CIG-STF: Change Information Guided Spatiotemporal Fusion for Remote Sensing Images

Spatiotemporal fusion has been attracting increasing attention in remote sensing applications, such as environmental monitoring and land cover change detection, due to its excellent ability to obtain high spatial and temporal resolution images. The land cover change has always been a great challenge in spatiotemporal fusion. Although most spatiotemporal fusion methods have demonstrated satisfactory performance in addressing phenological changes, the performance in terms of abrupt land cover type changes, such as floods or mudslides, falls short. To alleviate this issue, we propose a change information guided spatiotemporal fusion (CIG-STF) method. The proposed CIG-STF integrates change detection and spatiotemporal fusion in a unified framework, by taking advantage of change detection in capturing land cover changes to assist spatiotemporal fusion. Specifically, the CIG-STF comprises three modules: multiscale dilated feature extractor module (MDFE), spatiotemporal fusion-change detection integrated module (STF-CD), and reconstruction module. The MDFE employs multiscale dilated convolutions to comprehensively extract features, to prevent crucial information loss by increasing the convolutional receptive field. In the STF-CD, a change detection module on attention strategy is integrated into the spatiotemporal fusion task, by excavating land cover changes to further enhance the fusion performance. In addition, we design a dynamic decay loss function to further leverage change information, ensuring the accuracy of both change information and prediction results. The experiments were verified on the publicly available LGC and Daxing datasets with manual change labels. The experimental results demonstrate the superior performance of the proposed CIG-STF in both phenological variations and land cover type changes.