DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space

AI summary

Overview

Research area: Machine learning for weather forecasting and Earth observation, specifically direct satellite observation prediction (AI-DOP).

Technical level: Advanced — assumes familiarity with transformers, variational autoencoders, masked autoencoders, and meteorological data assimilation concepts, though the summary below explains them in plain terms.

Scope: The paper introduces DAWP, a two-stage framework that first fills in the gaps in sparse satellite observations and then forecasts them forward in time, sidestepping the reanalysis datasets that conventional AI weather models depend on.

What This Paper Is About

Most modern AI weather forecasting models are trained on reanalysis data — gridded, physically consistent snapshots of the atmosphere produced by blending raw observations with physics-based models. That blending introduces biases and a time lag of up to six hours, and it also resamples raw satellite measurements down to a coarse regular grid, discarding information. This paper asks whether an AI model can instead learn to forecast the atmosphere directly in the space of raw satellite observations, and answers yes by first "completing" the sparse, irregular observation field and then learning its dynamics.

Key Contributions

1. An end-to-end observation-space forecasting framework (DAWP) that combines an Artificial Intelligence Data Assimilation (AIDA) module with an AI weather prediction (AIWP) module, removing the dependency on reanalysis products altogether.

2. AIDA: a multi-modal masked autoencoder for observation-space data assimilation. AIDA imputes missing satellite measurements by attending across sensors and across neighbouring times, transforming irregular, partially observed data into a uniform, dense observation grid suitable for standard spatiotemporal learning.

3. A mask ViT-VAE encoder/decoder for multi-channel satellite data that ignores patches with insufficient observations via masked attention, compresses high-channel-count satellite imagery better than natural-image VAEs, and doubles as a precipitation mapping head.

4. A spatiotemporal decoupling transformer with Cross-regional Boundary Conditioning (CBC) that forecasts on sub-images while querying a global state cache for neighbouring regions, enabling efficient global rollout at 12 × 1152 × 2304 resolution, plus a 35 TB composite multi-sensor satellite dataset assembled for the task.

Main Findings

• A 12-hour lead-time advantage over existing methods. On AMSU-A channels, DAWP's error at 24–36 hours (3.66 and 5.80 MAE) beats EarthNet and Transformer-DOP's errors at 12–24 hours (3.84/6.14 and 4.12/6.65), meaning DAWP at 36 hours is roughly as accurate as baselines at 24 hours.

• Baselines degrade into persistence; DAWP does not. EarthNet is overtaken by the naive persistence baseline at around 16 hours of lead time, while DAWP holds out until roughly 64 hours — a large gap in usable forecast horizon.

• AIDA initialization is essential, not cosmetic. Without it, training loss diverges to ~1.0 MSE after about 20k steps. With it, loss converges below 0.1 after 200k steps. Applying AIDA to the competing Transformer-DOP model also substantially improves its multi-step rollout, indicating the benefit generalizes.

• Cross-regional Boundary Conditioning improves both accuracy and continuity. Adding neighbouring-region context cuts the center-area loss by about 15.6% and visually removes discontinuities at sub-image seams in rollout predictions.

• Precipitation forecasting shows clear promise. On 0–12 hour forecasts, DAWP reaches 0.807 CSI at the 30 mm total-column-water-vapor threshold versus 0.786/0.789 for baselines, with markedly lower false alarm ratios. On surface precipitation it improves CSI-0.5 and CSI-1.0 over Transformer-DOP by 44.8% and 78.9%, and is the only AI-DOP method that stays above the persistence baseline at the 2.0 mm/h threshold.

• Sensor importance is uneven. Dropping any sensor degrades accuracy, but ATMS carries the most predictive information (keeping only ATMS yields the lowest error ratios), AMSU-A carries the least spatiotemporal dynamics, and HIRS and MHS contain substantial redundancy. Errors follow a periodic pattern tied to the strip-scanning geometry of polar-orbiting satellites.

Methodology in Plain English

The framework runs in two stages.

Stage one — filling in the blanks. Polar-orbiting satellites only see a narrow swath beneath them at any moment, so at any given time most of the globe is unobserved. The researchers take a 12-hour window of satellite data, chop it into 144 × 144 sub-images, and encode each sensor's observations into tokens using a vision transformer that is told which patches are missing. All sensors' tokens are concatenated and fed into a masked autoencoder, which is trained by hiding some of the tokens it does have and learning to reconstruct them. Because reconstructing a token requires reasoning about nearby tokens in space and time, the trained model can also fill in genuinely absent areas. The result is a complete, uniform, multi-sensor observation field — this is the "assimilation" step.

Stage two — forecasting. With inputs and outputs now living in the same dense space, forecasting becomes a standard spatiotemporal prediction problem. The model splits attention into separate temporal and spatial passes (cheaper than full joint attention) and processes one sub-image at a time. To avoid seams between sub-images, each prediction is conditioned on tokens from its neighbouring regions, pulled from a global state cache that is updated after every 12-hour prediction step. This makes multi-day global rollout tractable without ever materializing the full globe in memory at once.

Why this design matters. Training directly on sparse observations fails because the input distribution (mostly empty) and the output distribution (fully dense) do not match — a problem the authors call input-output distribution shift. Filling in the blanks first removes that mismatch entirely.

Why This Matters

Research impact. The paper challenges the field's default assumption that reanalysis data is the natural training substrate for AI weather models. By demonstrating a working alternative that is trained purely on satellite radiances, it opens a route to models that respond in near-real-time and avoid inheriting the biases of numerical assimilation systems. The AIDA component is also a self-supervised, sensor-agnostic assimilation method, which is conceptually distinct from the graph- and mask-token approaches used by Transformer-DOP and Graph-DOP.

Real-world applications:

• Faster severe weather response. Because observations are available in near-real-time, a direct-observation model can begin forecasting hours before a reanalysis-based model would even have its inputs.
• Global precipitation monitoring. The demonstrated precipitation mapping gives a rapid-response path from raw microwave radiances to rainfall estimates, useful for flood and drought monitoring in regions with sparse ground radar.
• Wildfire and sea-ice tracking. The authors note the framework can plug in any downstream "retrieval operator," making burned-area detection and sea-ice mapping natural extensions.
• Earth system modeling. Because the framework absorbs arbitrary observation modalities, it could serve as a unifying backbone across atmospheric, oceanic, and land-surface measurements.

Industry relevance. Forecasting agencies, satellite operators, reinsurance and agricultural analytics firms, and aviation and logistics companies all depend on forecast latency and accuracy. A model that produces skillful forecasts directly from satellite feeds without a numerical assimilation pipeline in the loop has obvious operational appeal — fewer moving parts, faster turnaround, and a cleaner path to ingesting new instruments as they launch.

Future Directions

1. Broadening beyond satellite data. The authors explicitly flag homogeneous observation sources as the main limitation and plan to integrate ground station data and other in-situ measurements, which would require handling fundamentally different data geometries.

2. Direct prediction of physical variables. The paper raises the possibility of forecasting physical weather variables directly rather than only radiances, by incorporating station observations into the framework.

3. Extending to more downstream retrieval tasks. The precipitation mapping is presented as one instance of a general pattern; surface parameter estimation, wildfire monitoring, and sea-ice mapping are named as targets.

4. Understanding the modality trade-offs. The ablation showing ATMS dominance and HIRS/MHS redundancy invites a systematic study of how many sensors are actually needed, which has cost implications for future satellite missions and data pipelines.

Target Audience

This paper is most valuable to machine learning researchers working on spatiotemporal modeling and self-supervised representation learning, and to meteorological researchers and operational forecasters who want to understand where AI weather prediction is heading beyond reanalysis-trained models. Practitioners building Earth observation pipelines — particularly those working with satellite radiances, remote sensing, or precipitation products — will find the architecture and the 35 TB dataset construction details directly applicable. Readers should have some grounding in transformers and generative models to follow the method section comfortably, though the problem framing and results are accessible to a broader technical audience.

Authors’ abstract

Weather prediction is a critical task for human society, where impressive progress has been made by training artificial intelligence weather prediction (AIWP) methods with reanalysis data. However, reliance on reanalysis data limits the AIWPs with shortcomings, including data assimilation biases and temporal discrepancies. To liberate AIWPs from the reanalysis data, observation forecasting emerges as a transformative paradigm for weather prediction. One of the key challenges in observation forecasting is learning spatiotemporal dynamics across disparate measurement systems with irregular high-resolution observation data, which constrains the design and prediction of AIWPs. To this end, we propose our DAWP as an innovative framework to enable AIWPs to operate in a complete observation space by initialization with an artificial intelligence data assimilation (AIDA) module. Specifically, our AIDA module applies a mask multi-modality autoencoder(MMAE)for assimilating irregular satellite observation tokens encoded by mask ViT-VAEs. For AIWP, we introduce a spatiotemporal decoupling transformer with cross-regional boundary conditioning (CBC), learning the dynamics in observation space, to enable sub-image-based global observation forecasting. Comprehensive experiments demonstrate that AIDA initialization significantly improves the roll out and efficiency of AIWP. Additionally, we show that DAWP holds promising potential to be applied in global precipitation forecasting.