Limited-Data SAR ATR Causal Method via Dual-Invariance Intervention
Chenwei Wang, Renjie Xu, Yulin Huang, Jifang Pei, Chuan Huang, Wenqi Zhu, Jianyu Yang
Abstract
Synthetic aperture radar automatic target recognition (SAR ATR) with limited data has gained significant attention as practical application requirements change. Despite many proposed methods, key problems caused by the limited SAR data remain under-researched, hindering further performance improvement. In this article, we establish an SAR ATR model based on causal theory. It compares the causal effect of SAR ATR between cases with ample and limited data, showing that the negative impact of the confounder, which is blocked with ample data, is introduced with limited data, resulting in poor performance of limited-data SAR ATR. To address this, we propose a limited-data SAR ATR causal method via dual invariance intervention, which first derives the causal interventional solution. This solution is transformed into two optimizable objectives: inner-class feature invariance and the independence of features from the confounder. Subsequently, the dual invariance mechanism is designed to filter SAR outlier samples and noise features under limited-data conditions, accurately obtaining the intraclass invariant feature. It also alleviates the need for ample SAR data when optimizing the independence of features from the confounder, achieving the two objectives. Finally, the proposed method not only unravels the key problem caused by limited data but also derives an effective solution with precise recognition performance. Extensive experiments on three benchmark datasets validate the rationality of the causal SAR ATR (CSA) model, the effectiveness of the solution, and the soundness and recognition performance of the method. The codes and more experimental results will be released at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/cwwangSARATR/SARATR_Causal_Dual_Invariance</uri>.