MSFCN: A Multiscale Feature Correlation Network for Remote Sensing Image Scene Change Detection

Abstract

Scene-level change detection identifies land use changes and determines change types from a high-level semantic perspective, which is significant for monitoring urbanization. The existing advanced methods are generally based on Siamese networks that utilize the feature correlation of bitemporal scenes or introduce change information to enhance the feature representation. However, their extraction of feature correlation is insufficient to improve the model performance further. This article proposed a Siamese-based multiscale feature correlation network (MSFCN) to enhance the correlation extraction process. First, 1-D multiscale local features are obtained by the designed space-channel self-calibration module and multiscale local feature extraction module. Then, these features are inputted into the proposed multiscale feature correlation module to extract feature correlation. Finally, the dual-branch features are fused based on the feature correlation to generate more discriminative 1-D deep features. In addition, cosine embedding loss is used to constrain the scene binary change detection task and construct a multitask loss for model optimization. On the Hanyang and WH-MAVS datasets, MSFCN achieved average scene classification accuracies of 93.33% and 94.86%, scene-level binary change detection accuracies of 95.71% and 98.13%, and scene-level semantic change detection accuracies of 90.00% and 93.95%, respectively, significantly better than the comparison methods.