SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery

TL;DR

SkySense addresses the need for a universal RSFM by jointly modeling multi-modal temporal RSI and geo-context. It introduces a Factorized Multi-Modal Spatiotemporal Encoder, Multi-Granularity Contrastive Learning, and Geo-Context Prototype Learning, pre-trained on $21.5$M sequences to yield a $2.06$B-parameter model. Across $16$ datasets and $7$ tasks, SkySense achieves state-of-the-art results and outperforms $18$ RSFMs, with notable gains in both single- and multi-modal scenarios. The model is designed for flexible downstream use, and pre-trained weights are intended for public release to accelerate Earth observation research.

Abstract

Prior studies on Remote Sensing Foundation Model (RSFM) reveal immense potential towards a generic model for Earth Observation. Nevertheless, these works primarily focus on a single modality without temporal and geo-context modeling, hampering their capabilities for diverse tasks. In this study, we present SkySense, a generic billion-scale model, pre-trained on a curated multi-modal Remote Sensing Imagery (RSI) dataset with 21.5 million temporal sequences. SkySense incorporates a factorized multi-modal spatiotemporal encoder taking temporal sequences of optical and Synthetic Aperture Radar (SAR) data as input. This encoder is pre-trained by our proposed Multi-Granularity Contrastive Learning to learn representations across different modal and spatial granularities. To further enhance the RSI representations by the geo-context clue, we introduce Geo-Context Prototype Learning to learn region-aware prototypes upon RSI's multi-modal spatiotemporal features. To our best knowledge, SkySense is the largest Multi-Modal RSFM to date, whose modules can be flexibly combined or used individually to accommodate various tasks. It demonstrates remarkable generalization capabilities on a thorough evaluation encompassing 16 datasets over 7 tasks, from single- to multi-modal, static to temporal, and classification to localization. SkySense surpasses 18 recent RSFMs in all test scenarios. Specifically, it outperforms the latest models such as GFM, SatLas and Scale-MAE by a large margin, i.e., 2.76%, 3.67% and 3.61% on average respectively. We will release the pre-trained weights to facilitate future research and Earth Observation applications.

Figures (16)

• Figure 1: SkySense has achieved superior performance on 16 datasets over 7 distinct tasks compared with 18 state-of-the-art RSFMs and supports a board range of EO imagery interpretations.
• Figure 2: The overview of our SkySense model architecture.
• Figure 3: Overview of SkySense pre-training and downstream usage. SkySense employs data augmentations on the input and then feeds the augmented data into the student and teacher networks respectively. Multi-Granularity Contrastive Learning and Cross-Modal Alignment are proposed to pre-train the overall network. The region-specific prototype set $\mathcal{P}$ is learned on the student branch and it is frozen for downstream usage. Enhancing feature with $\mathcal{P}$ is optional. After pre-training, we adopt the parameters of the teacher branch for downstream tasks. Each pre-trained module can be used alone or combined with the others, with the chosen ones either frozen or fine-tuned.
• Figure 4: (a) Experiment on fine-tuning using different percentages of training data on the AID dataset. (b) The impact of S1-Ts data under varying cloud coverage conditions.
• Figure 5: Comparison between (a) ESRI LandCover Map and (b) Geo-Context Prototype. The visualization process of Geo-Context Prototype is illustrated in the upper part of this figure.