Presentation is loading. Please wait.

Presentation is loading. Please wait.

EarthTemp Arctic SST Workshop, MetOffice, Exeter, UK, 18.December 2013 Ice and Cloud masking in the Arctic Steinar Eastwood, Norwegian Meteorological Institute.

Similar presentations


Presentation on theme: "EarthTemp Arctic SST Workshop, MetOffice, Exeter, UK, 18.December 2013 Ice and Cloud masking in the Arctic Steinar Eastwood, Norwegian Meteorological Institute."— Presentation transcript:

1 EarthTemp Arctic SST Workshop, MetOffice, Exeter, UK, 18.December 2013 Ice and Cloud masking in the Arctic Steinar Eastwood, Norwegian Meteorological Institute (MET) Claire Bulgin (UoE), Sonia Pere and Pierre LeBorgne (M-F), Adam Dybbroe (SMHI)

2 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Outline ·Existing cloud/ice classification algorithms ·Probabilistic approach and PDFs ·Collection of training data ·Adapting existing PDFs to new sensors ·Twilight conditions AATSR, South-east coast of Greenland

3 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Conventional classification algorithms ·General cloud/ice/water classification algorithms often wants to provide an unbiased cloud classification ·Hence some clouds are undetected ·SST retrieval would like a more certain/conservative classification ·Need additional cloud and ice masking step after applying conventional cloud mask ·Especially for sea ice since this is often less accurately handled by general cloud classification algorithms ·Also need particulare focus on twilight conditions since more frequent at high latitudes

4 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Probabilistic approach ·Method is based on probability distribution functions (PDF) for channel features, often combined using Bayes theorem ·Gives the probability of the satellite data being cloudy, water and ice given the observations ·Provides flexible method, the additional masking can be tuned on the probabilities ·Need classified data to define/train PDFs

5 Classification example 2009-04-14 AATSR

6 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Ways of collection training data ·Good/representative PDFs depend on good training data ·Three methods: - Manual classification (eg. MET Norway) - Automatic classification with ground truth, such as synop, cloud satellite missions and sea ice products (eg. SMHI) - Using existing cloud mask as «truth», such as PPS, Maia, SADIS, CLAVR-X (eg. Chris, Owen, Claire)

7 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Automatic collection of training data with Calipso cloud lidar on Caliop ·Looking at Calipso cloud lidar data collocated with VIIRS data and ice concentration for collection of training data (cooperation with SMHI) ·Using VIIRS orbits which Caliop crosses close in time ·Use Calipso to define/classify cloudy and cloud free areas, even over ice ·Can then define the PDFs for cloud, ice and water ·Still need to separate ice from water, as Calipso only sees clouds

8 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Using PPS as «truth» ·Have used one year of collocated VIIRS data and PPS cloud mask ·Use PPS to define areas with clouds, clear sea, clear ice, clear land and clear snow ·At night, PPS only classify cloud and clear ·So apply simple test to split ice and sea: - ice if clear & T11 < 270 - water if clear & T11 > 275

9 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Using PPS as «truth» ·Build count matrix (multi-dim histogram) by looping through all scenes - Calculate the channel feature(s) we want (eg. 0.9/0.6, std(0.6)) - Summarize how many pixels in each class in defined intervals in these feature(s) - Save this matrix in NetCDF file ·Have made Python tool to visulize these histograms in 3D, 2D and 1D ·Can also compare matrixes from two different satellites

10 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Training data: Calipso vs PPS ·Comparing 2D histograms of training data ·Both are one year of locally received data at SMHI, Sweden ·Calipso in red ·PPS in blue ·r0.9um/r0.6um vs sunz

11 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute 3D view of nighttime matrix Sea Ice Water Cloud

12 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute How to define the PDFs ·Parameterize histograms as analytic probability functions ·Typically use Gaussian normal distribution ·Define look-up tables from training data, 1D or multi- dimensional ·Needs lots of data for multi-dim

13 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Defining PDFs for new instruments ·How to adapt existing coefficients / method to new instruments ·Options: - Repeat manual classification training – expensive... - Compare overlapping orbits from new and old instrument - Simulate difference between existing and new instruments ·Some examples...

14 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute New PDFs for NPP VIIRS ·The OSI SAF method had coefficients for AVHRRs on NOAA17,18,19 and METOP-A ·We needed to updated the OSI SAF method to work on VIIRS data (during day) ·Had no time to do an extended training of method by manual classification ·Wanted to adjust the existing N19 AVHRR coefficients to VIIRS ·Comparing N19 and NPP orbits close in time (+/- 0.5 hour)

15 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute NPP VIIRS N19 AVHRR 02.06.2012 r0.9um/r0.6um

16 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute NPP VIIRS PDFs ·Use the limited set of collocated VIIRS + AVHRR data to find average shift between PDFs for this limited set ·Apply this average shift to define overall VIIRS PDFs: PDF VIIRS = PDF AVHRR - shift(AVHRR- VIIRS) ·VIIRS chain at CMS/MF run with these new coefficients and work fine

17 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute VIIRS PDFs defined from Calipso, manual and PPS training r0.9um/ r0.6um PDF mean Shift N19CalipsoPPS Water0.490.500.52 Ice0.850.830.88 Cloud0.960.951.00

18 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Simulating ·ref Chris and Owen...

19 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Twilight ·Some of the features/channel combinations we use deviates from the normal day light behaviour when approaching twilight

20 www.esa-sst-cci.org AVHRR 0.8um/0.6um, sea, normalized

21 www.esa-sst-cci.org AVHRR 0.8um/0.6um, sea norm + PDF Sunz < 80: PDFmean = MB PDFstd = SB Sunz 80-90: PDFmean = MA*SOZ + MB PDFstd = SA*SOZ + SB

22 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Simulating twilight ·Have just started to simulate visible channels using libradtran ·Example with r0.9/r0.6 for water and snow

23 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Arctic Challenges ·How to train PDFs to cope with difficult conditions - Twilight - Ice vs clouds and wet ice vs cold water during night ·How thick is the ice before we can distinguish it from water ·Find good validation data for comparing algorithms under these challenging conditions ·How does different performance of cloud/ice mask at day / twilight / night influence SST accuracies? ·Should we try to see how SST accuracies varies with cloud masking skill from PDFs?

24 Questions?

25 Extra slides

26 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute From Metop02 to Metop01 ·The AVHRR on Metop02 and Metop01 should be very similar ·Will building channel count matrixes for both satellites with 6 months of Arctic data ·Classification is based on Maia cloud mask and OSI SAF sea ice concentration information

27 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute From Metop02 to Metop01 r0.9um/r0.6um June 2013 1 – Metop01 2 – Metop02

28 EarthTemp Arctic SST workshop, 18-19. December 2013 Norwegian Meteorological Institute Additional 0.6um test ·Some ice were still classified as open water, espesially melting ice with low 0.9um/0.6um values ·Sonia and Pierre suggested an additional test using 0.6um to mask undetected ice ·Defined static threshold test ·Could introduce PDF

29 ESA Climate Change Initiative Phase 1 Sea Surface Temperature (SST) www.esa-sst-cci.org Improving cloud and ice masking algorithm at high latitudes during ESA CCI on SST

30 www.esa-sst-cci.org Cloud and ice masking in ESA CCI SST Training of method for cloud and ice masking has been based on manual classification This has been done by visual inspection and classification of areas within satellite scenes ICE WATER CLOUD

31 www.esa-sst-cci.org Cloud and ice masking in ESA CCI SST Have collected scenes close to sea ice border from different satellites, months, day/night/twilight Also use this data set for validation This data set will be published for all interested to access Using GAC data 1/4: ice, 2: clouds, 3: water

32 www.esa-sst-cci.org Update of algorithm Want to improve masking during twilight and night Have looked at dependency on sun zenith angle for the PDFs Using existing cloud mask (CLAVR-X) from the HL MMD to study cloud signatures for all seasons, in areas close to the ice edge

33 www.esa-sst-cci.org Daytime, N19, r0.9/r0.6, cloud

34 www.esa-sst-cci.org Daytime, N19, r0.9/r0.6, cloud

35 www.esa-sst-cci.org VIIRS 0.8um/0.6um, sea

36 www.esa-sst-cci.org VIIRS 0.8um/0.6um, sea, normalized

37 www.esa-sst-cci.org VIIRS 0.8um/0.6um, sea norm + PDF Sunz < 80: PDFmean = MB PDFstd = SB Sunz 80-90: PDFmean = MA*SOZ + MB PDFstd = SA*SOZ + SB

38 www.esa-sst-cci.org VIIRS 0.8um/0.6um, cloud, normalized

39 www.esa-sst-cci.org VIIRS 0.8um/0.6um, sea ice, normalized Fewer data...

40 www.esa-sst-cci.org Daytime bt3.7-bt11 PDFmean = MA*SOZ + MB PDFstd = SA*SOZ + SB

41 www.esa-sst-cci.org Nighttime VIIRS, bt3.7-bt12 vs bt11-bt12

42 www.esa-sst-cci.org Nighttime VIIRS, distance to WTP Line defining water tiepoints Distance to water tiepoint

43 www.esa-sst-cci.org Histogram of dist(3.7-12) for N18,night

44 www.esa-sst-cci.org Nighttime VIIRS, bt3.7-bt12 vs bt11-bt8.5

45 www.esa-sst-cci.org Validation night time

46 www.esa-sst-cci.org Validation night time

47 www.esa-sst-cci.org Day time algorithms uses: r0.9/r0.6 and r1.6/r0.6 or r0.9/r0.6 and bt3.7-bt11 Must add r0.6... Night time algorithm uses: dist(bt3.7-bt11, bt11-bt12) LSD(bt3.7-bt11), log-normal distribution


Download ppt "EarthTemp Arctic SST Workshop, MetOffice, Exeter, UK, 18.December 2013 Ice and Cloud masking in the Arctic Steinar Eastwood, Norwegian Meteorological Institute."

Similar presentations


Ads by Google