EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting
Junwei Luo, Shuai Yuan, Zhenya Yang, Yansheng Li, Zhe Liu, Hengshuang Zhao
Read on arXiv →
Key claim
EO-WM improves vegetation forecasting under changing weather conditions.
In plain English
Imagine you're trying to predict how vegetation will respond to changing weather conditions using satellite data. Traditionally, models either make a single prediction that doesn't account for uncertainty or treat weather data too simplistically, leading to inaccurate forecasts. This is problematic because it can result in poor decision-making in agriculture, disaster response, and environmental management. This is what's called a failure to capture the complexity of weather impacts on land surfaces.
The approach in this paper, EO-WM, takes a fresh look at this problem by treating weather as a key factor that influences vegetation dynamics. Instead of just making one prediction, it uses a video diffusion transformer that incorporates detailed weather information, separating normal conditions from anomalies and accumulating stress signals over time. This allows the model to better understand how prolonged weather changes affect vegetation health.
What sets EO-WM apart from previous methods is its focus on how well forecasts respond to actual weather changes, rather than just how accurately they reconstruct past data. The authors introduce new benchmarks to evaluate this response behavior, leading to significant improvements in prediction accuracy. Practically, this means that if you're working in fields like agriculture or environmental monitoring, using EO-WM could lead to better-informed decisions based on more reliable forecasts.
The paper introduces a new approach to Earth Observation forecasting that significantly extends existing methods by incorporating a physically informed conditioning framework.
The claims are supported by experiments showing improved performance on multiple benchmarks, with a clear evaluation of weather-response behavior.
Deep reliability assessment
The methodology supports the claim that physically structured weather conditioning and repeated condition injection improve EO-WM on EarthNet2021-derived seasonal vegetation-stress benchmarks. It is less conclusive as evidence of a general-purpose EO “world model,” because the strongest tests are still curated around European summer vegetation degradation rather than long-horizon, multi-region, multi-crop, or true counterfactual interventions.
Reproducibility
Partially reproducible: the paper says the model and benchmarks will be open-sourced at the listed GitHub URL, and the benchmarks are derived from EarthNet2021. The provided excerpt does not confirm that the repository is already populated or that all preprocessing scripts/checkpoints are available.
Key figure
Figure 1 frames EO forecasting as a sparse, partially observed, weather-driven world modeling problem and illustrates the two proposed evaluations: extreme-summer vegetation degradation and matched-pair response under changed meteorological forcing.