Satellite Image Analysis by Pre-training Network

Abstract

Satellite image analysis is a crucial method for understanding climate change and vegetation growth. However, the relatively long acquisition periods of satellite images and the impact of weather conditions can make it challenging to determine accurate values related to vegetation and other factors. Therefore, this paper proposes a method that performs pre-training using a foundational model to convert synthetic aperture radar images into normalized difference vegetation index images, taking advantage of data from satellites other than the target satellite. Evaluation was conducted in a competition applied to vegetation understanding in a cabbage farm. As a result, our method achieved 1st place among 57 participants, with a 31% improvement in error compared to the 2nd-ranked method.