Hu Wenjie, Sun Guangtong, Zeng Xiangqiang, Tong Bo, Wang Zihao, Wu Xinyue, Song Ping
Institute of Disaster Prevention, Sanhe, 065201, China.
Hebei Key Laboratory of Resource and Environmental Disaster Mechanism and Risk Monitoring, Sanhe, 065201, China.
Sci Rep. 2025 Jul 1;15(1):21933. doi: 10.1038/s41598-025-08695-8.
Accurate landslide segmentation using remote sensing imagery is a critical component of geohazards response systems, particularly in time-sensitive tasks such as post-earthquake landslide damage assessment and emergency resource allocation. However, current methodologies struggle with two persistent challenges in sub-meter true-color imagery: fine-grained inter-class confusion between landslides and spectrally analogous terrain features, and within-landslide heterogeneity where localized damage signatures coexist with macro-scale deformation patterns within individual landslide bodies. To overcome these, we propose the Cross-Attention Landslide Detector (CALandDet), which improves the model's ability to distinguish between landslide and background features by sharply capturing global landslide feature information and integrating global landslide feature information with local information via a cross-attention feature enhancement mechanism. Ablation experiments show that CALandDet outperforms baselines, as evidenced by a 4.89% enhanced F1 score and an 8.73% greater Intersection over Union (IoU). In comparative experiments, it outperforms the other models by 8.05-10.78% in IoU and 1.05-8.9% in F1 score, achieving an IoU of 82.65% and an F1 score of 81.64%. Furthermore, the Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirm that the decision regions generated by the CALandDet model exhibit a higher spatial consistency with the actual landslide areas, effectively capturing indicative features including surface textures, sliding debris, accumulation bodies, and vegetation destruction. The proposed method may serve as a reference for future advancements in landslide segmentation and other remote sensing segmentation tasks.
利用遥感影像进行准确的滑坡分割是地质灾害响应系统的关键组成部分，特别是在诸如地震后滑坡灾害评估和应急资源分配等对时间敏感的任务中。然而，当前的方法在亚米级真彩色影像中面临两个长期存在的挑战：滑坡与光谱相似的地形特征之间存在细粒度的类间混淆，以及滑坡内部的异质性，即局部破坏特征与单个滑坡体内的宏观变形模式共存。为了克服这些问题，我们提出了交叉注意力滑坡检测器（CALandDet），它通过敏锐地捕捉全局滑坡特征信息，并通过交叉注意力特征增强机制将全局滑坡特征信息与局部信息相结合，提高了模型区分滑坡和背景特征的能力。消融实验表明，CALandDet优于基线模型，F1分数提高了4.89%，交并比（IoU）提高了8.73%。在对比实验中，它的IoU比其他模型高出8.05 - 10.78%，F1分数高出1.05 - 8.9%，IoU达到82.65%，F1分数达到81.64%。此外，梯度加权类激活映射（Grad-CAM）可视化结果证实，CALandDet模型生成的决策区域与实际滑坡区域具有更高的空间一致性，有效地捕捉了包括表面纹理、滑动碎屑、堆积体和植被破坏在内的指示性特征。所提出的方法可为未来滑坡分割及其他遥感分割任务的进展提供参考。
