Perspective on Explainable SAR Target Recognition
SAR Automatic Target Recognition (ATR) is a key task in microwave remote sensing. Recently, Deep Neural Networks (DNNs) have shown promising results in SAR ATR. However, despite the success of DNNs, their underlying reasoning and decision mechanisms operate essentially like a black box and are unkno...
| Main Authors: | , , , |
|---|---|
| 格式: | 文件 |
| 语言: | English |
| 出版: |
China Science Publishing & Media Ltd. (CSPM)
2020-06-01
|
| 丛编: | Leida xuebao |
| 主题: | |
| 在线阅读: | http://radars.ie.ac.cn/article/doi/10.12000/JR20059?viewType=HTML |
| _version_ | 1827614321578868736 |
|---|---|
| author | GUO Weiwei ZHANG Zenghui YU Wenxian SUN Xiaohua |
| author_facet | GUO Weiwei ZHANG Zenghui YU Wenxian SUN Xiaohua |
| author_sort | GUO Weiwei |
| collection | DOAJ |
| description | SAR Automatic Target Recognition (ATR) is a key task in microwave remote sensing. Recently, Deep Neural Networks (DNNs) have shown promising results in SAR ATR. However, despite the success of DNNs, their underlying reasoning and decision mechanisms operate essentially like a black box and are unknown to users. This lack of transparency and explainability in SAR ATR pose a severe security risk and reduce the users’ trust in and the verifiability of the decision-making process. To address these challenges, in this paper, we argue that research on the explainability and interpretability of SAR ATR is necessary to enable development of interpretable SAR ATR models and algorithms, and thereby, improve the validity and transparency of AI-based SAR ATR systems. First, we present recent developments in SAR ATR, note current practical challenges, and make a plea for research to improve the explainability and interpretability of SAR ATR. Second, we review and summarize recent research in and practical applications of explainable machine learning and deep learning. Further, we discuss aspects of explainable SAR ATR with respect to model understanding, model diagnosis, and model improvement toward a better understanding of the internal representations and decision mechanisms. Moreover, we emphasize the need to exploit interpretable SAR feature learning and recognition models that integrate SAR physical characteristics and domain knowledge. Finally, we draw our conclusion and suggest future work for SAR ATR that combines data and knowledgedriven methods, human–computer cooperation, and interactive deep learning. |
| first_indexed | 2024-03-09T08:54:17Z |
| format | Article |
| id | doaj.art-859e5ad4832347b3adaff0fb9b6b51a1 |
| institution | Directory Open Access Journal |
| issn | 2095-283X 2095-283X |
| language | English |
| last_indexed | 2024-03-09T08:54:17Z |
| publishDate | 2020-06-01 |
| publisher | China Science Publishing & Media Ltd. (CSPM) |
| record_format | Article |
| series | Leida xuebao |
| spelling | doaj.art-859e5ad4832347b3adaff0fb9b6b51a12023-12-02T13:39:12ZengChina Science Publishing & Media Ltd. (CSPM)Leida xuebao2095-283X2095-283X2020-06-019346247610.12000/JR20059Perspective on Explainable SAR Target RecognitionGUO Weiwei0ZHANG Zenghui1YU Wenxian2SUN Xiaohua3①(Center of Digital Innovation, Tongji University, Shanghai 200092, China)②(Shanghai Key Lab of Intelligent Sensing and Recognition, Shanghai Jiaotong University, Shanghai 200240, China)②(Shanghai Key Lab of Intelligent Sensing and Recognition, Shanghai Jiaotong University, Shanghai 200240, China)①(Center of Digital Innovation, Tongji University, Shanghai 200092, China)SAR Automatic Target Recognition (ATR) is a key task in microwave remote sensing. Recently, Deep Neural Networks (DNNs) have shown promising results in SAR ATR. However, despite the success of DNNs, their underlying reasoning and decision mechanisms operate essentially like a black box and are unknown to users. This lack of transparency and explainability in SAR ATR pose a severe security risk and reduce the users’ trust in and the verifiability of the decision-making process. To address these challenges, in this paper, we argue that research on the explainability and interpretability of SAR ATR is necessary to enable development of interpretable SAR ATR models and algorithms, and thereby, improve the validity and transparency of AI-based SAR ATR systems. First, we present recent developments in SAR ATR, note current practical challenges, and make a plea for research to improve the explainability and interpretability of SAR ATR. Second, we review and summarize recent research in and practical applications of explainable machine learning and deep learning. Further, we discuss aspects of explainable SAR ATR with respect to model understanding, model diagnosis, and model improvement toward a better understanding of the internal representations and decision mechanisms. Moreover, we emphasize the need to exploit interpretable SAR feature learning and recognition models that integrate SAR physical characteristics and domain knowledge. Finally, we draw our conclusion and suggest future work for SAR ATR that combines data and knowledgedriven methods, human–computer cooperation, and interactive deep learning.http://radars.ie.ac.cn/article/doi/10.12000/JR20059?viewType=HTMLsarautomatic target recognition (atr)deep learningexplainability and interpretabilityexplainable machine learning |
| spellingShingle | GUO Weiwei ZHANG Zenghui YU Wenxian SUN Xiaohua Perspective on Explainable SAR Target Recognition Leida xuebao sar automatic target recognition (atr) deep learning explainability and interpretability explainable machine learning |
| title | Perspective on Explainable SAR Target Recognition |
| title_full | Perspective on Explainable SAR Target Recognition |
| title_fullStr | Perspective on Explainable SAR Target Recognition |
| title_full_unstemmed | Perspective on Explainable SAR Target Recognition |
| title_short | Perspective on Explainable SAR Target Recognition |
| title_sort | perspective on explainable sar target recognition |
| topic | sar automatic target recognition (atr) deep learning explainability and interpretability explainable machine learning |
| url | http://radars.ie.ac.cn/article/doi/10.12000/JR20059?viewType=HTML |
| work_keys_str_mv | AT guoweiwei perspectiveonexplainablesartargetrecognition AT zhangzenghui perspectiveonexplainablesartargetrecognition AT yuwenxian perspectiveonexplainablesartargetrecognition AT sunxiaohua perspectiveonexplainablesartargetrecognition |