Semi-supervised semantic change detection of remote sensing images with dual-resolution semantic guidance network

Semantic change detection (SCD) for remote sensing images (RSIs) based on deep learning has made great progress, but usually requires massive labelled images. For bitemporal RSIs, pixel-level annotation usually consumes expensive resources. To alleviate the reliance on large amounts of annotated data, this article proposes a semi-supervised SCD method of RSIs with a dual-resolution semantic guidance network (DR-SGNet). Specifically, a concatenated aggregation pyramid pooling module (CAPPM) is designed to enlarge effective receptive fields and extract more abundant context information from low-resolution feature maps. Meanwhile, we propose a multi-level change feature fusion module (MCFF) which uses multiple fusion strategies to fuse the multilevel change features generated during the coding process. The feature maps extracted by CAPPM are combined with the semantic information extracted from ResNet34 to generate semantic segmentation maps and also combined with the change information extracted from MCFF to efficiently perform change detection (CD). In addition, a random data augmentation method is designed for a large number of unlabelled images, which is combined with the standard teacher-student framework to realize semi-supervision using a small number of labelled images, thus significantly improving the model performance. Comprehensive experiments show that the proposed DR-SGNet outperforms previous state-of-the-art performance. The code will be available at: https://github.com/RSMeng/DR-SGNet.