Multi-image Super-Resolution Algorithm Supported on Sentinel-2 Satellite Images Geolocation Error

Abstract

Every year, the hottest seasons are marked by forest fires. Monitoring these forest areas is more effective with the help of satellite imagery, since ground operations are hampered by vegetation density and height, making them less productive and more expensive. However, nowadays, freely available imagery from Sentinel-2 satellite has a maximum spatial resolution of 10 meters per pixel, a low resolution to identify small or thin structures in the image, such as roads, bridges, buildings, rivers, fuel breaks, among others.To improve image’s resolution, a new super-resolution algorithm, named KGEONP – K Geographically Nearest Pixels, is proposed in this paper. It benefits from Sentinel-2 regular observations (it has a revisit of 5 days) and the georeferencing error of its images (whose maximum value is 1.5 pixels). KGEONP seeks to add as much information as possible to the super-resolved image, by using data from K-nearest pixels and their spatial distance for computing each new pixel’s value.KGEONP was applied to Sentinel-2 images to increase resolution by a factor of 10 and was compared to state-of-the-art super-resolution techniques. It showed quite satisfactory results, with the capacity of increasing resolution and maintaining the structural data of the source images.