Remote Sensing Foundation Models

• RSFMs are large-scale neural networks leveraging self-supervised and multimodal training to extract robust feature representations from diverse remote sensing data.
• They enable rapid adaptation to downstream tasks such as classification, semantic segmentation, and change detection with minimal additional supervision.
• The models address computational bottlenecks by emphasizing efficiency, modular design, and domain specialization tailored to the unique constraints of geospatial data.

Remote Sensing Foundation Models (RSFMs) are large-scale neural network architectures designed to extract general-purpose, transferable feature representations from the diverse, high-dimensional, and often multi-modal data characteristic of Earth observation applications. These models synthesize advances in self-supervised learning, multimodal fusion, and scalable deep learning, and are engineered to support rapid adaptation to downstream remote sensing (RS) tasks—including classification, semantic segmentation, change detection, object detection, and beyond—with minimal additional supervision. Recent RSFM research explicitly seeks to circumvent the computational and accessibility bottlenecks of web-scale found in text and vision foundation models by designing efficient, modular, and domain-specialized approaches optimized for the unique constraints and opportunities of geospatial data (Adorni et al., 26 Nov 2025, Zhang et al., 18 Jul 2025, Bi et al., 4 Apr 2025, Xiao et al., 22 Oct 2024).

1. RSFM Definition, Motivation, and Taxonomy

RSFMs are defined as deep learning architectures pre-trained on large, heterogeneous collections of remote sensing data using either self-supervised or multi-task objectives, yielding a frozen encoder $f_\theta: \mathcal{X