Research map / Probabilistic and scientific modeling / Dynamical systems
Earthfarsser: Versatile Spatio-Temporal Dynamical Systems Modeling in One Model
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.
Proposes EarthFarseer, a spatio-temporal forecasting framework combining parallel local convolutions with a global Fourier-based transformer (FoTF) for space and a Fourier/multi-scale convolutional temporal module (TeDev) with a flexible-length decoder, evaluated on eight physical and social dynamics datasets.
Problem
Spatio-temporal dynamical system forecasting suffers from weak local fidelity, poor long-horizon prediction, low scalability across tasks and resolutions, and inefficient training.
Approach
Parallel local CNN and global Fourier-based token-mixing transformer branches with up/down-sampling interaction for space, plus a Fourier-domain multi-scale fully convolutional temporal module and a two-stage decoder that can output arbitrary target lengths and resolutions.
Key results
- Achieves best MSE on MovingMNIST (14.9 vs. 22.9 for the next best Vid-ODE) and lowest MAE on TaxiBJ+ (2.1 vs. 3.0 for SimVP).
- On SEVIR, reports CSI-M of 47.1 versus 44.2 for Earthformer.
- Reports best average ranking of 1.7 across the eight datasets; on 2DSWE it saves nearly 3/4 of training time compared to Vid-ODE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the local convolutional component (ablation) | EarthFarseer · TaxiBJ+ | MAE: 2.8 -> 2.1 | Helped |
| add the global Fourier-based Transformer (FoTF) (ablation) | EarthFarseer · TaxiBJ+ | MAE: 2.4 -> 2.1 | Helped |
| add TeDev (ablation) | EarthFarseer · MovingMNIST | MSE: 22.1 -> 14.9 | Helped |
| use the full decoder instead of a linear convolutional decoder (ablation) | EarthFarseer · RD | MSE: 10.2 -> 9.4 | Helped |
| use TeDev instead of ViT (ablation) | EarthFarseer · MovingMNIST | MSE: 23.5 -> 14.9 | Helped |
| use TeDev instead of SwinT (ablation) | EarthFarseer · MovingMNIST | MSE: 21.3 -> 14.9 | Helped |
Limitations
- Claimed scalability and efficiency rely on comparisons under the authors' own settings; ablation and scaling analyses are limited to few datasets.
- Some baseline entries in the results table are missing (e.g., '- -' for Rainformer and Earthformer on TaxiBJ+ and PD-like metrics), limiting full comparison.
Experimental setup
Datasets: MovingMNIST TaxiBJ+ KTH SEVIR RainNet Pollutant-Diffusion (PD) Reaction-Diffusion (RD) 2DSWE
Benchmarks: MovingMNIST TaxiBJ+ KTH SEVIR RainNet PD RD 2DSWE
Metrics: MSE MAE MAPE CSI-M RMSE
Baselines: ConvLSTM PredRNN-v2 E3D-LSTM SimVP ViT Swin Transformer Rainformer Earthformer PhyDNet Vid-ODE PDE-STD FourCastNet
Builds on: GFNet Earthformer SimVP
Closest papers by meaning
- Learning Spatiotemporal Dynamical Systems from Point Process Observations (ICLR 2025 · similarity 0.86)
- No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data (ICLR 2025 · similarity 0.85)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.85)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.85)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.85)
- Seahorse: A Unified Benchmarking Framework for Spatiotemporal Event Modeling (NeurIPS 2026 · similarity 0.84)
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.