EO-WM: A Physically Informed World Model for Probab...

EO-WM: A Physically Informed World Model for Probab... | AI Research

Key Takeaways

EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting Earth Observation (EO) forecasting aims to predict how the Earth's s...
Earth Observation (EO) forecasting aims to predict future Earth surface dynamics from satellite observations under changing meteorological conditions.
EO-WM incorporates a physically informed conditioning framework that represents meteorological forcing through a climatological baseline, weather anomalies, and cumulative physical stress signals.
Specifically, it separates baseline and anomaly through distinct conditioning pathways, and accumulates anomalous forcing over time to capture sustained heat and drought stress.
The benchmarks and model will be made open-source at this https URL .

Paper AbstractExpand

Earth Observation (EO) forecasting aims to predict future Earth surface dynamics from satellite observations under changing meteorological conditions. In this paper, we view this task as a partially observed, weather-driven world modeling problem, in which weather acts as a conditioning signal, while forecasting remains uncertain due to sparse observations and unobserved land-surface states. However, existing methods do not fully capture this setting: deterministic models collapse uncertainty into a single future prediction, while diffusion-based methods typically treat weather variables as undifferentiated conditioning signals, and existing benchmarks focus mainly on reconstruction accuracy rather than whether forecasts respond correctly to changed weather this http URL introduce EO-WM, a video diffusion transformer for multispectral EO forecasting. EO-WM incorporates a physically informed conditioning framework that represents meteorological forcing through a climatological baseline, weather anomalies, and cumulative physical stress signals. Specifically, it separates baseline and anomaly through distinct conditioning pathways, and accumulates anomalous forcing over time to capture sustained heat and drought stress. To evaluate weather-response behavior beyond standard metrics, we introduce two diagnostic benchmarks: an Extreme Summer Benchmark for severity-aware prediction of vegetation degradation under extreme weather, and a Seasonal Matched-Pair Benchmark for testing response fidelity under changed weather forcing. Experiments show that EO-WM reduces the error in predicted Normalized Difference Vegetation Index (NDVI) decline amplitude by a relative 5.63% and improves directional hit rate by a relative 7.80%, while remaining competitive on standard pixel-level metrics. The benchmarks and model will be made open-source at this https URL .

EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting
Earth Observation (EO) forecasting aims to predict how the Earth's surface changes over time based on satellite imagery and changing weather conditions. While current models are good at recreating images, they often struggle to understand the cause-and-effect relationship between weather and land-surface dynamics. This paper introduces EO-WM, a new model that treats EO forecasting as a "world modeling" problem. Instead of just predicting pixels, it learns how the environment responds to weather forcing, allowing it to better simulate how vegetation reacts to events like heatwaves and droughts.

A New Way to Process Weather Data

Most existing models treat weather information as a generic input, failing to distinguish between normal seasonal patterns and extreme events. EO-WM changes this by using a "physically informed" conditioning framework. It breaks down weather data into three distinct components: a climatological baseline (what is expected for that time of year), weather anomalies (the difference from the expected norm), and cumulative physical stress (the buildup of heat or drought over time). By separating these, the model can better understand that vegetation degradation is often the result of sustained environmental stress rather than just a single day of bad weather.

Probabilistic Forecasting for Uncertain Futures

Because satellite observations are often sparse or obscured by clouds, the future state of the land is inherently uncertain. EO-WM uses a video diffusion transformer architecture to address this. By framing the task as a probabilistic model, it can represent multiple plausible futures rather than collapsing uncertainty into a single, potentially inaccurate prediction. This allows the model to handle the gaps in satellite data while maintaining a physically consistent response to the meteorological conditions provided as input.

Evaluating Real-World Response

To ensure the model actually understands the impact of weather, the researchers introduced two new diagnostic benchmarks. The "Extreme Summer Benchmark" tests whether the model can accurately predict the severity of vegetation decline during heat and drought events. The "Seasonal Matched-Pair Benchmark" compares how the model predicts the same location across different years with different weather, testing if the model’s predictions change in the correct direction and magnitude when the weather forcing changes.

Key Findings

Experiments show that EO-WM significantly outperforms existing methods in predicting vegetation health. It reduced the error in predicting the amplitude of vegetation decline by 5.63% and improved the directional hit rate—the ability to correctly predict whether vegetation will improve or decline—by 7.80%. These results demonstrate that by incorporating physical knowledge about how weather drives environmental change, the model achieves a much higher level of "forcing-response fidelity" than models that rely solely on pixel-level reconstruction.