关键词: 深度学习, 高分遥感, 变化检测, Cesium, WebGIS

Abstract:  Aiming at the problems that deep learning change detection methods are sensitive to illumination and registration errors and lack of samples, this paper proposes and implements a set of intelligent extraction methods and technical processes for surface remote sensing image change patches in coal mining areas based on Cesium WebGIS and deep learning theory from theoretical and practical perspectives. Firstly, a multi-task spatio-temporal attention change detection network (MSTACDN) is constructed by improved fusion of spatio-temporal feature matching consistency detection and large convolutional kernel spatio-temporal attention, to realize intelligent extraction of remote sensing image change patches. Then, the constructed WebGIS platform is used for visual display, management, and multi-user human-computer collaborative editing of the extracted change detection patches, so as to improve the correctness and accuracy of detection results. Finally, the manually edited results are stored in the sample library as new samples and fed back to the deep learning model for autonomous learning to further improve the model accuracy. The practical application results of change detection for three typical features (buildings, water bodies, and roads) show that compared with FC-SD (fully convolutional-siamese difference), SNUNet (siamese nested UNet), BITNet (bi-temporal image transformer network), and ChangeFormer algorithms, the proposed deep learning algorithm achieves the optimal detection accuracy. The two accuracy indicators of IoU (intersection over union) and F1 scores for buildings, water bodies, and roads are 82.26% and 91.79%, 81.68% and 91.50%, and 72.13% and 87.59%, respectively, which are significantly better than other models. Autonomous learning with the enhanced sample library effectively compensates for the shortage of samples. After one round of sample enhancement, the IoU and F1 scores of buildings, water bodies, and roads are increased to 89.53% and 91.86%, 85.74% and 91.78%, and 82.32% and 89.77% accordingly.

Key words: deep learning, high resolution remote sensing, change detection, Cesium, WebGIS