Resolution Aware Change Detection: Analysis, Training Strategy and Module Design

Change detection (CD) in remote sensing imagery has become an increasingly substantial task in the monitoring of natural resources and disaster response. However, existing CD methods face a notable challenge in terms of resolution flexibility, wherein the resolution gap during training and testing stages results in a substantial performance decline. To address this issue, this paper first constructs a synthetic multiresolution benchmark through the resampling of existing public CD datasets. Subsequently, the impact of varying resolutions on CD models is systematically analyzed, accompanied by the introduction of a novel multi-resolution training (MRT) strategy aimed at improving the robustness of CD methods. Additionally, a plug-and-play module, termed the resolution pyramid constraint (RPC), is designed to facilitate alignments between multi-resolution features and original single-resolution features. Experiments demonstrate that small resampling factors could improve the performances of conventional CD methods, while large factors cause accuracies sharp decline. MRT and RPC are validated to be effective in enhancing robustness. Code and dataset will be available at https://github.com/flying318/MRCD.git.