Beyond Accuracy: Evaluating Model Reliability for Target Radar Image Classification Using SHAP

Deep Neural Networks (DNNs), which play a significant role in a wide range of both civilian and military applications, are often regarded as "black-boxes" due to their opaque decision-making processes, and this raises concerns about their reliability. Explainable AI (XAI) aims to address this issue by providing human-understandable explanations for their decisions. In low-risk tolerance applications like automatic target recognition (ATR), ensuring reliable predictions is imperative. While some reliability measures exist, such as scalar thresholds and model ensembles, they often fall short in practical settings. This paper addresses the challenge of enhancing reliability in the classification of Synthetic Aperture Radar (SAR) images by incorporating XAI techniques. Specifically, the SHapley Additive exPlanations (SHAP) algorithm is employed to distinguish between positive and negative contributions of individual pixels to the outcomes of Convolutional Neural Networks (CNNs). Reliability is associated with the model’s ability to focus on target features rather than spurious correlations from background clutter. The paper proposes a novel numerical reliability index based on the SHAP analysis to enhance the decision-making process. This index quantifies the reliability of predictions based on pixel contributions. The theoretical findings are evaluated using the OpenSARShip dataset, which contains ship patches extracted from SAR images.