SAR Image Classification using Transfer Learning

Abstract

Automatic Target Recognition (ATR) is a prominent step in Maritime Domain Awareness (MDA). These missions are aided by data from multiple sensors like Electro Optic-Infrared (EO/IR) cameras, Synthetic Aperture Radar (SAR), etc. Recent studies of classifying SAR images have been accomplished using supervised deep learning techniques to enhance the performance of classification. In this paper, we use transfer learning on VGG-16 architecture pre-trained on ImageNet dataset. The VGG-16 architecture is extended using a 2-step modular approach for the data classes in MASATI-V2 binary and MSTAR-SAR multiclass datasets. The model is then trained by suitable tuning of hyperparameters, and we report accuracies of 100% for binary classification and 99.18% for SAR multiclass classification. This approach reduces the overfitting in case of smaller dataset volume that is typical in SAR imagery in maritime domain, while reducing the time taken for training and deployment of the model.