Pre-Training and Fine-Tuning for Satellite Image-to-Image Translation

Anonymous

Keywords: Sentinel-1, Sentinel-2, pix2pix

TL;DR: NDVI from SAR by Image-to-Image translation with Pre-training

Abstract: Geospatial technologies are increasingly important for various applications worldwide, including vegetation monitoring. Collecting ground truth data for specific geospatial tasks are challenging and time-consuming. Recently, the foundation model research have been explored, then pre-training on large-scale data and fine-tuning for specific tasks are important components of this technique. Although this approach can enhance the performance in downstream tasks such as satellite image translation, directly fine-tuning models pre-trained on natural images like ImageNet is suboptimal for geospatial data due to the inherent domain differences. In this paper, we propose a novel image translation approach with pre-training on geospatial-specific data and data augmentation. We present a case study where our method achieved outstanding results in a competition for inferring the Normalized Difference Vegetation Index from satellite-based Synthetic Aperture Radar data of cabbage farms. Our approach outperformed other methods with a 31% higher score than the second-ranked team and a 44% higher score than the average of the top five teams.