FROM SURFACE FORECASTING TO OBSERVABILITY FORECASTING: A LATENT WORLD MODEL FOR CLOUD-AWARE EO MONITORING

Earth Observation (EO) pipelines often fail not due to a lack of data, but because of cloud cover and atmospheric haze that render incoming imagery unusable. This paper reframes the standard EarthNet2021 surface-forecasting task into an 'observability forecasting' problem. Instead of attempting to reconstruct pixel-level details, the author adapts the LeWorldModel (LeWM)—a Joint-Embedding Predictive Architecture (JEPA)—to predict whether the next satellite acquisition will be usable and, if not, when a clear view is likely to return.

The model is trained on 23,904 episodes of multispectral imagery and meteorological covariates. It learns to map these inputs into a 192-dimensional latent space, where it predicts future states conditioned on exogenous weather drivers. The evaluation follows a strict 'locked' protocol, where linear probes are fitted only on training data and then frozen for testing across IID, OOD, and extreme weather scenarios.

LeWM demonstrates that latent world models are highly effective for tracking the temporal evolution of scene visibility. Key results include:

• Superior Forecasting: LeWM consistently outperforms persistence-based baselines in both next-step usability (balanced accuracy 0.769–0.887) and predicting the exact horizon for the first usable observation (accuracy 0.602–0.806).
• Task Specialization: When compared to a frozen LightGBM baseline, LeWM proves superior for continuous regression tasks and recovery timing. However, the LightGBM baseline remains more robust for binary classification tasks (e.g., whether any usable observation occurs within six steps) and shows higher stability on OOD data.
• Diagnostic Utility: Beyond forecasting, the model provides strong anomaly-ranking signals, allowing for the detection of temporal inconsistencies in satellite data streams.

This work shifts the focus of EO processing from expensive, often ambiguous pixel-level gap-filling to actionable observability forecasting. By predicting when a scene will be visible, operators can optimize data acquisition schedules and improve the reliability of monitoring pipelines. The study highlights that for operational EO, a latent model that understands temporal state transitions is often more valuable than a model that attempts to synthesize high-resolution spatial imagery.