Deriving Bathymetry From Optical Images With a Localized Neural Network Algorithm

We present a localized neural network algorithm for water depth estimation from optical remote sensing images. Our new model is called a locally adaptive back-propagation neural network (LABPNN). In an LABPNN, the neural networks were trained at regularly distributed normative sites. For each unit of LABPNN, training data samples were identified by a specified search radius from the normative sites. Water depth was estimated by an ensemble of LABPNNs, with weights assigned inversely by their distances to the point of estimation. The water depth prediction accuracy from the LABPNN models doubled compared to the regular back-propagation network that uses all of the samples without considering nonstationarity. We also compared the LABPNN model with the regression-based inversion model with the localization feature. The prediction error of LABPNN is less by about 5% in the first case study, and 7% less in the second case study. It is because of the better performance of neural networks than that of the regression models when the sample data are relatively sparse. The experiments suggest that the LABPNN model is a viable solution to water depth retrieval from optical images.