A Novel Remote Sensing Spatiotemporal Data Fusion Framework Based on the Combination of Deep-Learning Downscaling and Traditional Fusion Algorithm

Traditional remote sensing spatiotemporal data fusion algorithms generally use upsampled low-resolution images (MODIS) to be fused with high-resolution images (Landsat); this makes both images less spatially consistent and many hybrid image elements in low-resolution images, so uncertainty errors propagate into the fusion results. To address this issue, we propose a framework for combining deep-learning-based super-resolution techniques with traditional spatiotemporal fusion methods. By reconstructing low-resolution images using super-resolution image reconstruction techniques, we obtain low-resolution images with more spatial details and better spatial consistency with high-resolution images. These reconstructed images are then fused using spatiotemporal fusion methods. In this study, we selected flexible spatiotemporal data fusion (FSDAF) and residual channel attention network (RCAN) to carry out a detailed study to prove the effectiveness of this kind of framework. That is, a new RCAN-FSDAF model is developed. After testing, RCAN-FSDAF has the following advantages: First, the band reflectance predicted by RCAN-FSDAF is closer to base reflectance than FSDAF, DMNet, and GAN-STFM, as shown by greater correlation and smaller error. Second, RCAN-FSDAF better decomposes image elements among heterogeneous features and more accurately identifies boundaries between different features and changes in land-cover type. Third, high spatial and temporal resolution NDVI data obtained by the inversion of the prediction results of RCAN-FSDAF are more accurate. The framework developed in this study can be extended to other spatial and temporal data fusion applications.