Artificial Neural Networks-Based Radar Remote Sensing to Estimate Geographical Information during Oil-Spills

Abstract

In this paper, we demonstrate how an artificial neural network (ANN) of low complexity can be used on radar data in the remote sensing field to estimate geographical information from oil-spills scenes. The work aims at the extraction of two features that are important for an effective contingency plan: the thickness of thick oil slicks, and their relative dielectric constant physical parameter. The adopted system model assumes reflectivities measured by wide-band radar sensors operating in C-and X- frequency bands and mounted on nadir-looking systems such as drones. It extracts the thickness of oil slicks being in the 1–10 millimeter range and the dimensionless relative dielectric constant (permittivity) of the heavy oil material in the 2.8-3.3 range. We test the accuracy of the ANN model using simulated and in-lab experimental data. Finally, we validate the low complexity of our approach by providing FPGA implementation results of the inference. To the best of our knowledge, ANNs in combination with the active radar sensor have not been used for oil-spills parameters' estimation so far.