DCFF-Net: A Densely Connected Feature Fusion Network for Change Detection in High-Resolution Remote Sensing Images

Change detection is one of the main applications of remote sensing image analysis. Due to the strong capabilities of neural networks in other fields, a growing number of research works of automatic remote sensing change detection focus on deep learning algorithms. The network architectures of change detection are mostly based on the encoder-decoder architecture. Although the encoder-decoder architecture can acquire high-level semantic information for change detection, it still exists some problems in high-resolution remote sensing images, such as the loss of high-resolution location information during the down-sampling process, insufficient high-resolution information during the up-sampling reconstruction process, and small changes are challenging to detect. To address these issues, we propose a densely connected feature fusion network (DCFF-Net) for change detection. First, we extract the multiscale raw image features by two-stream network architecture with the same weights. At the same time, bitemporal images are concatenated as one input with six channels to generate the change map by a difference extraction network based on encoder-decoder architecture. In order to better reconstruct the edge details of the change map and the changes with the small region, an attention mechanism is employed in each up-sampling process to fuse the previously extracted raw image features with difference features. The deep supervision strategy is adopted to alleviate the problem of gradient vanishing. In addition, a novel weighted loss is proposed by combining self-adjusting dice loss and binary cross-entropy loss to alleviate the data imbalance issue. We perform extensive experiments on two public change detection datasets. The visual comparison and quantitative evaluation confirm that our proposed method outperformsother state-of-the-art methods.