NEURAL NETWORKS IN ATMOSPHERIC REMOTE SENSING

By KENNETH GONZALEZ

CONTENTS INTRODUCTION PRESENT CHALLENGES SOLUTIONS BASED ON NEURAL NETWORKS MATHEMATICAL NOTATION REFERENCES PHYSICAL BACKGROUND OF ATMOSPHERIC REMOTE SENSING OVERVIEW OF THE COMPOSITION AND THERMAL STRUCTURE OF THE EARTH’S ATMOSPHERE CHEMICAL COMPOSITION OF THE ATMOSPHERE VERTICAL DISTRIBUTION OF PRESSURE AND DENSITY THERMAL STRUCTURE OF THE ATMOSPHERE CLOUD MICROPHYSICS ELECTROMAGNETIC WAVE PROPAGATION MAXWELL’S EQUATIONS AND THE WAVE EQUATION POLARIZATION RE?ECTION AND TRANSMISSION AT A PLANAR BOUNDARY ABSORPTION OF ELECTROMAGNETIC WAVES BY ATMOSPHERIC GASES MECHANISMS OF MOLECULAR ABSORPTION LINE SHAPES ABSORPTION COEF?CIENTS AND TRANSMISSION FUNCTIONS THE ATMOSPHERIC ABSORPTION SPECTRA SCATTERING OF ELECTROMAGNETIC WAVES BY ATMOSPHERIC PARTICLES MIE SCATTERING THE RAYLEIGH APPROXIMATION COMPARISON OF SCATTERING AND ABSORPTION BY HYDROMETEORS RADIATIVE TRANSFER IN A NONSCATTERING PLANAR-STRATI?ED ATMOSPHERE EQUILIBRIUM RADIATION: PLANCK AND KIRCHHOFF’S LAWS RADIATIVE TRANSFER DUE TO EMISSION AND ABSORPTION INTEGRAL FORM OF THE RADIATIVE TRANSFER EQUATION WEIGHTING FUNCTION PASSIVE SPECTROMETER SYSTEMS OPTICAL SPECTROMETERS MICROWAVE SPECTROMETERS SUMMARY AN OVERVIEW OF INVERSION PROBLEMS IN ATMOSPHERIC REMOTE SENSING OPTIMALITY METHODS THAT EXPLOIT STATISTICAL DEPENDENCE THE BAYESIAN APPROACH LINEAR AND NONLINEAR REGRESSION METHODS PHYSICAL INVERSION METHODS THE LINEAR CASE THE NONLINEAR CASE HYBRID INVERSION METHODS IMPROVED RETRIEVAL ACCURACY IMPROVED RETRIEVAL EF?CIENCY ERROR ANALYSIS ANALYTICAL ANALYSIS PERTURBATION ANALYSIS SIGNAL PROCESSING AND DATA REPRESENTATION ANALYSIS OF THE INFORMATION CONTENT OF HYPERSPECTRAL DATA SHANNON INFORMATION CONTENT DEGREES OF FREEDOM PRINCIPAL COMPONENTS ANALYSIS (PCA) NONLINEAR PCA LINEAR PCA PRINCIPAL COMPONENTS TRANSFORMS THE PROJECTED PC TRANSFORM EVALUATION OF RADIANCE COMPRESSION PERFORMANCE USING TWO DIFFERENT METRICS REPRESENTATION OF NONLINEAR FEATURES INTRODUCTION TO MULTILAYER PERCEPTRON NEURAL NETWORKS A BRIEF OVERVIEW OF MACHINE LEARNING SUPERVISED AND UNSUPERVISED LEARNING CLASSI?CATION AND REGRESSION KERNEL METHODS SUPPORT VECTOR MACHINES FEEDFORWARD NEURAL NETWORKS FEEDFORWARD MULTILAYER PERCEPTRON NEURAL NETWORKS NETWORK TOPOLOGY NETWORK TRAINING SIMPLE EXAMPLES SINGLE-INPUT NETWORKS TWO-INPUT NETWORKS EXERCISES A PRACTICAL GUIDE TO NEURAL NETWORK TRAINING DATA SET ASSEMBLY AND ORGANIZATION DATA SET INTEGRITY THE IMPORTANCE OF AN EXTENSIVE AND COMPREHENSIVE DATA SET DATA SET PARTITIONING MODEL SELECTION NUMBER OF INPUTS NUMBER OF HIDDEN LAYERS AND NODES ADAPTIVE MODEL BUILDING TECHNIQUES NETWORK INITIALIZATION CALCULATION OF THE ERROR GRADIENT USING BACKPROPAGATION FIRST-ORDER OPTIMIZATION: GRADIENT DESCENT SECOND-ORDER OPTIMIZATION: LEVENBERG-MARQUARDT UNDER?TTING AND OVER?TTING REGULARIZATION TECHNIQUES TREATMENT OF NOISY DATA WEIGHT DECAY PERFORMANCE EVALUATION PRE- AND POST-PROCESSING OF ATMOSPHERIC DATA MATHEMATICAL OVERVIEW DATA COMPRESSION FILTERING OF INTERFERING SIGNALS THE WIENER FILTER STOCHASTIC CLOUD CLEARING DATA WARPING FUNCTION OF TIME OF DAY FUNCTION OF GEOLOCATION FUNCTION OF TIME OF YEAR NEURAL NETWORK JACOBIAN ANALYSIS CALCULATION OF THE NEURAL NETWORK JACOBIAN NEURAL NETWORK ERROR ANALYSIS USING THE JACOBIAN THE NETWORK WEIGHT JACOBIAN THE NETWORK INPUT JACOBIAN USE OF THE JACOBIAN TO ASSESS NOISE CONTRIBUTION RETRIEVAL SYSTEM OPTIMIZATION USING THE JACOBIAN NOISE SMOOTHING VERSUS ATMOSPHERIC SMOOTHING OPTIMIZATION APPROACH OPTIMIZATION RESULTS NEURAL NETWORK RETRIEVAL OF PRECIPITATION FROM PASSIVE MICROWAVE OBSERVATIONS STRUCTURE OF THE ALGORITHM PHYSICAL BASIS OF PREPROCESSING PHYSICAL BASIS OF POST-PROCESSING SIGNAL PROCESSING COMPONENTS LIMB-AND-SURFACE CORRECTIONS PRECIPITATION DETECTION CLOUD CLEARING BY REGIONAL LAPLACIAN INTERPOLATION TEMPERATURE-PRO?LE AND WATER-VAPOR-PRO?LE PRINCIPAL COMPONENTS IMAGE SHARPENING DEVELOPMENT OF THE ALGORITHM RETRIEVAL PERFORMANCE EVALUATION IMAGE COMPARISONS OF NEXRAD AND AMSU/HSB NUMERICAL COMPARISONS OF NEXRAD AND AMSU/HSB RETRIEVALS GLOBAL RETRIEVALS OF RAIN AND SNOW NEURAL NETWORK RETRIEVAL OF ATMOSPHERIC PRO?LES FROM MICROWAVE AND HYPERSPECTRAL INFRARED OBSERVATIONS THE PPC/NN ALGORITHM 10 1 1 NETWORK TOPOLOGY 10 1 2 NETWORK TRAINING RETRIEVAL PERFORMANCE COMPARISONS WITH SIMULATED CLEAR-AIR AIRS RADIANCES 10 2 1 SIMULATION OF AIRS RADIANCES 10 2 2 AN ITERATED MINIMUM-VARIANCE TECHNIQUE FOR THE RETRIEVAL OF ATMOSPHERIC PRO?LES 10 2 3 RETRIEVAL PERFORMANCE COMPARISONS 10 2 4 DISCUSSION VALIDATION OF THE PPC/NN ALGORITHM WITH AIRS/AMSU OBSERVATIONS OF PARTIALLY CLOUDY SCENES OVER LAND AND OCEAN 10 3 1 CLOUD CLEARING OF AIRS RADIANCES 10 3 2 AIRS/AMSU/ECMWF DATA SET 10 3 3 AIRS/AMSU CHANNEL SELECTION 10 3 4 PPC/NN RETRIEVAL ENHANCEMENTS FOR VARIABLE SENSOR SCAN ANGLE AND SURFACE PRESSURE 10 3 5 RETRIEVAL PERFORMANCE 10 3 6 RETRIEVAL PERFORMANCE SENSITIVITY ANALYSES 10 3 7 DISCUSSION AND FUTURE WORK SUMMARY AND CONCLUSIONS