Accession Number : ADA236731
Title : Back Propagation Neural Networks for Bathymetry Modeling Using Multispectral Data,
Corporate Author : NAVAL OCEANOGRAPHIC AND ATMOSPHERIC RESEARCH LAB STENNIS SPACE CENTER MS
Personal Author(s) : Rosche, Henry, III
Report Date : 05 MAY 1991
Pagination or Media Count : 9
Abstract : The Naval Oceanographic and Atmospheric Research Laboratory (NOARL) has investigated using back propagation neural networks to model bathymetry from multispectral data. By using three multispectral bands, bathymetry estimation is possible. Traditional computational models fail to handle wide ranges in bathymetry sue to the highly non-linear nature of the input data. Back propagation neural networks, which are highly non-linear in nature, were selected to determine if faster and more accurate results could be achieved over larger ranges of bathymetry values. To evaluate this approach, 211 known spectral values for the three bands and their registered bathymetry were used as control and experimental sets. The data were sorted by increasing values of depth. The training set was obtained by taking every other value from the initial data set and by running over several thousand training cycles. After the network has shown sufficient adjustment of the internal weights, the control set is presented to the network. The rms error for this set is 1.8 meters.
Descriptors : ACCURACY, BATHYMETRY, CONTROL, CYCLES, DATA BASES, DEPTH, ESTIMATES, INPUT, INTERNAL, MATHEMATICAL MODELS, MODELS, MULTISPECTRAL, NAVAL RESEARCH LABORATORIES, NEURAL NETS, PROPAGATION, RANGE(EXTREMES), SPECTRA, TRAINING, VALUE, WEIGHT.
Subject Categories : Physical and Dynamic Oceanography
Distribution Statement : APPROVED FOR PUBLIC RELEASE