关键词: 遥感影像, 变化检测, 差分增强, 并行编码, 多尺度融合

Abstract: Remote sensing image change detection aims at identifying target differences in remote sensing images of different periods, and methods based on convolutional neural networks have made great progress in remote sensing image change detection tasks in recent years. However, the problem of pseudo-changes in remote sensing images in different periods is still difficult to be solved due to the influence of light changes and seasonal changes. Meanwhile, multi-scale features are not fully utilized in most of the methods, resulting in a certain degree of limitation in the performance and accuracy of the models. To address the above problems, a multi-scale differential feature-enhanced change detection method is proposed. Firstly, a parallel coding framework consisting of a twin network encoder and a differential network encoder is used to extract features at different levels respectively, and the same level of diachronic features and differential features are spliced to establish a complementary relationship between them. Then, the differential feature enhancement module is introduced to obtain more discriminative feature maps as supplementary inputs to the differential network encoder, which enriches the change information and increases the attention of the model to the change area, so that it can accurately distinguish the real change from the pseudo change of the features. Finally, in order to enhance the diversity and expressiveness of the features, the feature mismatch fusion module is used to achieve the cross-fertilisation of the semantic features, so that the semantic information in each feature is fully and differently interacted. The F1 scores of this method on the CDD dataset and LEVIR-CD dataset reached 95.45% and 92.04%, and the cross-merging ratios reached 92.26% and 82.93%, which were optimal compared with the remaining eight mainstream methods, and the experimental results proved the effectiveness of this method.

Key words: remote sensing images, change detection, differential enhancement, parallel encoding, multi-scale feature fusion