A novel spatio-temporal fusion approach combining deep learning downscaling and FSDAF method

Flexible spatio-temporal data fusion (FSDAF) is usually used to fuse high spatial resolution images with ordinary up-sampling methods processed low spatial resolution images. However, ordinary up-sampling methods lead to spatial inconsistency between high and low spatial resolution images, as well as the presence of many mixed pixels in the low spatial resolution images, which reduces the fusion accuracy. In this study, a novel method by combining deep learning downscaling and the FSDAF is proposed. This method (RCAN-FSDAF) firstly downscales low spatial resolution images by using residual channel attention network (RCAN), and then fuses the high spatial resolution images and the downscaled low spatial resolution images by using FSDAF to finally obtain high spatio-temporal resolution data. The results show that RCAN-FSDAF presents several advantages comparing with the conventional FSDAF method. Firstly, the band reflectance predicted by RCAN-FSDAF is closer to the base reflectance than that predicted by FSDAF, suggests its higher correlation and smaller errors. Secondly, RCAN-FSDAF is superior in decomposing image into different features, accurately identifying boundaries between different features, and detecting changes in land cover types. Lastly, the high spatio-temporal resolution NDVI data, obtained from the prediction results of RCAN-FSDAF, has higher accuracy.