Better Data for Satellite Super Resolution

Miguel Castells, Jules Salzinger, Oliver Zendel

Keywords: super-resolution, satellite imagery, Sentinel-2, CLIP embeddings, cross-sensor learning, S2-NAIP dataset, image fidelity, remote sensing, data alignment, near-infrared

TL;DR: We improve cross-sensor super-resolution for satellite imagery by evaluating CLIPScore models, pretraining, and extending the S2-NAIP dataset.

Abstract: Reliable satellite data is needed for many large-scale tasks in urban planing, agriculture, and disaster relief. However, high resolution satellite data is restricted or expensive. ESA's Sentinel-2 provides free satellite data with global coverage but only at a coarse level of detail. In this work we use super-resolution models trained to create high-resolution versions of Sentinel-2 data. We compare the feasibility of various CLIP embeddings to evaluate similarity between hallucinated satellite data and extend the existing S2-NAIP dataset. We automatically clean unreliable data and add new NIR band data. Our experiments show clear improvement in fidelity and quality of single image cross-sensor super resolution for satellite images.

Submission Number: 88