Research map / Time series / Long-term forecasting
DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space
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 DAWP, a framework that runs AI weather prediction directly in satellite observation space, initialized by a masked multimodal autoencoder for observation assimilation and forecasted on sub-images with cross-regional boundary conditioning.
Problem
Weather prediction models trained on reanalysis data inherit data assimilation biases and temporal lag; direct observation prediction from irregular, high-resolution satellite data is hard because sparse, irregular inputs mismatch dense outputs during rollout.
Approach
Impute irregular satellite observations into a uniform dense observation space with a masked multi-modal MAE (AIDA initialization) over mask ViT-VAE tokens, then forecast on sub-images using a spatiotemporal decoupling transformer conditioned on neighbouring sub-image states from a global state cache (cross-regional boundary conditioning).
Key results
- DAWP's 0-12h MAE on AMSU-A ch0 is 1.92 (1e-5 units) versus 2.67 for Transformer-DOP and 2.93 for EarthNet.
- On precipitation at 12h, DAWP reaches CSI-0.5 of 0.197 vs 0.136 for Transformer-DOP (about 44.8% higher).
- Removing AIDA initialization made training loss jump at about 20k steps to MSE around 1.0, while the AIDA-initialized model converged below 0.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| initialize Transformer-DOP with AIDA (ablation) | Transformer-DOP · satellite observation forecasting, 0-72h rollout | MAE | Helped |
| remove AIDA initialization (ablation) | DAWP · observation-space forecasting training | MSE loss: less than 0.1; 1.0 | Hurt |
| add cross-regional boundary conditioning (ablation) | DAWP · center-area prediction | convergence loss: 0.128 -> 0.106 | Helped |
| mask the ViT-VAE attention computation (ablation) | ViT-VAE · AMSU-A and HIRS reconstruction | reconstruction error: AMSU-A: 0.92 -> 0.78; HIRS: 4.28 -> 4.11 | Helped |
| mask the ViT-VAE attention computation (ablation) | ViT-VAE · ATMS and MHS reconstruction | reconstruction error: ATMS: 1.36 -> 1.29; MHS: 2.45 -> 2.41 | Hurt |
| drop any one satellite modality before AIDA (ablation) | DAWP · multistep satellite observation prediction | relative MAE error ratio | Hurt |
| keep only ATMS observations before AIDA (ablation) | DAWP · multistep satellite observation prediction | relative MAE error ratio | Helped |
| use DAWP instead of Transformer-DOP | direct observation forecasting · AMSU-A ch0, 0-12h | MAE: 2.67 -> 1.92 | Helped |
| use DAWP instead of Transformer-DOP | direct observation forecasting · AMSU-A ch0, 24-36h | MAE: 4.91 -> 3.66 | Helped |
| use DAWP instead of EarthNet | global precipitation forecasting · TCWV, 12h, CSI-30 | CSI: 0.786 -> 0.807 | Helped |
| use DAWP instead of Transformer-DOP | global precipitation forecasting · SP, 12h, CSI-0.5 | CSI: 0.136 -> 0.197 | Helped |
Surprising findings
- Without AIDA initialization, training of the observation forecaster becomes unstable and the loss rises sharply after about 20k steps.
Limitations
- Only satellite observations are used; the framework has not been tested with other observation sources such as station data.
- Baselines EarthNet and Transformer-DOP were reimplemented by the authors because no code was released, so comparisons depend on their reproduction.
- No error bars or statistical significance tests are reported.
Experimental setup
Datasets: AMSU-A ATMS HIRS MHS ATMS-Precipitation
Benchmarks: ATMS observation forecasting (AMSU-A, ATMS, HIRS, MHS channels) ATMS-Precipitation (TCWV, SP)
Metrics: MAE CSI FAR
Baselines: Persistence EarthNet Transformer-DOP ConvLSTM PredRNN RainFormer EarthFormer SimVP
Builds on: Masked autoencoder (MAE) MultiMAE Transformer-DOP EarthNet
Closest papers by meaning
- DiffDA: a Diffusion model for weather-scale Data Assimilation (ICML 2024 · similarity 0.86)
- Towards a Self-contained Data-driven Global Weather Forecasting Framework (ICML 2024 · similarity 0.86)
- Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution (CVPR 2025 · similarity 0.86)
- FNP: Fourier Neural Processes for Arbitrary-Resolution Data Assimilation (NeurIPS 2024 · similarity 0.84)
- FlowDAS: A Stochastic Interpolant-based Framework for Data Assimilation (NeurIPS 2025 · similarity 0.84)
- Align-DA: Align Score-based Atmospheric Data Assimilation with Multiple Preferences (NeurIPS 2025 · 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.