Spatiotemporal Fusion of Satellite Images via Very Deep Convolutional Networks

Spatiotemporal fusion provides an effective way to fuse two types of remote sensing data featured by complementary spatial and temporal properties (typical representatives are Landsat and MODIS images) to generate fused data with both high spatial and temporal resolutions. This paper presents a very deep convolutional neural network (VDCN) based spatiotemporal fusion approach to effectively handle massive remote sensing data in practical applications. Compared with existing shallow learning methods, especially for the sparse representation based ones, the proposed VDCN-based model has the following merits: (1) explicitly correlating the MODIS and Landsat images by learning a non-linear mapping relationship; (2) automatically extracting effective image features; and (3) unifying the feature extraction, non-linear mapping, and image reconstruction into one optimization framework. In the training stage, we train a non-linear mapping between downsampled Landsat and MODIS data using VDCN, and then we train a multi-scale super-resolution (MSSR) VDCN between the original Landsat and downsampled Landsat data. The prediction procedure contains three layers, where each layer consists of a VDCN-based prediction and a fusion model. These layers achieve non-linear mapping from MODIS to downsampled Landsat data, the two-times SR of downsampled Landsat data, and the five-times SR of downsampled Landsat data, successively. Extensive evaluations are executed on two groups of commonly used Landsat–MODIS benchmark datasets. For the fusion results, the quantitative evaluations on all prediction dates and the visual effect on one key date demonstrate that the proposed approach achieves more accurate fusion results than sparse representation based methods.