ENHANCED POLSAR IMAGE CLASSIFICATION USING DEEP CONVOLUTIONAL AND TEMPORAL CONVOLUTIONAL NETWORKS

Abstract

A new framework in the form of Polarimetric Synthetic Aperture Radar (PolSAR) image classification, where deep Convolutional Neural Networks (CNNs) were integrated with the traditional Machine Learning (ML) techniques under a Temporal Convolutional Network (TCN) architecture, was introduced in the paper. The main aim behind this new approach is to overcome the severe limitations inherent in both deep CNN and conventional ML approaches. The application of the sliding-window strategy eliminates the necessity of requiring extensive feature extraction procedures while reducing computational complexity simultaneously. Experiments on four benchmark PolSAR datasets for C-Band, L-Band, AIRSAR, and RADARSAT-2 data attest to the framework's remarkable classification accuracies in the range of 94.55% to 99.39%. This integrated framework is thus a significant advancement in PolSAR image analysis in offering an efficient methodology that combines the strengths of deep CNNs and traditional ML, by mitigating their respective limitations. It also combines the sliding-window technique with the architecture of TCN and then yields excellent classification accuracy with no much additional computational overhead. The results obtained thus indicate a good chance of revolutionizing the state of the art in PolSAR image classification, providing crucial efficiency improvements and making applications in environmental applications stronger, across almost all kinds of fields.