Detection of Fishing Boats in SAR Image using Linear Moments and Machine Learning

Abstract

Synthetic Aperture Radar is a popular instrument used in maritime surveillance due to its all weather imaging capabilities and ability to discriminate objects from its backscatter. One of the important concern in maritime safety is detection of fishing boats for rescue and other strategic operations. Popular approaches such as, Constant False Alaram Rate (CFAR) and its variations, distribution based approaches such as K-Wishart, Guass, G0 and Deep Learning approaches have been widely employed but with limitations such as false detection or no detection due to non-convergence or high computational requirements. Keeping in view of the above, in this article we propose a simple approach based on Linearized Moments in association with machine learning methods to detect fishing boats. Our approaches is examined on a data set acquired from a Airborne SAR platform with frequencies in L and C bands operating in multiple polarization modes and having a spatial resolution of 1.2m. The proposed approach showed better accuracy in identifying fishing boats from Sea clutter with an area under curve (AUC) of 0.98 using a C4.5 decision tree classifier. Our study had shown that Linearized Moments can better model the Sea clutter thus improving the ability of machine classifers to detect fishing boats with improved accuracies.