www.nr.no Advancements and validation of the global CryoClim snow cover extent product Rune Solberg1, Øystein Rudjord1, Arnt-Børre Salberg1 and Mari Anne Killie2 1) Norwegian Computing Center (NR), P.O. Box 114 Blindern, NO-0314 Oslo, Norway 2) Norwegian Meteorological Institute (MET Norway), P.O. Box 43 Blindern, NO-0313 Oslo
The CryoClim service The CryoClim project (2008-2013) initiated by the Norwegian Space Centre (NSC) and administrated by ESA ESTEC under the PRODEX programme developed algorithms, products and a service for cryospheric climate monitoring Operational service from November 2013: www.cryoclim.net Products: Snow cover (MET/NR): 1982-present, global Snow cover extent (SCE) Sea ice (MET): 1979-present, global Sea ice concentration (SIC) Sea ice edge (SIE) Glaciers Norway (NVE): 1952/1988 present Glacier area outline (GAO) Glacier lake outline (GLO) Glacier lake outburst flood (GLOF) Glacier periodic photo series (GPP) Glaciers Svalbard (NPI): 1936/1992 present Glacier area outline (GAO) Glacier surface type (GST) Extended with Greenland in 2014 (GEUS): 2000-present Glacier surface type (GST)
Sub-service snow Snow Cover Extent (SCE) (snow/no snow) Developed three competing prototype products: SCE from PMR (10 km) Based on SMMR (1978-1987) and SSM/I (1987- present) SCE from optical (5 km) Based on AVHRR GAC (1982-present) SCE multi-sensor/temporal (5 km) Combination of optical and PMR Final product: Multi product global time series 1982 present Aggregation levels: Day, month, year Projection/files: EASE-Grid, NetCDF CF, Northern & Southern Hemisphere Multi-sensor multi-temporal snow cover 1 March 2005 Climate-change indicator products: Snow season length, first and last day of snow
CryoClim Snow Phase 2 objectives 1. Mitigate weaknesses in the single sensor components of the algorithm (optical and passive microwave radiometers) and multi-sensor/multi-temporal data fusion to further increase the accuracy and robustness of the product. 2. Extend the product with uncertainty estimates at the product and per-pixel levels. 3. Advance the algorithms and processing chains with the inclusion of Sentinel-3 OLSI and SLSTR data. 4. Perform more extensive validation of the product in space and time, including focus on inter-sensor issues in the time series. 5. Include the results in the CryoClim processing chain for snow and advance the operational level of the processing. Project period 2015-2016 4
The PMR SCE algorithm is based on an estimate of the probability of snow P( S k x 1, x 2,, x n )= P x 1 S k P x 2 S k P( x n S k ) P( S k )/ m=1 C P x 1 S m P x 2 S m P( x n S m ) P( S m ) SMMR Snow classes: Snow & no snow Features: x1=t18v-t37v x2=t18h-t37h SSM/I Snow classes: Dry snow, wet snow, no snow & no snow with a large portion of water Features: x1=t37v-t37h x3=t22v-t85v x2=t19v-t37v x5=t22v x4=(1.95 T19v-0.95 T19h)/0.95 SMMR 1978 1987 SSM/I F8 1991 SSM/I F11 1999 SSM/I F15
The optical AVHRR GAC SCE algorithm is based on an estimate of probability of snow P( S k x 1, x 2,, x n )= P x 1 S k P x 2 S k P( x n S k ) P( S k )/ m=1 C P x 1 S m P x 2 S m P( x n S m ) P( S m ) prob. for snow prob. for cloud prob. for bare ground
A state model based on fusion of single-sensor state models Optical Multi-sensor PMR
Implemented the model applying the Hidden Markov Model framework In HMM we observe a system assumed to evolve through a series of different states Transitions from one state to another happen with certain probabilities While in a given state the system will produce observables with a certain probability density States: Observables: Prob. distr.: Q= { S 1, S 2,, S v } X T ={ X 1, X 2,, X T } p X t E t = S i, i=1, 2,,v Transition probabilities.: p E t = S i E t 1 = S j, i,j=1, 2,,v Initial conditions: p( E 1 = S i ), i=1, 2,,v
Note that a there is one state model per grid cell p X t E t = S i p E t = S i E t 1 = S j
Estimating the probabilities Pr(S) p(snow O ) 0 p(snow O ) 1 p(bare ground snow) 0 p(snow O ) 0.5 Time Climatological probability of snow Per grid cell daily climatological probability of snow computed from Savitzky-Golay smoothed PMR snow probabilities Used to estimate transition probabilities
Using the Viterbi algorithm to determine the model sequence best explaining the temporal observations The Viterbi algorithm is a dynamic-programming algorithm for finding the most likely sequence of hidden states (the Viterbi path) that results in a sequence of the observables The algorithm requires as input the state probability density functions, the transition probabilities between the different states and the initial probability of each state V 1,k =p X 1 k p( E 1 = S k ) V t,k =p X t k max i (p E t = S i E t 1 = S j V t 1,k ) Final state model chosen
Examples from snow season 2005 1 February 2005 1 May 2005 1 March 2005 1 July 2005
Mitigation of algorithm weaknesses and other errors Some known problems with the products: To much snow in the summer in mountain regions in Asia (in particular Tibet) False snow in Sahara (currently with a preliminary fix) Systematic lack of snow along some (complex) shorelines Sometimes processing errors create artefacts Mitigation work is currently ongoing 13
Artificial snow Too much snow Himalaya and surrounding regions in summer: Coming from the PMR Related to thin snow layer and extremely dry soil, maybe also permafrost? Originally snow was often detected in Sahara: From PMR (and clouds in optical?) Temporarily fixed by setting a surface temp. threshold, with PMR at T=20 C Works most of the time, but not always Himalaya, 15 July 2005 North Africa, 17 January 1995 14
Some permanently snow-free shorelines Systematic lack of snow in some coastal regions Enlargement Typical winter 15
Processing errors Occasional irregular blocks of false snow In both hemispheres Time discontinuities around July/August Seem to be related to model start for a new snow season North America, 30 Aug 2007 South America, 1 January 2001 Asia, 24 July 2002 Asia, 25 July 2002 16
Optical: Solar angle influences the PDFs No snow High probability for snow The spectral signatures applied are relatively independent of solar angle, but change character for low solar angles Current version: static PDFs. Can give false snow cover for low solar angles OSI SAF SST work
Product uncertainty Direct modelling through e.g. error propagation too complex with the HMM method Currently investigating alternative approaches, including bootstrapping Bootstrap estimates on the individual bootstrap datasets θ ( ) = 1/B b=1 B θ (b) where θ (b) is the estimate on bootstrap sample b. E.g. the bootstrap estimate of the variance is var boot (θ)= 1/B b=1 B [ θ (b) θ ( ) ] 2 18
Validation Datasets for validation: Snow depth from the Global Historical Climatology Network Daily (GHCN-D) SYNOP database applied in Phase 1 of product development Daily snow depth observations, including recent years Historical Soviet Daily Snow Dataset (HSDSD) Daily measurements of snow depth and snow cover from meteorological stations in former Soviet Union Former Soviet Union Hydrological Snow Surveys (FSUHSS) Snow transects from meteorological stations in former Soviet Union Ideally, a third new dataset covering more recent years, and area outside former Soviet Union, e.g.: GHCN-D NCDC U.S. Daily Snow Depth Data NRCS SNOTEL 19
Global Historical Climatology Network Daily (GHCN-D) Data set applied in the first phase of snow product development Daily snow depth observations from 2005 Filtering out obvious errors: Like stations with suspicious behaviour related to that zero snow depth not reported explicitly Validation results (2005): Very high accuracy in summer Somewhat lower accuracy in November-January Slightly lower accuracy in April Month Accuracy Number of samples January 0.82 1298 February 0.90 1318 March 0.94 1385 April 0.88 1008 May 0.93 1254 June 0.99 1437 July 1.00 1488 August 1.00 1488 September 1.00 1427 October 0.98 1225 November 0.78 702 December 0.76 1170 Total 0.924 15200 20
Historical Soviet Daily Snow Dataset (HSDSD) Around 280 stations in former Soviet Union Daily snow depth and snow cover measurements Until 1995 (currently 3.5 years of overlap) Filtering and assumptions: Using only stations with coordinates in GHCND data (more accurate). Consider less than 50% snow cover as bare ground. Remove anything slightly suspicious. All fields flagged as humidity measurements are assumed to be snow free. Validation results for the period 1 Aug 1992 31 Dec 1995: High accuracy in summer Lower accuracy in October-November Slightly lower accuracy in April Month Accuracy Number of samples January 0.92 10876 February 0.94 9810 March 0.93 10230 April 0.90 9452 May 0.92 10006 June 0.99 9592 July 1.00 10602 August 0.99 13940 September 0.97 13447 October 0.85 13375 November 0.84 12652 December 0.92 13475 Total 0.928 138021 21
Former Soviet Union Hydrological Snow Surveys (FSUHSS) Large number of snow transects in former Soviet Union measured until 1996 (1345 stations in total, not all in recent years) 0.5 2 km long transects, usually three per month Several parameters measured, including snow cover and snow depth. Filtering: Using only stations with coordinates in GHCND data (more accurate). Contains no observations of bare ground! Validation results for the period 1 Aug 1992 31 Dec 1996: High accuracy in winter Lower accuracy in spring and autumn No data in summer Month Accuracy Number of samples January 0.98 3067 February 0.99 3924 March 0.99 6084 April 0.90 4198 May 0.77 753 June 0.00 7 July 0.00 0 August 0.00 0 September 0.36 58 October 0.61 1232 November 0.89 2990 December 0.95 3089 Total 0.930 25431 22
Conclusions and way forward Status: Overall validation results from Phase 1 using GHCN-D confirmed in Phase 2 with HSDSD and FSUHSS data Overall accuracy 93% Transition periods with large areas of both wet and dry snow and thin snow layer in autumn reduces the accuracy somewhat ( 85-90%) Some weaknesses of false snow, lack of snow and processing errors discovered in Version 1.0 Way forward: Processing errors being investigated and fixed now Ways to mitigate false snow and lack of snow under investigation Pixel-wise uncertainty estimate under development Porting optical algorithm to Sentinel-3 will start next year Validation data set will be extended with more recent observations (e.g. GHCN-D) for studying a longer period (sensor degradation and inter-sensor issues) cryoclim@cryoclim.net www.cryoclim.net 23