Unified multiscale fusion network for change detection in high-resolution remote sensing images

Change detection in remote sensing images holds significant application value across various fields. Deep learning methodologies have evolved into mainstream strategies for constructing change detection models. However, due to the limits of local receptive fields, convolutional neural network (CNN)-oriented models in change detection often struggle to effectively incorporate the comprehensive structural and semantic information of images. Moreover, they tend to lose information when dealing with data exhibiting intricate contextual interconnections. Consequently, in this research documentation, we present a Unified Multiscale Fusion U-Net, which synergistically combines the strengths of CNNs and transformers to achieve precise change detection within remote sensing imagery. A feature aggregation module is designed to address the issue of detailed and semantic information loss in CNNs, thereby strengthening the capacity to capture global data. We also developed a global spatial attention module to compensate for the transformer’s local perception shortcomings, thus enhancing the capacity to model long-range dependencies and details. In addition, a cross-layer aggregation module is devised to enrich the features and augment the model’s expressive potential, particularly in tasks involving multiscale feature information. A comprehensive range of experiments has been conducted on three diverse change detection datasets. The results indicate that our model achieved F1 scores of 0.9004, 0.9456, and 0.8846 on the LEVIR-CD, CDD, and WHU datasets, respectively, whereas the Kappa scores were 0.8974, 0.9414, and 0.8683. It significantly outperformed other comparative algorithms across various evaluation metrics and detection results, demonstrating outstanding performance in the field of change detection.