Error assessment of sea surface temperature satellite data relative to in situ data: effect of spatial and temporal coverage. Aida Alvera-Azcárate, Alexander Barth, Charles Troupin, Jean-Marie Beckers GHER University de Liège FNRS National Fund for the Scientific Research, Belgium
Objectives Establish error of satellite data respect to in situ data different in situ sensors/platforms Night-time / day-time SST? Merging of satellite SST with several in situ data sources Should we use all data available? (or limit to a predefined time interval)
In situ data Domain chosen: Mediterranean Sea World Ocean Database 2005 (WOD05, http://www.nodc.noaa.gov/) MEDAR/MedAtlas (MEDAR-Group (2002), http://www.ifremer.fr/medar/) Coriolis Data Center (http://www.coriolis.eu.org/ International Council for the Exploration of the Sea (ICES, http:// www.ices.dk/). Year chosen: 1999 (higher number of data) Tests: Check for doubles Depth 5 m Total number of data: 3775
Spatial and temporal distribution In situ data Sensors and number of data CTD 320 XBT 1043 floats/drifters 1994 low-resolution CTD 260 BATHY 141 TESAC 13 Unknown 4
Satellite data AVHRR SST data (http://podaac.jpl.nasa.gov) day-time and night-time passes ~5 km spatial resolution Interpolated to in situ positions (both day-time and night-time passes) From 3775 in situ data, 2552 match-ups 50% night-time satellite data 50% day-time satellite data Day-time SST on 11 April 1999
#observations / month Data statistics (I) #observations / hour of day Monthly mean
Data statistics (II) Temperature distributions Temperature distribution by sensor Monthly distribution by sensor
Satellite data: day-night difference Data statistics (III) Satellite in situ monthly difference
Error assessment of satellite vs. in situ data ( C) Bias RMS r r anomaly Day-time 0.33 0.73 0.98 0.84 Night-time 0.17 0.75 0.98 0.84
Error assessment of satellite vs. in situ data Comparison by month Day-time SST vs. in situ Night-time SST vs. in situ Reference: satellite data
Error assessment of satellite vs. in situ data Comparison by sensor Day-time SST vs. in situ Night-time SST vs. in situ Reference: satellite data
Error assessment of satellite vs. in situ data Summary of satellite-in situ comparison Homogeneous day/night coverage of in situ data Heterogeneous spatial/temporal (monthly) in situ data distribution Drifter data show a higher bias (warm) CTD data present higher RMS error Day-time SST data: higher errors in summer months Diurnal cycle present in satellite data, but not in satellite-in situ data difference
Comparison between satellite and in situ data Night-time SST (October average) Analysis made through Diva-on-web (http://gher-diva.phys.ulg.ac.be) Day-time SST (October average) Inverse variational analysis Correlation length: 1 Signal-to-noise ratio: 0.98
Comparison between satellite and in situ data All XBT data September 1999 Night-time SST (September average) Day-time SST (September average)
Comparison between satellite and in situ data 13 January 12 October
DINEOF analysis of in situ and satellite data DINEOF (Data Interpolating Empirical Orthogonal Functions): parameter-free, EOF-based technique to reconstruct missing data Determines optimal number of EOFs by cross-validation Univariate and multrivariate analyses Extension to merge different data sources Near-real time reconstruction of the Western Mediterranean SST http://gher-diva.phys.ulg.ac.be/dineof
1st: Demeaned matrix: missing data flagged and set to zero DINEOF Some data are set aside for cross-validation 2nd: EOF decomposition with N=1 EOF Calculate missing values: Improved guess for missing values X k T i, j = ρp ( up ) i ( vp ) j p= 1 Convergence: best value for missing data with 1 EOF cross validation: error EOF decomposition with N=2 EOFs Calculate missing values error Improved guess for missing values Then we repeat with N= 3 EOFs N EOFs and so on
Multivariate DINEOF: examples SST + HF-Radar currents C m/s Radial currents: positive values: current towards the antenna negative values: current away from the antenna
Error maps and outlier detection Error maps are calculated for TSM reconstructions using: i) the EOF basis from DINEOF as background covariance ii) the location of valid data Outliers (pixels with value larger than the statistically expected misfit calculated during the analysis) will be objectively identified and removed from initial data
A tough case When too few data are present: temporal EOFs poorly constrained: unrealistic discontinuities Sharp transition
Temporal covariance matrix filter C=F T CF F is a Laplacian filter Filter on C instead of X: C is much smaller and less sensitive to missing data Filter applied iteratively: more iterations, further reach of the filter Alvera-Azcárate et al, 2009
Tough case (resolved!)
Conclusions DINEOF to fill missing information in satellite data sets Univariate and monovariate reconstructions Error maps and outliers Free available code: http://modb.oce.ulg.ac.be/mediawiki/index.php/dineof Extension to incorporate in situ data Comparison of satellite and in situ data: dependent of platform/sensor Biases and RMS errors need to be taken in to account before merging