ESTIMATING DEPTH AND BOTTOM CONTRIBUTIONS USING OPTIMIZATION ANDNEURAL NETWORK METHODS FOR HYPERSPECTRAL IMAGERY

摘要

Optimization and neural networks methods were applied to hyperspectral imagery (Rrs) over theLEO-15 and Tampa Bay sites to determine water depth and bottom composition. Both methodsattempt to solve the inverse optics problem where the optimization method utilizes forwardsolutions provided by a semianalytical method and the neural network utilizes HYDROLIGHT.The analysis of the respective inversion procedures and results is accomplished using sensitivityanalysis, error distribution analysis, and observations on convergence. The methods arecompared on the basis of accuracy, efficiency, and robustness. As a possible alternative, ahybrid method that uses both supervised and unsupervised neural network approaches in aniterative, optimization scheme is proposed to address the specific problem of bottomcomposition. For validation, the remotely sensed hyperspectral data collected over the sites werecompared to available in situ data. The final product is a quantitative comparison of depth andbottom properties to show the robustness and applicability of the optimization and neuralnetwork methods applied to diverse, coastal environments using hyperspectral data.