparable at near 90 GHz and very different at 6 GHz while the contrast in polarisation difference between ice and water is comparable at 6 GHz and near 90 GHz. If there is an ambiguity between the thin ice concentration and its thickness this would explain that all polarisation type of algorithms are yielding similar results. The simulations in chapter 8 show that the transition from bare thin ice to snow covered thin ice is dramatic radiometrically and some of the ice thickness signatures for some of the algorithms may just be related to the snow cover. We do not know if the thin ice we have identified in our observations are snow covered or not. The fact that there are nearly no detection of ice which is near 0.35 m thick in Figure 6-5 could point to the fact that there really is a thin ice thickness and concentration ambiguity. Ice which is >0.35m thick covers the surface while thinner ice is also low concentration ice with open water in between. page 95 of 189

96 7 Melt ponds In terms of radiometric signatures open melt ponds are similar to cracks and leads in the ice. This means that the ice concentration of melting sea ice is one minus the melt pond fraction and the lead fraction. Melt ponds are not open water all the time. Sometimes they freeze up with a thin ice cover. This further complicates the interpretation and when there are melt ponds a raised level of uncertainty is expected. When sea ice is melting melt water is draining into the melt-ponds when they have formed and into the leads between the floes. In the leads there is often a very stable stratification with nearly fresh water at the surface. Even at L-band which is sensitive to salinity it will therefore be difficult to distinguish melt ponds from leads. Gridded independent SIC estimates from MODIS has been collocated with the coincident melt pond cover fraction (DARD: ID 2.08 and 2.09) and SICCI SIC products to investigate the impact of melt ponds and the performance of the algorithm in such melt-pond cover infested areas [Roesel et al., 2012b]. Reflectances measured by bands 1, 3, and 4 of the MODIS sensor and provided by NASA as MODIS Surface Reflectance 8-Day L3 Global 500m SIN Grid V005 - product (MOD09A1) and MODIS Surface Reflectance daily L2G Global 500 m and 1 km - product (MOD09GA) are projected onto a polarstereographic grid with 500 m grid resolution, applying land, cloud and other flags provided with the original MODIS product. The analysis is limited to tiles covering the Arctic Ocean, i.e. north of 60 N. A spectral un-mixing process is applied together with an artificial neural network to classify the reflectances of each grid cell into the three surface classes: open water, melt ponds, and snow / ice. For the final melt pond fraction product, surface class distributions are interpolated onto a 12.5 km grid resolution polarstereographic grid; resulting netcdf files contain melt pond fraction (MPF), number of valid (non-flagged) 500 m grid cells per 12.5 km grid cell (N), standard deviation of MPF according to N (MPF-SD), and the open water fraction (OWF). This is the standard product available via ICDC ( The RRDP requires high quality, high ice concentration data at 100 km grid resolution. For the RRDP product, 8x km grid cell averages are computed for MPF, MPF-SD, and OWF. For this a binary mask is produced, comprising 12.5 km grid cells with above 90% ice concentration (computed as 100% - OWF) and with clear-sky conditions (100% valid grid cells). This mask is applied and 100 km grid resolution average values are computed for those 8x8 grid-cell boxes where at least 90% of the cells belong to the binary mask. For the grid cells selected this way, the RRDP contains 8-day composite MPF data for years for the entire Arctic and daily MPF data for the Kara Sea and the area North of Greenland for 2009; the period for the latter subset was chosen in accordance with melt pond retrieval activities at FMI. For the RRDP a special 1-day dataset has been produced in order to enable accurate collocation in time with satellite PMR data. page 96 of 189

97 7.1 Comparison between SIC and open water fraction From the MODIS MPF and SIC data the area fraction of ice (C) is calculated as C=(1-W) = 1-(1-MSIC)-(MSIC*MPF) (7-1) where C is surface fraction of ice W is surface fraction of water (leads + melt ponds) MSIC is sea ice concentration from the MODIS data MPF is melt pond fraction derived from MODIS (the fraction of ice that is covered by melt ponds) The following plots show SIC calculated by our implemented algorithms as a function of C. A total of 8152 datapoints were selected for the analysis where standard deviation of melt pond fraction over the 100x100 kilometers was less than 5%, standard deviation of SIC < 5% and SIC > 95% from the MODIS MPF dataset. 0.6 Melt pond fraction Figure 7-1: Melt pond fraction throughout the dataset. X-axis goes from June 1 to August 31. July 1 is at 5310, August 1 at 8122, so very few points from August. page 97 of 189

98 ASI Near90Lin Figure 7-2: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) Bootstrap F NORSEX Figure 7-3: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) page 98 of 189

99 NASA Team Bristol Figure 7-4: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) Bootstrap_P OSISAF Figure 7-5: SIC derived from selected algorithms as a function of C computed according to (Eq 7.1) Ideally the lines in figures 7-2 to 7-5 should have a slope close to 1 (the black line) for sea ice concentration as a function of melt-pond fraction because open melt ponds are seen as open water. The Bootstrap F family have slopes around 1 compared to the Hybrid type of algorithms e.g. the Bristol slightly below 1 and the polarisation type of algorithms e.g. the NASA Team less than 1. The smaller sensitivity to the melt-pond fraction for the polarisation type of algorithms with slope less than 1 indicate that they have a relatively high sensitivity to snow and ice emissivity and temperature variability and/or atmospheric effects. The ASI_NWF does not have sensitivity to melt-ponds. When selecting dynamical tie-points and initial estimate of the ice concentration is needed. This estimate should be provided by an algorithm with sensitivity to melt-ponds not to bias the tie-points with meltpond infested measurements. The tables below show more quantitative details of the data analysis. page 99 of 189

100 Note that all algorithms have a positive bias relative to the y=x line (black). This indicates that there may be a bias in the MODIS MPFs. Since the bias is present over the whole range of C=(1-W) and SIC>0.95 for all data points the bias cannot only be in SIC even though the bias can be adjusted using seasonally varying tie-points. All June July August avg stdev avg stdev avg stdev avg stdev SIC SIC std MPF MPFstd (1 W) V H V H V H V H V H V H Table 7-1: Average values and stdevs for MODIS derived SIC and MPF and for AMSR TBs Table 7-1 shows a summary of SIC and MPF derived from MODIS data split in the 3 summer months. Note the melt pond fraction is 13% in June increasing to 30% in July and 27% in August for the analysed areas. page 100 of 189

101 All June July August avg stdev avg stdev avg stdev avg stdev C=1 W Near90_lin_dyn Near90GHz ASI_NWF P P Bootstrap_p P Bristol PR NASA_Team NORSEX Bootstrap_f CalVal UMass_AES P One_channel TUD (NT+BF)/ (NT+BF+N90lin_dyn)/ (P37+N90lin_dyn)/ (P37+N90lin_dyn+BF)/ (BF+BF*N90lin_dyn)/(1+BF) (BF+BF3*N90lin_dyn)/(1+BF ) (BF+N90lin_dyn)/ (BF+BF*N90lin_dyn)/(1+BF) osisaf osisaf osisaf CMY Table 7-2: Calculated SIC and their stdevs for the tested algorithms. Upper row in red are MODIS derived reference values. The bottom row is multi-year ice concentration as derived with the Nasa Team algorithm page 101 of 189

102 8 Simulated data 8.1 The simulated sea ice concentration Microwave brightness temperatures used for computing the ice concentration are sensitive to noise from the atmosphere and surface emission variability. Even though the sensitivity to noise is minimized in sea ice concentration algorithms in general the estimated ice concentration may still have some sensitivity to noise left. Over open water the dominating noise sources are wind roughening of the water surface, water vapor in the atmosphere and cloud liquid water. Over ice the atmosphere play a minor role except at near 90 GHz where liquid water is a noise source. Over ice the noise is dominated by snow and ice temperature variability, the snow depth and grain size variability, and to some extent snow surface density variability as a proxy for layering in the snow. Snow layering and surface roughness effects can be investigated on a case by case basis but this is beyond the scope here. These parameters mentioned above are investigated in this chapter. In these simulations the real ice concentration is 1 over ice and 0 over open water. All deviations from 1 and 0 in the estimated ice concentrations are caused by sensitivity to noise. Traditionally weather filters have been applied to avoid noise over open water. A weather filter is an ice - water classifier which is truncating pixels classified as open water to 0 % ice concentration. This removes noise in open water regions and it may remove real low concentration ice or new-ice along the ice edge. It doesn t work over ice. Some processing facilities e.g. EUMETSAT s OSI SAF is using explicit correction of the brightness temperatures before computing the ice concentration. This is a spatially/temporally varying noise reduction which is working over both ice and open water. The correction is using NWP data of wind, temperature, and water vapor and an atmospheric radiative transfer model to correct the brightness temperatures. This procedure requires dynamical tie-points to avoid potential biases from the model. Even though there are very good radiative transfer models for the atmosphere it is not possible to correct for all noise sources. For example, the representation of cloud liquid water in NWP models is not suitable for correction. The parameters in the snow and ice are difficult to measure or quantify in numerical models and they have not been used for explicit correction. It is therefore important to find algorithms with low sensitivity to physical parameters which are difficult to correct for. This includes cloud liquid water and snow and ice parameters in general. 8.2 The simulated data Combined thermodynamic and emissivity models have the potential to build long snow/sea ice/microwave time-series that can be used for statistical analysis of radiometer sea ice data sensitivities (Mätzler et al., 2006). However, Wiesmann et al. (2000) show that one-dimensional thermodynamic models for snow and frozen ground including microphysical parameters and a vertical stack of layers, for example, SNTHERM (Jordan, 1991) and Crocus (Brun et al., 1989), underestimate the formation of thin crusts or weak layers in the snow pack. Comparison to snow pit page 102 of 189

103 measurements showed that the density of thin layers is underestimated in Crocus and thin layers are not represented properly in SNTHERM. Further, when the thermodynamic model output is used as input to a microwave emission model this leads to underestimation of the simulated polarization difference. These models were developed for other applications such as avalanche risk assessment. The thermodynamic model used here treats layers from individual precipitation events and retains all layers even when thin (1 mm) in an attempt to alleviate earlier problems in microwave modeling applications (Tonboe, 2005). Representing the layering in the snow is very important for simulating realistic Tb s and in particular the Tv and Th polarization difference (Wiesmann et al., 2000). To ensure reasonable initial snow layer thickness precipitation events less than 1 kg/m2 over 6 h are retained and released only when the next precipitation event exceeds the threshold. While this may not be totally realistic it does produce simulated snow depths which are comparable to climatology (Warren et al., 1999). An earlier investigation of different sea ice concentration algorithms by Tonboe and Andersen (2004) was using measured profiles as input to the emission model thus changing the density and grain size of specific layers in the snow-pack. It showed that the frequency algorithms had low sensitivity to these changes while the algorithms using near 90 GHz channels had a high sensitivity to the density of the snow surface. 8.3 The microwave emission models Sea ice emission models relate physical snow and ice properties such as density, temperature, snow crystal and brine inclusion size to microwave attenuation, scattering and reflectivity. The model used here is a sea ice version of MEMLS (Wiesmann and Mätzler, 1999) described in Mätzler et al. (2006) and hereafter called the emission model. The theoretical improved Born approximation was used here, which validate for a wider range of frequencies and scatterer sizes than the empirical formulations (Mätzler and Wiesmann, 1999). MEMLS, using an empirical scattering formulation, has earlier been validated for snow on land in the GHz region (Wiesmann and Mätzler, 1999). The concern at higher frequencies is the validity of this empirical scattering formulation. The theoretical improved Born approximation is in principle also valid in the GHz range and for large scatterers (Mätzler and Wiesmann, 1999). Using the improved Born approximation the shape of the scatters is important for the scattering magnitude (Mätzler, 1998). We assume spherical scatters in snow when the correlation length pec, a measure of grain size, is less than 0.2 mm and the scatters are formed as cups when greater than 0.2 mm to resemble depth hoar crystals. The sea ice version of MEMLS includes models for the sea ice dielectric properties while using the same principles for radiative transfer as the snow model. Again the scattering within sea ice layers beneath the snow is estimated using the improved Born approximation. The scattering in first-year ice and multi-year ice is assumed from small brine pockets and air bubbles within the ice respectively. All simulations are at 50 0 incidence angle similar to AMSR, SSMIS and other conically scanning radiometers. In addition to the surface emission model for sea ice we use a Wentz model for simulating the atmospheric emission and absorption and the open water emissivity. The Wentz model is an open water model specifically developed for different radiometers SMMR (Wentz, 1983), SSM/I, excluding the 85GHz channels (Wentz, 1997) and AMSR (Wentz and Meissner, 2000). page 103 of 189

104 Each of them has been modified so that they can be used over ice by including the sea ice emissivity and effective temperature as input. 8.4 The snow and sea ice thermodynamic and mass model In order to produce input to the emissivity model a one-dimensional snow/ice thermodynamic model has been developed. Its purpose is not necessarily to reproduce a particular situation in time and space rather to provide realistic microphysical input to the emissivity model. Earlier thermodynamic models such as Maykut and Untersteiner (1971) or even simple degree day models (see e.g. Maykut, 1986) are useful for simulating ice thickness. However, for microwave emission modeling applications additional parameters such as temperature, density, snow grain size and ice salinity at very high vertical resolution are needed. Because the thermal conductivity is a function of temperature the model uses an iterative procedure between each time step of 6 h. The thermodynamic model is fed with ECMWF ERA40 data input at these 6 h intervals. In return the thermodynamic model produces detailed snow and ice profiles which are input to the emission model at each time step. A Wentz model is used for simulating the atmospheric emission and absorption. This gives a picture of significant emission processes in sea ice even though the one-dimensional thermodynamic model is not capturing the spatial variability of the sea ice cover caused for example by ice convergence resulting in deformation, ice divergence resulting in new-ice formation, and wind redistribution of the snow cover affecting snow depth, density and grain size. The thermodynamic model has the following prognostic parameters for each layer: thermometric temperature, density, thickness, snow grain size and type, ice salinity and snow liquid water content. Snow layering is very important for the microwave signatures therefore it treats snow layers related to individual snow precipitation events. For sea ice it has a growth rate dependent salinity profile. The sea ice salinity is a function of growth rate and water salinity (Nakawo and Sinha, 1981). 8.5 Simulation procedure The input at 6 hour intervals to the thermodynamic model is ECMWF ERA 40 re-analysis climate data: air pressure, air temperature, wind speed, humidity, precipitation, incoming short wave radiation, and incoming long wave radiation. The data are used in the model for computing the surface energy balance and snow accumulation. In addition to the parameters which are used for the thermodynamical model also the cloud liquid water and water vapor are extracted from the ERA40 data. These are used as input to the Wentz model. The output from the thermodynamic model at 6 hourly time-steps is a snow and sea ice profile including for each layer: The temperature, the density, the correlation length, the salinity, the snow or ice type is input to the snow and ice emission model. The emission model is computing the emissivity and the effective temperature which is used together with the cloud liquid water and water vapor as input to the Wentz atmospheric model. Over open water the Wentz model is used alone. The output is the top of the atmosphere brightness temperatures which can be used as input to the ice concentration algorithms and all other physical parameters describing the system. 8.6 The simulated data (Antarctic cases) page 104 of 189

105 Two sites were selected in Antarctica: One grid point in the Ross Sea (75 S, 200 E) for a first-year ice floe and one in the Bellinghausen Sea (64 S, 280 E) in open water. Two sites were selected in the Arctic: One in the Lincoln Sea (85 N, 240 E) in 100% multiyear se ice and one in the Norwegian Sea (70 N, 0 E) in open water. Even though we have identified and constructed about 20 different sea ice concentration algorithms it is clear that there is a limited number of algorithm families where the sensitivity to noise is almost similar. There is the gradient (e.g. CF and CalVal), the polarization (e.g. NT and CP), the high frequency (e.g. ASI_NWF and N90LIN), the single channel (e.g. OneChannel - Tb19h, OneChannel - Tb6h) and the hybrid combining gradient and polarization (e.g. Bristol and TUD) families. The single channel algorithms are not real candidates in the algorithm selection but the OneChannel - Tb6h is included because of its low sensitivity to the atmosphere and the surface emissivity variability. Radiometers measuring at 6 GHz were on SMMR and on AMSR and now on AMSR onwards. The ESMR radiometer on NIMBUS 5 covered the important period before modern multi frequency radiometers from 1972 to It measured at a single channel at Tb19h. The Tb19h channel is included on all multi frequency radiometers from 1978 until today. We have therefore selected the 9 algorithms representing different families in table 1. Algorithm Category NASA Primarily 19 GHz polarization Bristol 19 and 37 GHz hybrid Bootstrap F 19 and 37 GHz gradient Bootstrap P 37 GHz polarization TUD 19, 37 and near 90 GHz hybrid ASI_NWF Near 90 GHz high frequency N90LIN Near 90 GHz high frequency OneChannel - Tb6h Low frequency single channel OneChannel - Tb19h ESMR single channel Table 8-1: Categorization of the 9 selected algorithms. The polarization algorithms are using the polarization difference or ratio. The gradient algorithms are using the spectral gradient e.g. at Tb19v and Tb37v. The hybrid refers to a combination of polarization and gradient. Here low frequency is 6 GHz and high frequency is near 90 GHz. ESMR is the single channel (Tb19h) radiometer on NIMBUS 5 The surface parameters in snow and sea ice affecting the thermal microwave emission and the estimated ice concentration are a stretch target for NWP and sea ice models at present. The parameters are even difficult to measure in the field and simulate using detailed process models. The complexity of the atmosphere snow ice system makes it difficult to identify and define parameters and the parameters are often not independent as seen in the correlation matrix e. g. the average snow correlation length and the snow depth and the snow temperature gradient. Other parameters such as the atmospheric water vapor and the snow surface temperature are also correlated page 105 of 189

106 Dens appc Ti ist St Snowg Vapor Liquid NT Bristol BF BP TUD ASI N90L Dens Apcc Ti Ist St Snowg Vapor Liquid NASA Bristol BF BP TUD ASI N90L 1.00 Table 8-2: The Ross Sea correlation matrix. The snow surface density: Dens, The average snow correlation length: appc, The snow surface temperature: Ti, The snow ice interface temperature: ist, The snow depth: St, the snow temperature gradient: snowg, The atmospheric water vapor: Vapor, The cloud liquid water: Liquid. page 106 of 189

107 Figure 8-1: The simulated snow and ice profile in the Ross Sea (75 S, 200 E) The ice in the simulated profile from the Ross Sea shown in Figure 8-1 is growing from the initial 2 cm at the beginning of the season to about 150 cm at the end of the cold season. The snow cover gradually accumulates during several snow precipitation events to a thickness of about cm. page 107 of 189

108 Figure 8-2: The simulated Ross Sea ice concentration After the initial ice growth and snow accumulation phase the simulated sea ice concentration from the different algorithms gives near 100% estimates as shown in Figure 8-2. It is unclear whether the initial phase is a realistic representation of thin ice or a shortcoming of the thermodynamic model to provide a realistic snow profile to the emission model. Therefore the initial 120 iterations (30 days) have been excluded from the analysis in all ice profiles and we only include mature ice where the simulated ice concentration is near 100%. page 108 of 189

109 Figure 8-3: The 9 different ice concentration algorithms sensitivity to the snow ice interface temperature in the Ross Sea profile (see figure 8-1). Figure 8-3 shows the 9 different algorithms sensitivity to the snow - ice interface temperature. For cold temperatures (<260 K) the estimated ice concentration is increasing with temperature for the Bristol, Bootstrap-F, the TUD, the OneChannel Tb6h, and the OneChannel - Tb19h. For all algorithms except the ASI_NWF the ice concentration is decreasing with temperature between 260 K and 270 K. The high frequency algorithms do not penetrate to the snow ice interface. Even though the ice concentration variability is plotted as a function of the snow ice interface temperature, water vapor etc. it is often difficult to see the specific sensitivity to each of the parameters since all the parameters are varied at the same time. Further, some of the parameters are correlated as will be discussed later. Figure 8-3 is a good example of why a slope may not indicate a particular sensitivity. page 109 of 189

110 Figure 8-4: Sensitivity of 9 sea ice concentration algorithms to cloud liquid water at the Ross Sea ice simulation Over sea ice cloud liquid water is not one of the primary noise sources except perhaps at the near 90 GHz algorithms i.e. N90LIN and TUD. Other algorithms shown in Figure 8-4 seem insensitive to cloud liquid water such as the Bristol. page 110 of 189

111 Figure 8-5: The sensitivity of 9 algorithms to snow depth at the Ross Sea ice profile The estimated ice concentration is increasing with snow depth for the polarization algorithms the NASA Team, the Bootstrap P, and the OneChannel Tb6h. page 111 of 189

112 Figure 8-6: The sensitivity of 9 sea ice concentration algorithms to the snow temperature gradient in the Ross Sea ice profile The snow temperature gradient is not one of the primary parameters affecting the estimated sea ice concentration noise. None of the algorithms show sensitivity to the snow temperature gradient as seen in Figure 8-6. The snow temperature gradient is defined as the Tsnow surface - Tsnow-ice interface difference over the snow depth. page 112 of 189

113 Figure 8-7: The sensitivity of 9 sea ice concentration algorithms to the snow surface density in the Ross Sea ice profile The snow surface density is not one of the primary noise sources affecting the estimated sea ice concentration. Of the 9 algorithms shown in Figure 8-7 none shows a clear relationship between snow surface density and the estimated ice concentration. page 113 of 189

114 Figure 8-8: The sensitivity of 9 sea ice concentration algorithms to atmospheric water vapor in the Ross Sea ice profile Water vapor is not one of the primary parameters affecting the estimated sea ice concentration noise over ice. Other parameters dominate. However, the high frequency algorithms i.e. the TUD and the N90LIN estimated ice concentration increase with water vapor. page 114 of 189

115 Figure 8-9: The sensitivity of 9 sea ice concentration algorithms to the average snow correlation length in the Ross Sea ice profile The estimated sea ice concentration is not sensitive to the average snow correlation length as shown in Figure 8-9. Several of the algorithms may be sensitive to scattering in the snow. However, the average snow correlation length is not good measure of the scattering magnitude. page 115 of 189

116 Figure 8-10: The sensitivity of the 9 sea ice concentration algorithms to the snow surface temperature in the Ross Sea ice profile Algorithm The TUD, the N90LIN and the OneChannel Tb19h ice concentration estimates increase with the snow surface temperature as shown in Figure Observed Winter std mean IC incl. atm. IC std incl. atm. IC std excl. atm. NASA Bristol Bootstrap F Bootstrap P TUD ASI N90LIN OneCannel - Tb6h OneCannel - Tb19h mean atm/no atm IC ratio Table 8-3: mean ice concentration, the standard deviation with or without atmosphere and the mean atmosphere/ no atmosphere ice concentration ratio for the 9 algorithms in the Ross Sea ice profile page 116 of 189

117 Table 8-3 shows the mean ice concentration, the standard deviation with or without atmosphere and the mean atmosphere/ no atmosphere ice concentration ratio for the 9 algorithms in the Ross Sea ice profile. The no atmosphere case is the surface emission only. The atmosphere included is the atmospheric emission absorption and reflection computed with a modified Wentz model in addition to the surface emission. This includes oxygen absorption. Most of the algorithms are well tuned for this purpose with small overall biases only the TUD with a bias of 9%, the N90LIN with a 12% bias and the OneChannel Tb19h with a 16% bias shown in Table 8-3. For this example an acceptable bias is one where the absolute ice concentration estimate is near 100% for evaluation and comparison of the algorithms at this point. Anyway, these relatively small biases are not posing a problem for the analysis and we retain the set-up with the algorithms using the round robin northern hemisphere tie-points. The standard deviation of all 9 algorithms is either unchanged or reduced with the atmospheric contribution included. This means that the atmosphere acts as a smoother on the brightness temperature and ice concentration variability. The mean ratio of the ice concentration computed with or without atmosphere is greater than one except the Bootstrap F (0.99) and the ASI_NWF (0.99) meaning that the atmospheric contribution generally increase the estimated ice concentration. The Bootstrap P (1.06) and the N90LIN (1.07) have the highest atmosphere/ no atmosphere ice concentration ratios. Ice Surface T Snow Surf. Vapor Liquid surface T depth density NASA Bristol CF CP TUD ASI N90L Table 8-4: The partial correlation of each of the 6 parameters (the snow ice interface temperature, the snow surface temperature, the snow depth, the snow surface density, atmospheric water vapor, cloud liquid water) and the ice concentration with the effects of the other physical parameters (snow surface density, the average snow correlation length, snow surface temperature, the snow ice interface temperature, snow depth, snow temperature gradient, atmospheric water vapor, cloud liquid water, excluding one of the 6) removed Several of the physical parameters in the atmosphere snow ice system are correlated as shown in Table 8-1. Partial correlations can be computed as shown in Table 8-4. The correlation coefficients marked with red have page 117 of 189

118 correlation coefficients greater than 0.5. This may give an indication of some of the relationships, which are difficult to see in figures 8-3 to Figure 8-11: Snow and ice profile in the Lincoln Sea multiyear ice 8.7 The simulated data (Arctic cases) The multiyear ice profile is initialized with a 2.5 m ice profile and a 10 cm snow layer with relatively coarse grains resembling summer snow. The snow depth at the end of the winter season is about 30 cm. page 118 of 189

119 Figure 8-12: The sea ice concentration for the Lincoln Sea multiyear ice profile After the initial phase with snow accumulation and metamorphosis all algorithms give near 100% ice concentration estimates as seen in Figure The initial 120 (30 days) iterations are excluded from the analysis. page 119 of 189

120 Figure 8-13: The snow - ice interface temperature in the Lincoln Sea multiyear ice profile Both the gradient and hybrid and OneChannel algorithms are sensitive to the snow ice interface temperature as seen in Figure The estimated ice concentration is increasing with increasing snow ice interface temperature. However, the polarization algorithms seem insensitive to the snow ice interface temperature. page 120 of 189

121 Figure 8-14: The sea ice concentration estimate from 9 different algorithms sensitivity to snow surface temperature in the Lincoln Sea multiyear ice profile The Bootstrap-F, the TUD, the Bristol and the OneChannel algorithms are sensitive to the snow surface temperature as seen in Figure The estimated ice concentration is increasing with temperature. page 121 of 189

122 Figure 8-15: The 9 sea ice concentration algorithms sensitivity to cloud liquid water in the Lincoln Sea multiyear ice profile It seems from Figure 8-15 that the gradient, the hybrid, the N90LIN and the OneChannel algorithms are sensitive to the cloud liquid water and show an increase of the estimated ice concentration with increasing cloud liquid water; this increase seems to be non-linear and is larger at low cloud liquid water levels than at high ones. The NASA and the Bootstrap-P ice concentration seem unaffected by increasing levels of cloud liquid water. page 122 of 189

123 Figure 8-16: The 9 sea ice concentration algorithms sensitivity to snow depth in the Lincoln Sea multiyear ice profile From Figure 8-16 it is difficult to see the estimated ice concentration sensitivity to snow depth. Other sensitivities dominate. page 123 of 189

124 Figure 8-17: The 9 sea ice concentration algorithms sensitivity to snow temperature gradient in the Lincoln Sea multiyear ice profile None of the algorithms are sensitive to the snow temperature gradient in the snow as seen in Figure page 124 of 189

125 Figure 8-18: The 9 sea ice concentration algorithms sensitivity to snow surface density in the Lincoln Sea multiyear ice The polarization algorithms at all frequencies i.e. NASA, Bootstrap-P, and N90LIN are sensitive to the snow surface density. The estimated ice concentration is decreasing with density. The ASI_NWF has some sensitivity to the snow surface density but it is much smaller than for the other polarization algorithms. Other sensitivities dominate for the hybrid, gradient and OneChannel algorithms. page 125 of 189

126 Figure 8-19: The 9 sea ice concentration algorithms sensitivity to atmospheric water vapor in the Lincoln Sea multiyear ice profile Atmospheric water vapor is correlated with e.g. the surface temperature and these two sensitivities are difficult to separate in Figure It is therefore difficult to estimate the sensitivity to atmospheric water vapor over ice in Figure However, the sensitivity to the surface temperature was a linear one, while here we tend to have a non-linear one - similar to the cloud liquid water. page 126 of 189

127 Figure 8-20: The 9 sea ice concentration algorithms sensitivity to average snow correlation length in the Lincoln Sea multiyear ice profile Other sensitivities dominate over the average snow correlation length in Figure Observed winter std Mean IC incl. atm. std incl. atm. std excl. atm. NASA Bristol Bootstrap - F Bootstrap - P TUD ASI N90LIN Mean atm/no atm IC ratio OneCannel - Tb6h OneCannel page 127 of 189

128 Tb19h Table 8-5: Shows the mean ice concentration, the standard deviation with or without atmosphere and the mean atmosphere/ no atmosphere ice concentration ratio for the 9 algorithms in the Lincoln Sea ice profile. The no atmosphere case is the surface emission only. The atmosphere included is the atmospheric emission absorption and reflection computed with a modified Wentz model in addition to the surface emission. This includes oxygen absorption Ice surface T Surface T Snow depth Surf density Vapor Liquid NASA Bristol Bootstrap - F Bootstrap - P TUD ASI N90LIN Table 8-6: Partial correlations: The partial correlation of each of the 6 parameters (snow ice interface temperature, surface temperature, snow depth, snow surface density, atmospheric water vapor, cloud liquid water) and the ice concentration with the effects of the other physical parameters (snow surface density, snow correlation length, snow surface temperature, snow ice interface temperature, snow depth, snow temperature gradient, atmospheric water vapor, cloud liquid water, excluding one of the 6) removed 8.8 Open Water The sensitivity of the different sea ice concentration algorithms over open water to wind roughening of the water surface, atmospheric water vapor, cloud liquid water, and sea surface temperature was tested using a Wentz model. The input data were extracted from the ECMWF ERA40 re-analysis from two selected grid points one in the Bellinghausen Sea in Antarctica and one in the Norwegian Sea in the Arctic. These two one year time series are representing the annual cycle of meteorological conditions near the ice edge. The target ice concentration is zero and in general the algorithms have small biases around zero except the ASI_NWF seen in Table 8-7. Algorithm Observed STDEV Mean STDEV NASA page 128 of 189

129 Bristol Bootstrap - F Bootstrap - P TUD ASI_NWF N90LIN Table 8-7: The 64 S 280 E sample MEAN and STDEV The Antarctic open water point at 64 S 280 E Table 8-8 is showing the correlation matrix of the four physical parameters and 7 selected sea ice concentration algorithms. Most of the algorithms are correlated with the four physical parameters. Strangely the TUD is not correlated and the Bootstrap-F is correlated. Figures 8-21 to 8-24 show the plots. The plots show that the correlation coefficient can falsely indicate sensitivity where there is none. page 129 of 189

130 Wind Vapor Liquid Ts NT Bristol BF BP TUD ASI N90 Wind Vapor Liquid Ts NT Bristol BF BP TUD ASI N90L 1.00 Table 8-8: The Bellinghausen Sea open water correlation matrix, r page 130 of 189

131 Table 8-9 shows the partial correlations between the sea ice concentration algorithms and the four meteorological parameters. Partial correlation measures the degree of association between two random variables, with the effect of a set of controlling random variables removed. None of the algorithms are correlated with the sea surface temperature. The NASA team is correlated with the three other meteorological parameters and the other algorithms are correlated with one or two parameters except the TUD which is not correlated with any of the meteorological parameters. Algorithm Parameter Wind Vapor Liquid Ts NASA Bristol Bootstrap-F Bootstrap-P TUD ASI_NWF N90LIN Table 8-9: The partial correlation, r, of each of the 4 input parameters to the Wentz model (The surface wind speed, atmospheric water vapor, the cloud liquid water, and the sea surface temperature) and the ice concentration with the effects of the other physical parameters removed (the one in question excluded) Figure 8-21 shows the sensitivity of the 9 different algorithms to cloud liquid water over open water at a point in the Bellinghausen Sea. Most of the algorithms are to some extent sensitive to cloud liquid water except the Bootstrap F the TUD and the OneChannel Tb6h. The Bootstrap P, ASI_NWF and N90LIN are very sensitive to cloud liquid water. page 131 of 189

132 Figure 8-21: The simulated sensitivity of the 9 sea ice concentration algorithms to cloud liquid water at 64 S 280 E over open water Figure 8-22 is showing the simulated sea ice concentration sensitivity to sea surface temperature over open water. It is difficult to judge if the 9 different ice concentration algorithms are sensitive to sea surface temperature. page 132 of 189

133 Figure 8-22: The simulated sensitivity of the 9 sea ice concentration algorithms to sea surface temperature at 64 S 280 E over open water Figure 8-23 shows the 9 sea ice concentration algorithms sensitivity to atmospheric water vapor. Most of the algorithms seem sensitive to water vapor over open water especially the Bootstrap P, the ASI_NWF and the N90LIN. page 133 of 189

134 Figure 8-23: The simulated sensitivity of the 9 sea ice concentration algorithms to atmospheric water vapor at 64 S 280 E over open water Figure 8-24 is showing the 9 sea ice concentration algorithms sensitivity to wind over open water. Some of the algorithms show sensitivity to wind e.g. the NASA and the Bristol, but it is difficult to judge because other noise sources seem to dominate. page 134 of 189

135 Figure 8-24: The simulated sensitivity of the 9 sea ice concentration algorithms to surface wind at 64 S 280 E over open water The Arctic open water point at 70 N 0 E The Arctic open water point at 70 N 0 E in the Norwegian Sea near Jan Mayen. Algorithm Wind Vapor Liquid Ts NASA Bristol Bootstrap F Bootstrap P TUD ASI_NWF N90LIN Table 8-10: The partial correlation, r, of each of the 4 input parameters to the Wentz model (The surface wind speed, atmospheric water vapor, the cloud liquid water, and the sea surface temperature) and the ice concentration with the effects of the other physical parameters removed (the one in question excluded) page 135 of 189

136 The atmospheric water vapor and the sea surface temperature, Ts, are correlated as seen in Table All of the algorithms are correlated to some extent. The ice concentration is not correlated to wind speed. In general the results from the Norwegian Sea open water point confirm the results from the Bellinghausen Sea open water point. Wind Vapor Liquid Ts NT Bristol BF BP TUD ASI N90L Wind Vapor Liquid Ts NT Bristol BF BP TUD ASI N90L 1.00 Table 8-11: Correlation matrix at 70 N 0 E, r, with the four input parameters to the Wentz model and the ice concentration from selected algorithms Algorithm Observed STDEV MEAN STDEV NASA Bristol Bootstrap - F Bootstrap - P TUD ASI_NWF N90LIN page 136 of 189

137 Table 8-12: The mean, MEAN, and standard deviation, STDEV, of the simulated data at 70 N 0 E Figure 8-25: The simulated sensitivity of the 9 sea ice concentration algorithms to cloud liquid water at 70 N 0 E over open water page 137 of 189

138 Figure 8-26: The simulated sensitivity of the 9 sea ice concentration algorithms to sea surface temperature at 70 N 0 E over open water page 138 of 189

139 Figure 8-27: The simulated sensitivity of the 9 sea ice concentration algorithms to atmospheric water vapor at 70 N 0 E over open water page 139 of 189

140 Figure 8-28: The simulated sensitivity of the 9 sea ice concentration algorithms to wind at 70 N 0 E over open water page 140 of 189

141 9 Weather filters 9.1 Description of weather filters To deal with weather effects many algorithms use weather filters or atmospheric correction in form of radiative transfer model with input from numerical weather prediction models or standard atmosphere profiles. There are two main weather filters: 1. Gloersen and Cavalieri (1986) filter for SMMR and Cavalieri et al. (1995) for SSM/I: SMMR : C 0 if GR (9-1) SSM /I : C 0 if GR and /or GR (9-2) This filter was developed for the NASA Team algorithm and can be used with other algorithms (e.g. NRL, PR, Two-channel) that do not have weather filters originally. 2. In Comiso and Sullivan (1986) filter, the data points in the TB19V TB22V scatter plot that are below the open-water filter line (see dotted line in figure 3) are set to 0 %. An additional mask is applied to deal with extreme atmospheric conditions. Data above a certain threshold for the difference of TB22V and TB19V is also set to 0 %. This filter is used in the Comiso Bootstrap algorithm. In some algorithms (like one-channel) atmospheric effects are dealt with through adjustment of tie point signatures which allows incorporation of average atmospheric effects. For example the water tie point is the apparent temperature of the water surface observed through an average atmosphere (Pedersen, 1991; Kaleschke et al., 2001). Others use more complex approach, like ECICE algorithm correction of 85 GHz brightness temperature (Shokr et al., 2008). Influence of integrated water vapour W and cloud liquid water content L is subtracted from the original brightness temperature using n 4 TB corrwl TB C k a ji V W i L j k 1 i, j 0 k (9-3) where C k is concentration of each ice type k, a ji - coefficients that depend on the surface emissivity which in turn is a function of wind speed V. The term between the brackets in the right-hand side of the above equation is a fourth-order polynomial. Correction for the influence of wind speed over ocean surface is carried out using the following equation: page 141 of 189

142 TB corrwlv TB corrwl C OW b i V i (9-4) where the coefficients b i are obtained using regression of results from the microwave radiative transfer model obtained at fixed of W, L, and V. Other algorithms account for atmospheric influence by using more complete version of the radiative transfer equation. This is the case for NORSEX, Bristol, NASA Team 2, Near 90GHz, UMass-AES. In order to test the algorithms performance with weather filter we used the first filter (Cavalieri et al., 1995), see Eq Results for selected algorithms 4 i Summer, SIC = 15% AMSR SSM/I Algorithm Without WF With WF Without WF With WF Northern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team ECICE Southern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team ECICE Table 9-1: SIC = 15%, summer. Standard deviations of concentrations with and without weather filters page 142 of 189

143 AMSR SSM/I Algorithm Without WF With WF Without WF With WF Northern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team ECICE Southern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team ECICE Table 9-2: SIC = 15%, summer. Average concentrations with and without weather filters Winter, SIC = 15% AMSR SSM/I Algorithm Without WF With WF Without WF With WF Northern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team page 143 of 189

144 AMSR SSM/I Algorithm Without WF With WF Without WF With WF ECICE Southern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team ECICE Table 9-3: SIC = 15%, winter. Standard deviations of concentrations with and without weather filters AMSR SSM/I Algorithm Without WF With WF Without WF With WF Northern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team ECICE Southern Hemisphere Near 90GHz lin, dyn ASI Bootstrap P Bristol NASA Team Bootstrap F OSISAF NASA Team page 144 of 189

145 AMSR SSM/I Algorithm Without WF With WF Without WF With WF ECICE Table 9-4: SIC = 15%, winter. Average concentrations with and without weather filters 9.3 Simulated data (15, 20, 25, 30% ice and cut-off) page 145 of 189

146 Figure 9-1: Illustration of weather filter performance with AMSR-E data from 2008 Northern hemisphere. X-axis is GR1, and y-axix is GR2 from equation 9-2 The data used in this are artificial data created using the same mechanism as described under the SIC15 tests in chapter 3 (equation 3-1). The figure shows how the weather filters remove some ice up to 25-30%. This effect is considered undesirable and we therefore investigate next (chapter 10) how atmospheric correction of TBs using our Radiative transfer and surface emissivity model can reduce ocean surface and atmospheric effects. In this chapter we investigated whether and how much the weather filters cut-off sea ice concentrations - even though they might be real. For this purpose we computed GR1 and GR2 (see Eq. 9-2) from AMSR-E brightness temperatures created artificially using the same mechanism as described under the SIC15 tests in chapter 3 (equation 3-1). We used cases of SIC = 15%, 20%, 25% and 30%. Figure 9-1 illustrates which GR1 - GR2 data pairs that belong to the respective SIC test case remain untouched (shaded box) and which are filtered out by the weather filter. page 146 of 189