A Neural Network Approach to Deriving Optical Properties and Depths of Shallow Waters

A Neural Network (NN) approach is studied in deriving information of bathymetry for optically shallow waters. In this study, more than 7000 remote-sensing reflectance (ratio of waterleaving radiance to downwelling irradiance above the surface) spectra for shallow waters were created with a semi-analytical model. This synthetic data base covered chlorophyll-a concentrations from 0.2 to 6 mg/m^3, water depths from 0.5 to 20 meters, and dark to bright-sand bottom albedos. The multi-layer NN is trained with the synthetic data using a back-propagation algorithm, and tested with both synthetic and field data. One advantage of using NN approach is that it reduces the calculation time greatly compared to an early optimization method. INTRODUCTION Recently, an optimization method has been developed to retrieve bottom depth and in-water properties from measured remote-sensing reflectance. The method is proved accurate and successful, however, it is too slow for image processing using current computers. A quick and reliable method for bathymetry is desired. Over the past several years, more attention have been paid for artificial neural networks (NN) for remote sensing applications. For example, Key used the Advanced Very High Resolution Radiometer (AVHRR) data in conjunction with the Scanning Multichannel Microwave Radiometer (SMMR) for the classification of four land surface and eight cloud classes in the Arctic. However, most applications of NN were toward the qualitative classifications of remotely sensed images, few researches were focused on the quantitative derivation of properties of interest. In this study, we investigate the application of using neural networks to derive bottom depth. To design a neural network, large data base is required to well train the neural network, which is not available yet from field measurement, however. Computer models are proved that they can generate shallow water spectra, but it is very time consuming. We generate our data base of shallow-water remote sensing reflectance by a semi-analytical model, which is simple and easy with high accuracy. We used this synthetic data base to train a neural network, and used both synthetic and field data to test it.