Research map / Time series / Foundation models for time series
Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Terra, a large multimodal spatio-temporal Earth dataset (45 years of hourly/3-hourly, daily, monthly meteorological time series on global grids, plus LLM-generated text descriptions and geo-images), and benchmarks forecasting and spatial prediction baselines on it.
Problem
Existing public spatio-temporal datasets are limited in spatial coverage, temporal span, resolution, and lack multimodal supplementary information, hindering spatio-temporal general intelligence research.
Approach
A global, multi-resolution gridded Earth dataset combining long-span meteorological time series with LLM-derived grid text (using metadata-grounded spatial prompt engineering) and geo-images/satellite imagery, plus benchmark experiments showing current models do not clearly exploit spatio-temporal structure.
Key results
- TimesNet and STID achieved the best and second-best average precipitation forecasting results on the UK, USA, and CN subsets, with spatio-temporal models not outperforming time-series models.
- SatCLIP outperformed GeoCLIP and CSP in location-based prediction, reaching MSE 0.010 for precipitation versus 0.162 for GeoCLIP and 0.997 for CSP.
- UrbanVLP without its street-view branch achieved the highest R2 (e.g., 0.745 for CN precipitation) among CLIP, UrbanCLIP and UrbanVLP in vision-language spatial prediction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SatCLIP location embeddings instead of CSP | Global location-based spatial prediction · Global precipitation prediction | MSE: 0.997 -> 0.010 | Helped |
| Use SatCLIP location embeddings instead of GeoCLIP | Global location-based spatial prediction · Global precipitation prediction | MAE: 0.249 -> 0.052 | Helped |
| Use SatCLIP location embeddings instead of CSP | Global location-based spatial prediction · Global wind speed prediction | MAE: 0.787 -> 0.116 | Helped |
| Use SatCLIP location embeddings instead of GeoCLIP | Global location-based spatial prediction · Global temperature prediction | MSE: 0.021 -> 0.002 | Helped |
| Use UrbanVLP* vision-language features instead of CLIP | Vision-language spatial prediction · CN precipitation prediction | R2: 0.483 -> 0.745 | Helped |
| Use UrbanVLP* vision-language features instead of CLIP | Vision-language spatial prediction · USA wind speed prediction | R2: 0.519 -> 0.750 | Helped |
| Use UrbanVLP* vision-language features instead of UrbanCLIP | Vision-language spatial prediction · AUS temperature prediction | R2: 0.210 -> 0.352 | Helped |
Surprising findings
- Spatio-temporal prediction models did not outperform time-series models, even ConvLSTM designed for precipitation.
- Spatio-temporal models such as STAEformer, GWNet, and STGCN failed on the global setting due to high spatial memory consumption.
Limitations
- Image and text modalities only available at 1° and 0.5° resolutions, not 0.1°.
- LLM-generated text may hallucinate and becomes outdated as LLMs evolve; illusion rate estimated at roughly 30% attribute error.
- Satellite image redistribution restrictions may make the imagery outdated or unstable to share.
Experimental setup
Datasets: MSWX MSWEP Köppen climate classification ETOPO2v2 C3S Global Land Cover GMT geographic images ArcGIS satellite imagery
Benchmarks: Terra precipitation forecasting (Global and five country subsets) Terra location-based spatial variable prediction Terra vision-language spatial variable prediction
Metrics: MAE RMSE MSE R2
Baselines: TimesNet FEDformer PatchTST DLinear STAEformer STID GWNet STGCN ConvLSTM HI SatCLIP GeoCLIP
Builds on: MSWX MSWEP ConvLSTM CLIP
Closest papers by meaning
- Unifying Knowledge from Diverse Datasets to Enhance Spatial-Temporal Modeling: A Granularity-Adaptive Geographical Embedding Approach (ICML 2025 · similarity 0.87)
- No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data (ICLR 2025 · similarity 0.86)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.86)
- SPACETIME: Causal Discovery from Non-Stationary Time Series (AAAI 2025 · similarity 0.86)
- Earthfarsser: Versatile Spatio-Temporal Dynamical Systems Modeling in One Model (AAAI 2024 · similarity 0.85)
- On the Integration of Spatial-Temporal Knowledge: A Lightweight Approach to Atmospheric Time Series Forecasting (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.