Spatiotemporal Fusion via Conditional Diffusion Model

Spatiotemporal fusion aims to reconstruct sequence remote sensing images in an economically efficient way, for which we observe that the sensor and scale errors can approach the distribution of Gaussian noise. To model the random noise, a spatiotemporal fusion method based on a conditional diffusion model is proposed. A new encoder-decoder network is designed to fuse multisource images. The new model learns the noise distribution at the forward diffusion stage and employs an iterative removal of the noise at the backward diffusion stage, which enhances the model against the Gaussian noise. The proposed method is evaluated on two datasets and compared with seven state-of-the-art algorithms, in which the average root mean square errors (RMSEs) decrease from 0.0198 to 0.0188 for Landsat-7 and from 0.0155 to 0.0141 for Landsat-5, respectively. The experimental results also demonstrate that the proposed method can preserve clearer details and adapt better to abrupt phenological changes.