A lightweight deep learning model for classification of synthetic aperture radar images

Images acquired by SAR radars are massively being used for various earth observations, such as land and ocean surveillance, oil spill detection, and military and maritime vigilance. Classifying SAR images is challenging due to the noisy and unreadable picture quality of the images. Earlier, the classification of SAR images was timeconsuming since it involved manual participation, and automating such tasks has become an area of research. Numerous works have been proposed, focusing on the application of deep learning in SAR image classification. However, most of them are computationally expensive and result in misclassification. With an aim to curtail these issues, we studied the performance of the known deep learning models by implementing each model on SAR image classification. Based on the observed results, we have proposed a new lightweight classification model that is computationally efficient on SAR data. Experiments on the MSTAR benchmark show that the accuracy attained by the proposed model is at par with that of the high-computational models. The proposed model could scale down the parameters by up to 25 times compared to models such as VersNet while still achieving a classification accuracy of 97%. Our work, therefore, concludes that the use of the single-unit kernel for feature mapping contributes to a reduction in the number of convolutional computations. Additionally, the use of depthwise convolutions in the proposed model enables superior feature discrimination.