Dataset and Benchmark for Thin Cloud Removal of Optical Remote Sensing Images and Beyond
Abstract: Cloud cover poses a significant challenge to the utility of spaceborne optical Earth imagery, greatly limiting the effectiveness of the data for various applications. While there are several publicly available datasets for cloud removal have been introduced by researchers, most of them are confined to specific cloud-removal tasks, raising the question of whether an approach can be effectively applied to remote sensing images obtained under various cloud and different ground object in different time periods and different regions. We present a large novel dataset named We present a large dataset named Cloud ImageRy Remote Sensing Simulation (CIRRUS) and a high-fidelity cloud simulation algorithm to target the challenge of generalization. The dataset simulates cloud and haze formations based on their real-world appearances in images, enhancing the realism of the cloud simulation. We further provide a variety of criteria for cloud removal algorithm evaluation, including full-reference metrics, no-reference metrics, and subjective evaluation. Experiments conducted on CIRRUS shed light on the comparisons and limitations of the state-of-the-art single image remote sensing cloud removal algorithm, and suggest promising future directions. The dataset and simulation source codes will be public soon.
Loading