Foundation Models in Remote Sensing: Insights from Multispectral and Hyperspectral Self-Supervised Learning

Abstract

Self-supervised learning (SSL) has triggered a paradigm shift in computer vision and remote sensing, enabling the development of foundation models that generalize across diverse downstream tasks with minimal or no fine-tuning. This talk will be structured in three parts. The first part provides a concise overview of SSL in remote sensing and its applications. The second part discusses a use case of SSL-pretrained models for forest monitoring, focusing on practical aspects for semantic segmentation problems: foundation models vs. specialized models, inference cost, and the importance of qualitative evaluation of model outputs. The final part introduces SpectralEarth, a large-scale dataset designed for pretraining hyperspectral foundation model, and its potential in advancing hyperspectral and multi-sensor SSL.