SLFFormer: A Network for Remote Sensing Change Detection with Spatial-Temporal Feature Fusion and Local Residual Learning

Detecting changes in images captured at various times of the same geographical region, known as Remote Sensing Change Detection (RSCD), is crucial for urban expansion, agricultural oversight, environmental and disaster tracking, and land use evolution. The progress in remote sensing technology, particularly the access to high-resolution satellite imagery, along with the evolution of deep learning methods, especially convolutional neural networks (CNNs) and transformers, has considerably improved RSCD's effectiveness. Nonetheless, RSCD encounters challenges like the requirement for substantial computational power, intricate data patterns, and the struggle to attain high precision and prompt execution in varied and dynamic conditions. To surmount these obstacles, we introduce the SLFFormer model with three key innovations: the Spatial-Temporal Feature Fusion (STFF) module for enhanced detection of changes in dual-temporal images; the Local Residual Learning (LRL) module to focus on significant changes by filtering out less relevant details; and the Feature Attention Fusion (FAF) module to dynamically emphasize important features and pixels, improving pattern representation. Our model outperforms several state-of-the-art change detection models on the LEVIR-CD dataset in terms of Precision, Recall, F1 Score, IoU, and OA, validating its effectiveness.