高光谱成像
遥感
计算机科学
人工智能
变更检测
模式识别(心理学)
计算机视觉
目标检测
调制(音乐)
特征提取
雷达跟踪器
合成孔径雷达
频率调制
杂乱
信号处理
图像处理
噪音(视频)
匹配滤波器
探测理论
作者
Shufang Xu,Wenxuan Fei,Shuyu Fei,Runhua Sheng,Hongmin Gao
标识
DOI:10.1109/tgrs.2026.3732009
摘要
Hyperspectral change detection (HSCD) remains challenging due to spectral inconsistency across bitemporal observations, heterogeneous spatial change patterns, and ambiguous decision boundaries that induce predictive uncertainty. Existing methods mainly rely on deterministic representation learning, lacking mechanisms to model uncertainty and resolve such conflicts, often resulting in overconfident and unreliable predictions. To address these issues, we propose CCEM-Net, a conflict-aware evidential deep learning (EDL) framework for reliable HSCD. Specifically, a change-aware interaction (CAI) module is introduced to model cross-temporal dependencies, suppress pseudo-changes, and enhance change-relevant responses via dual gating. A parallel adaptive multiscale spectral and spatial (AMSS) module further captures heterogeneous patterns through temporal-difference-guided multiscale feature aggregation. To handle decision ambiguity, we develop a conflict-aware cross evidential modulation (CCEM) mechanism that models branch predictions as Dirichlet distributions to quantify uncertainty. By measuring interbranch conflict via Jensen–Shannon (JS) divergence, CCEM adaptively modulates evidence strength. Subsequently, reliable decision-level fusion is performed using the Yager rule. Experiments on benchmark datasets demonstrate that CCEM-Net consistently achieves the superior performance in both accuracy and reliability compared with state-of-the-art (SOTA) methods. The code will be publicly available at:https://github.com/fwx0314/CCEM-Net
科研通智能强力驱动
Strongly Powered by AbleSci AI