基于双时相遥感影像差异信息的深度学习滑坡检测

目前利用高分辨率卫星影像进行滑坡等地质灾害识别逐渐成为研究热点，滑坡目视解译依赖于解译人员的经验，耗时费力且提取精度低，而传统的滑坡自动识别方法易将滑坡和道路、裸地、建筑等多种具有相似光谱信息的地物混淆。针对以上问题，文章使用一种双时相高分辨率卫星影像差异信息的深度学习滑坡检测算法，获取时序影像各个波段和归一化植被指数（Normalized Difference Vegetation Index,NDVI）的差异影像作为深度学习的输入特征。为充分挖掘滑坡前后影像多种信息差异特征，采用了U-net网络模型耦合空洞空间金字塔池化和嵌入注意力机制模块相结合进行滑坡特征提取的方法，该方法增强了滑坡边界信息的保存，能够有效地提取滑坡边界信息和发生剧烈变化的区域。利用上述方法对恩施市和九寨沟进行了滑坡检测，实验结果显示，所取得的综合评价指标值（F1-Score）分别为88.4%和90.53%，误差较小、精度较高。表明该方法能够准确检测出高分卫星数据的滑坡边界，且能保持滑坡的完整性。

Deep Learning Landslide Extraction Based on Difference Information of Dual-phase Remote Sensing Images

Current using of high-resolution satellite images to identify geological hazards such as landslides has gradually become a research hotspot. The visual interpretation of landslides relies on the experience of the interpreter, and is time-consuming and labor-intensive, and the extraction accuracy is low. However, the traditional landslide automatic identification method is easy to confuse the landslide with various ground objects with similar spectral information, such as roads, bare ground and buildings. In response to the above problems, this paper uses a deep learning technology landslide detection algorithm based on dual-phase high-resolution satellite image difference information, obtain each band of time series images and the normalized difference vegetation index (NDVI) difference image as the input feature of deep learning. To fully excavate the characteristics of various information differences in the images before and after the landslide, a method for landslide feature detection with U-net network model coupled with atrous spatial pyramid pooling and embedded attention mechanism module, this method enhances the preservation of landslide boundary information, and can effectively extract landslide boundary information and areas with drastic changes. Landslide detection in Enshi and Jiuzhaigou by the method in this paper, the experimental results show that the obtained F1-Scores are 88.4% and 90.53%, respectively, with small errors and high precision. The method in this paper can accurately detect the landslide boundary of high-resolution satellite data, and can maintain the integrity of the landslide.