Optical Image Translation Using Diffusion Models in Support of Heterogeneous Change Detection

We propose a novel deep learning-based method that adapts the domains of images acquired by different remote sensing sensors. It adapts a lower resolution image to the domain of an an higher resolution targeted sensor. This is effective in the case of change detection, where differences between sensors, such as spatial resolution and radiometry, can hinder the detection performance and where model hallucination artifacts are unwanted. The proposed technique divides the input image into patches and uses a diffusion-based model to generate translated patches in the style of the target sensor. The translated patches are stitched together to form the output image, which provides global generative consistency. Our approach can handle images with different resolutions and tonalities. We show its effectiveness on a Sentinel-II + Planet Dove data set and demonstrate its high generation quality and contribution to enhance change detection performance.