Evidential Uncertainty in SAR Target Recognition for Reliable Prediction

Deep learning techniques have shown promising results in SAR Automatic Target Recognition (ATR). However, most SAR ATR methods use traditional deterministic neural networks supervised by loss functions based on maximum likelihood estimation, often leading to overconfident predictions and neglecting uncertainty, posing significant risks to the field. This paper delineates common sources of uncertainty in neural networks within the SAR ATR context, including small samples, speckle noise, unfocused blur, and low resolution. A thorough simulation of these scenarios is conducted to elucidate their impact on traditional deterministic models and to explore mitigation strategies. We introduce Evidential Deep Learning (EDL) as a novel approach in SAR ATR to enhance prediction reliability. Our experimental findings reveal that uncertainty significantly compromises prediction accuracy, whereas integrating EDL bolsters prediction reliability and enhances model robustness by uncertainty quantification and probability calibration. The data and code are available at https://github.com/GhostWCY/Evidential-SAR-ATR.git