470 likes | 606 Views
Correction of the influence of the atmosphere using forward and inverse neural networks. Roland Doerffer Retired from Helmholtz Zentrum Geesthacht Institute of Coastal Research roland.doerffer@hzg.de. Space Shuttle 335 km 20060720 , NASA. Atmospheric correction using models.
E N D
Correction of the influence of the atmosphere usingforward and inverse neural networks Roland Doerffer Retired from Helmholtz Zentrum Geesthacht Institute of Coastal Research roland.doerffer@hzg.de Space Shuttle 335 km 20060720 , NASA
Atmospheric correction using models The NN associates a large number of independent components (IOPs) with the dependent reflectances Top of Atmosphere (TOA) reflection Atmosphere: molecule & aerosol scattering, gas abs.: Pathradiance,transmittance Air/sea interface, refractive index: Rfresnel Water reflection Rw Forward model inverse model Optical properties: a(),b (),Fl., Water constituents Phytoplankton, SPM etc Replace a complex radiative transfer model by a neural network, which is trained with simulations using the RTF Bottom reflection
Type of Neural networks Angular dependent Reflectances + aux angles,T, S, Angular dependent reflectances out in Further NNs: autoNN, errorNN, Normalisation NN Forward NN Inverse NN Combination of Inverse and forward NN In out Independent: IOPs, angles, T, S, Independent: IOPs Fixed to bands, conditions, Harder to train, incl. ambiguities But: very fast, less noise Very flexible, can be adapted To bands, conditions But: slow, more noise, ambiguities During run
Inverse Modellierung using NN in Optimization Procedures Start by Inverse NN Reflexion- Spektra Satellite Start values Modell Parameters IOP / Konz Do spectraagree ? Parameters are the IOPs / Konz. yes Reflexion- Spektra simulated Radiative transfer Model or Forward NN no Change Parameters Determine Search direction Downhill in cost function Test = ∑(Rsim(i) – Rsat(i))2 Also measure of quality !
Training of a neural network for atmospheric correction 7 Atmosphere - optical model Bio - optical model Selection Max sunglint 1 Max tau_aerosol NNforward water 8 Min. Rlw(560) ( based on Hydrolight simulations ) Etc. 9 RLw 12 Transmittance 10 L_up 5 RLtosa RLpath_noglint RLpath RLpath Training & RLpath_glint Ed_boa Ed_boa MC code Test data set Ed_boa Tau_aerosole RLw Tau_aerosole 2 Tau_aerosole 13 4 Optional 11 Polarisation 6 correction
Aerosol Optical Properties used for NN Training data set • RTF models used for • simulation of training • data sets: • Monte Carlo photon tracing • 6Sv • Aeronet based model by • R. Santer • Water: Hydrolight C. Mobley Angström Coefficient Aerosol Optical Thickness (AOT) at 550 nm
Turbid water reflectances Amazone: Rw 0.18 at 681nm Plata: Rw 0.17 at 681nm Yangtse: Rw 0.2 at 681nm Cocco: Rw 0.12 at 560 nm
Simulations with Hydrolight • Hydrolight 5.1 for computation of bi-directional water leaving radiance reflectance spectra (RLw) • Extension of Hydrolight with the pure water model of this project • Temperature • Salinity • Refractive index • Uncertainties • bio-optical model: • 5 optical components • Absorption coefficients of pigment, detritus, yellow substance • Scattering coefficients of particles and white particles
Inverse NN for atmospheric correction – CC version Input (18) Output (43) RLtosa 12 bands Tau_aerosol 412, 550, 778, 865 Sun_glint ratio a_tot, b_tot Neural Network 18->25x30x40->43 sun zenith view x View y MERIS band 1-10, 12,13 View z Ed_surface temperature Path radiance reflectance RLw salinity RLw(,) =Lw (,) /Ed For each water type
Scheme for forwardNN based procedure RLtosa‘ MERIS RLtoa Compare RLtosa RLtosa‘ RLath, trans forNNatmo 3 par Compute RLtosa aaNN OOS forNNwater n_lam n_var (5) Optim, procedure Results: RLw IOPs uncertainty Select Bands Variables parameters constraints Start values
Testof NN 17x27x17 Training with 5% random noise on rlw 560 nm
Test of water algorithm forward NN Relationship between the measured and derived total absorption coefficient a_total at 443 nm using the neural network algorithm. Red is the 1 by 1 line, blue the regression. n=498 points
Separation atmosphere / water using inverse NN MERIS band 5, 560 nm
Transect Test using forward NN Inverse NN Forward NN
Separation test: RLw vs RLpath MERIS scene of Hawaii 20040406 TOA radiance reflectance band 13 with transect Top of atmosphere, path radiance and water leaving radiance reflectance along transect of Hawaii scene Band 5 560 nm Rlw *10
Comparison as scatter plot Neural network AGC /C2R Standard AC MERIS MOBY 73 no glint cases, log10 scale
Comparison as scatter plot Neural network AGC /C2R 73 no glint cases Neural network AGC /C2R 85 high glint cases
Identify spectra, which are out of scope of the training set
Detection of out of scope conditions • 2 Procedures have been developed • Combination of an inverse and forward Neural Network • Use of an autoassociative Neural Network • Both produce a reflection spectrum, which is compared with the input spectrum • Deviation between input and output spectrum can be computed as a chi2 • A threshold can be used to trigger an out of scope warning flag • Combination • of inverse • and forward NN R output R Input Inverse NN IOPs Forward NN R output • Auto-associative • NN R Input autoassociative NN
Detection of out of scope conditions (MERIS processor) Top of atmosphere radiance spectra at normal and critical locations
Detection of out of scope conditions (MERIS processor) Exeptional bloom, Indicated by high Chi_square value Chi_square is computed by comparing The input reflectance spectrum with the output of the forward NN
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 For the GAC training data Set of ~ 1Mio. Cases Bottleneck minimum was 4-5 Bottle- neck Input layer output layer Hidden 1 Hidden 3
Detection of out of scope conditions aaNN:example for L1 (TOA) data High SPM Sun glint Transect
Detection of out of scope conditions aaNN: example rel. deviation Histogram of deviations shows 2 maxima, around 1 in sun glint 0.9 in high SPM area, which out of scope significant deviation in area with high SPM concentrations, but not in sun glint area
TOSA, Path, Water Reflectance 708 nm Inverse NN
Cross section Hawai scene radiance_9 [mW/(m^2*sr*nm)] 200 180 160 140 120 radiance_9 [mW/(m^2*sr*nm)] 100 80 60 40 20 0 -160 -158 -156 -154 -152 -150 -148 -146 longitude (deg)
Simulated Rayleigh path radiance reflectance and sun glint radiance reflectance 0.04 nadir view 0.035 sun zenith 20 deg wind 3 m/s 0.03 Rayleigh path radiance 0.025 0.02 Radiance refelctance [sr-1] sun glint 0.015 0.01 0.005 0 400 450 500 550 600 650 700 750 800 850 900 wavelength [nm]
Specular refleced radiance differences, ratio rel. to Lw T5/S35 T25/S0
MERIS full resolution: Baltic and North Sea, 20080606 Stockhom Sun glint Oslo Baltic Sea Sun glint North Sea Copenhagen Spatial resolution: 300 m Swath: 1200 km, 4800 pixel Hamburg
MERIS FR, Area of Gotland, TOA RLw RGB Stockholm Estonia sunglint Lettland Gotland Ca. 100 km Baltic Sea
Water leaving radiance reflectance RGB Gotland Ca. 100 km
MERIS 20070505: TOA reflectances RGB New York
Path radiance+ Fresnel reflectance RLpath MERIS band 5 (560 nm)
Water leaving radiance reflectance RLw MERIS band 5 (560 nm)
MERIS FR USA East Coast 12.6.2008, Signal depth z90 Philadelphia Chesapeake Bay Washington North Atlantic
Overview • Artificial neural networks (NN) can be used for atmospheric correction in different ways • As a forward model to determine the path radiance and transmittance as a function of aerosol optical properties, wind (sun glint), angles • As an inverse model, which determines water reflectances, transmittances, path radiance from TOSA reflectance spectra • NN AC is based on association between water reflectance and top of atmosphere reflectance • No negative reflectances • The relationship between TOSA reflectance, path radiance, transmittance and the independent parameters (pressure, aerosols, wind/waves) must be described with a radiative transfer model • NNs are then trained by a large number of simulated cases (> 1 Mio) by minimizing the difference between the output of the NN and • For turbid coastal waters reflectance must include reflectance by water • For coastal waters include all bands for atmospheric corrections, no extrapolation