TiMo: Spatiotemporal Foundation Model for Satellite Image Time Series

TL;DR

TiMo tackles the challenge of learning robust multiscale spatiotemporal representations from satellite image time series by introducing a four-stage hierarchical vision transformer equipped with spatiotemporal gyroscope attention. It couples this attention with a more efficient differential variant and trains on a new MillionST dataset using spatiotemporal masked image modeling, enabling strong generalization across diverse downstream tasks. Empirical results across deforestation monitoring, land cover segmentation, crop type classification, and flood assessment demonstrate state-of-the-art performance, as well as favorable data efficiency and scalability. This work advances SITS foundation models, offering a practical, scalable approach for comprehensive Earth observation analysis.

Abstract

Satellite image time series (SITS) provide continuous observations of the Earth's surface, making them essential for applications such as environmental management and disaster assessment. However, existing spatiotemporal foundation models rely on plain vision transformers, which encode entire temporal sequences without explicitly capturing multiscale spatiotemporal relationships between land objects. This limitation hinders their effectiveness in downstream tasks. To overcome this challenge, we propose TiMo, a novel hierarchical vision transformer foundation model tailored for SITS analysis. At its core, we introduce a spatiotemporal gyroscope attention mechanism that dynamically captures evolving multiscale patterns across both time and space. For pre-training, we curate MillionST, a large-scale dataset of one million images from 100,000 geographic locations, each captured across 10 temporal phases over five years, encompassing diverse geospatial changes and seasonal variations. Leveraging this dataset, we adapt masked image modeling to pre-train TiMo, enabling it to effectively learn and encode generalizable spatiotemporal representations.Extensive experiments across multiple spatiotemporal tasks-including deforestation monitoring, land cover segmentation, crop type classification, and flood detection-demonstrate TiMo's superiority over state-of-the-art methods. Code, model, and dataset will be released at https://github.com/MiliLab/TiMo.

Figures (10)

• Figure 1: Overview of TiMo's Architecture: TiMo follows a four-stage hierarchical design, where each stage comprises attention layers and feed-forward networks following a downsampling layer. To enhance spatio-temporal representation learning, TiMo replaces Multi-head Self-Attention (MHSA) with a novel and efficient Differential Spatiotemporal Gyroscope Attention (D-STGA) in the first two stages.
• Figure 2: Illustration of Spatiotemporal Gyroscope Attention.
• Figure 3: Diagram of calculating spatial similarity in D-STGA.
• Figure 4: City distribution for MillionST data sampling.
• Figure 5: Experimental results of different SITS FMs trained with varying sample sizes on the KuroSiwo dataset.