Uncertainties in ocean colour remote sensing

Transcription

1 ENMAP Summer School on Remote Sensing Data Analysis Uncertainties in ocean colour remote sensing Roland Doerffer Retired from Helmholtz Zentrum Geesthacht Institute of Coastal Research Now: Brockmann Consult

2 What determines the radiance spectrum at TOA TOA Radiance Spectra Radiance (Wm-2 sr-1 µm-1) North Sea wavelength (nm) Pin 1 Pin 2 Pin 3 Pin 4 Pin 5 Pin 6 air molecules different aerosols thin clouds contrails Sky reflectance Sun glint foam floating material Suspended particles dissolved organic matter Vertical distribution chlorophyll Bottom reflection different phytoplankton species

3 Accuracy and Precision bias Reliability Accuracy Precision Stability Reproducibility RMS error Bias Linearity Scatter Standard deviation

4 Calibration of the Space Sensor Overview On board calibration Vicarious calibration Radiometric calibration Spectral calibration Dark signal Linearity stability Requirement for radiometric calibration very high (~ 1% accuracy) Long term performance and stability has to be monitored

5 Calibration of MERIS

6 Long term gain development of MERIS Camera 1 Camera 4 Camera 2 Camera 5 Camera 3 mean per camera optics degradation for all bands. Cameras 1 to 5 from top to bottom.

7 Atmosphere and water surface Large variety of aerosols Use of selected aerosol types (models) Thin clouds (cirrus, contrails) Sub-pixel clouds Cloud shadows Absorption by atmospheric gases: water vapour, oxygen, ozone, NO2 Foam and white caps Waves Sky and sun glint

8 Radiances at Top of Atmosphere (TOA) W m-2 sr-1 µm Laer Lray Lwat wavelength nm

9 MERIS Top of atmosphere radiance reflectance RLtoa RGB New York

10 Path radiance+ Fresnel reflectance RLpath MERIS band 5 (560 nm)

11 Water leaving radiance reflectance RLw MERIS band 5 (560 nm)

12 Water leaving radiance reflectance RLw MERIS band 2 (443 nm)

13 Chlorophyll

14 Water Nature of water Pure water optical properties only partly known with required accuracy Temperature and salinity effects Many different water constituents with different and varying inherent optical properties Vertical distribution not homogenous Sub pixel patchiness Model of water Definition of a bio-optical model Optical components and their similarity and variability Methods to separate different components

15 Pure water absorption II courtesy of R. Roettgers, HZG

16 Impact of salinity on reflectance spectrum Water leaving radiance reflectance of oligotrophic water for temperature 15 deg C, salinity 0 and 35, chl 0.1 mg m-3, ys(440 nm) 0.01, SPM 0.01 g m-3.

17 Water constituents Variable optical properties How to define a bio-optical model How to measure IOPs separately for different constituents Relationship between IOPs and concentrations Influence of the vertical distribution Saturation and masking effects

18 Case 1 water algorithm based on reflectance ratio model R445 / R555 Chl =a[r(445) / R(555)] b Case 1 water: Morel / Antoine MERIS Case 1 water ATBD

19 Phytoplankton Photos by Marion Rademaker

20 Suspended Matter and Phytoplankton in Coastal Water Photos by K. Heymann

21 Uncertainties due to the bio-optical model Optically relevant variables in the water Optical components with IOPs and variability Proxy concentration variable Scattering material Absorbing material Phytoplankton pigments Scattering by all particles Absorption by organic material Absorption by phyto pigments Dry weight of suspended matter [ g m-3] Concentration of DOC / POC [mg m-3] Concentration of chlorophyll a [mg m-3]

22 Uncertainties due to variability of optical properties Heincke187 a443 pigment n= a (m -1 )) norm chlorophyll a [mg/l] wavelength (nm) Normalized absorption spectra of North Sea phytoplankton Summer period a(443) (m -1 )) Variability in the relationship between a_pig and chlorophyll concentration Conversions: Chl. a [mg m-3] = 21 * a_pig_442 ^1.04

23 Sensitivity at different concentration ranges and spectral bands RLw for MERIS bands 1 (412 nm), 6 (560 nm), 10 (708 nm) 0.03 RLw [sr -1 ] band 1 = 412 nm band 5 = 560 nm band 9 = 708 nm Sensitivity of the reflectance at a spectral band depends on the concentration To cover a large concentration range many bands from the blue to NIR range are necessary a_gelb_440: 0.2, a_part_440: SPM/25, pig: 2 mg m SPM [g m -3 ]

24 Sensitivity at different concentration ranges and spectral bands 9.00E-003 Remote Sensing reflectance TSM E-002 Remote Sensing reflectance TSM E E E E E-002 Rsr (sr-1) 5.00E E E-003 Rsr 5_01_1 Rsr 10_01_1 Rsr (sr-1) 3.00E E-002 Rsr 5_01_100 Rsr 10_01_ E E E E wavelength (nm) 0.00E wavelength (nm) Chl. 5/10 mg m-3 TSM 1 g m-3 ays(443) 0.1 m-1 Chl. 5/10 mg m-3 TSM 100 g m-3 ays(443) 0.1 m-1

25 Signal depth at different spectral bands Multiband algorithms: the information for each band may come from a different water layer water depth m turbid coastel ocean Signal depth z90 z 90 =1/k coastal: TSM=5 mg/l Chlor.=5µg/l Gelb=a 380 =1m -1 open ocean: Chlor.=1µg/l -50 pure water wavelength µm

26 The inverse problem reflectance spectrum inverse model pure water phytoplankton suspended matter dissolved org. matter sun zenith view zenith azimuth diff forward model sun zenith view zenith azimuth diff Matrix inversion Inversion by optimization Inversion by neural network Success depends on: Bio-optical model ambiguities

27 How to reduce / determine uncertainties Exclude pixels with doubtful results -> flagging Check if spectrum is within the scope of your algorithm Comparison with in situ data (global, per water type) determine conficence range from model results

28 Searching for minimum: principle, 1D case Search for minimum: Deviation between measured and simulated spectrum deviation Acceptance level deviation Acceptance level Independent variable Independent variable Width can be estimated from the 2nd order derivative (Hessean matrix)

29 Error due to masking and ambiguities simulated reflectance test error 0.01 stdev RLw [sr-1] wavelength [nm] rlw_true rlw_err rlw_retr True spectrum simulated measured spectrum = true * random error Retrieved spectrum when LM has found solution

30 Results and errors of retrieval Variable conc true conc retr. stdev of log_conc err. estimated % err true % chlorophyll [mg m-3] detritus [g m-3] gelbstoff a443 [m-1] min. SPM [g m-3] kdmin_true: kdmin_ret: error: % kd490_true: kd490_ret: error: -1.04%

31 Uncertainties due to ambiguities for different concentration mixtures 10 2 case 2 water chlorophyll retrieval with NN nn-derived µg/l model chlorophyll µg/l All cases of turbid water 2 mg m-3

32 Ambiguities Pigment in North Sea Water (gelb < 0.2 m-1, MSM < 5 mg/l) nn µg/l model µg/l Typical North Sea coastal water: ay_443: < 0.2 m-1, TSM < 5 mg /l

33 Detection of out of scope conditions (MERIS processor) r g Reflectance spectrum r g i nvnn r, r log of reflectances c log of concentrations g geometry information q quality indicator c IOPs or concentrations f or wnn r NN input NN output q c q c Comparator to trigger out of scope flag C: log(b_tot) log(a_ys+a_btsm) log(a_pig) All at 443 nm Out of scope flag Angles: Sun zenith View nadir Azimuth diff If RLw < , RLw = ( ρ 0.003) Reflectance spectrum Output of NN flag PCD1_16,17 q true if sum (r(i) / r (i)) > 4.0 courtesy of H. Schiller

34 Detection of out of scope conditions (MERIS processor) Top of atmosphere radiance spectra at normal and critical locations

35 Detection of out of scope conditions (MERIS processor) Exceptional bloom, Indicated by high Chi_square value Chi_square is computed by comparing The input reflectance spectrum with the output of the forward NN

36 Detection of out of scope conditions using an aann Important to detect toa radiance specta which are not in the simulated training data set These are out of scope of the atmospheric correction algorithm Autoassociative neural network with a bottle neck layer Functions also as nonlinear PCA i.e. bottle neck number of neurons Provide estimate of Independent components Input layer Hidden 1 Bottleneck Hidden 3 output layer For the GAC training data Set of ~ 1Mio. Cases Bottleneck minimum was 4-5

37 Detection of out of scope conditions aann: example for L1 (TOA) data High SPM Sun glint Transect

38 Detection of out of scope conditions aann: example AutoNN test 12x5x12 Yellow Sea transect, MERIS band 7, nm AutoNN test 12x5x12 Yellow Sea transect, MERIS band 7, nm rltoa rltoa_ann difference ratio rel. deviation RLtoa longitude longitude significant deviation in area with high SPM concentrations, but not in sun glint area rel. frequency AutoNN test Yellow Sea transect 12x5x12, MERIS band 7, nm Histogram of deviations shows 2 maxima, around 1 in sun glint 0.9 in high SPM area, which out of scope rel. radiance reflectance ratio to true

39 Validation NOMAD data set as reference Compiled, quality checked and maintained by OC group of NASA In situ observations from different cruises, different teams, instruments, procedures, sky and wave conditions Includes RLw at 6 MERIS bands (412,443,490, 510, 560,665) a_total, b_total / bb_total at443 Note: in situ data have their own variabilities and uncertainties! Relationship between chlorophyll a concentration and the absorption coefficient of phytoplankton pigments

40 a443 log10_a443_nn = log10_a443_measured * , stdev = 0.141

41 Frequency distribution Frequency distributions of measured and derived a443 after removing outliers with sum_sq > 1.0 e-5

42 Measured and nn-derived a443 for all cases with sd <1.0e-5

43 Test of NN based on measurements for chlorophyll Log10 scale, red: 1 by 1 line

44 Comparison of histograms: measured, NN computed

45 NN for kd489

46 Histogram kd489 measured and NN derived

47 Helgoland transect C Rtosa Rpath aann

48 Rw water reflectance 443 nm 560 nm 756 nm

49 Helgoland transect C atotal_443 TSM

50 Helgoland transect C apig_443 and uncertainty chlorophyll

51 Uncertainties related to comparison with in situ data Error in method and handling, e.g. HPLC for chlorophyll determination Sample not representative for water volume of pixel Vertical distribution: water comes from a certain depth, e.g. 4 m for FerryBox Temporal difference between sample and satellite overpass Sub-pixel patchiness Scatter in bio-optical data, e.g. relationship between concentration and IOPs

52 Summary and conclusions Uncertainty in coastal water products can be large due to the large number of factors in atmosphere and water, which determine the reflectance spectrum Conditions where algorithms (AC & water) fail Prerequisite for a successful retrieval are optical models of the atmosphere and the water, which meet the actual conditions Regional models might be necessary Reflectance spectra have to be tested if they are within the scope of these models Out of scope spectra have to be flagged, treated with special algorithms or excluded from further processing Limited sensitivity of reflectance spectrum and ambiguities lead to an uncertainty even for spectra, which are in scope Uncertainties have to be quantified on a pixel-by-pixel basis Validation in coastal waters by match up in situ samples can be difficult due to patchiness and fast changes Uncertainties in in situ match up data have to be quantified Validation should be complemented by statistical analysis of larger areas, transects and time series