A HRNet-Transformer Network Combining Recurrent-Tokens for Remote Sensing Image Change Detection

Deep learning is developing rapidly and has achieved significant results in the field of remote sensing image change detection. Manual inspection is time-consuming and labor-intensive compared to deep learning, which makes it an inevitable trend to replace manual labor. In this paper, we introduce HTRNet (A HRNet-Transformer Network Combining Recurrent-tokens). Our approach addresses the following challenges: We use HRNet to retain spatial and channel information in image features, avoiding the loss of spatial information that occurs with convolution-based methods. To mitigate the model’s bias caused by non-uniformly distributed bitemporal semantic information, we propose a Recurrent-tokens module to enrich contextual information. Our model generates a binary mask map considering both pixel classification and position. We use the Cosine-embedding loss to measure similarity between the generated mask and ground truth. Experimental results on LEVIR-CD & DSIFN-CD datasets demonstrate that HTRNet outperforms SOAT methods in various metrics. Additionally, our model exhibits smoother edges and robustness in predictions. In summary, HTRNet effectively addresses key challenges in change detection, achieving superior performance compared to existing methods.