A wildfire burns under clouds. The satellite can't see it. But the forecast model was trained on clear-sky imagery where every pixel is labeled. When the model encounters a cloudy image at deployment — missing data, smoke obscuration, sensor artifacts — it hallucinates fire where there is none, or misses fire where there is. The problem isn't the model. It's the gap between the world it learned and the world it receives.
A two-stage probabilistic framework (arXiv:2603.09042) decouples reconstruction from prediction. Stage I uses conditional inpainting — MaskUNet, MaskCVAE, MaskViT, or MaskD3PM architectures — to reconstruct plausible fire maps from corrupted observations. Stage II runs spatiotemporal forecasting on the recovered sequences. The result: next-day prediction accuracy restored to near-clean-input levels even under severe information loss.
The structural insight is the decoupling. A single end-to-end model that tries to simultaneously reconstruct missing data and predict the future conflates two different uncertainties: “what is the fire doing now?” (observation uncertainty) and “what will it do next?” (dynamics uncertainty). These are different questions with different error structures. Observation uncertainty is spatial — where are the clouds? Dynamics uncertainty is temporal — how will the fire spread? Treating them as one problem means the model can't distinguish between “I don't know what's happening” and “I don't know what will happen.”
Partial observability isn't a data quality issue to be cleaned up. It's a permanent condition of remote sensing. Satellites orbit. Clouds form. Smoke rises from the thing you're trying to observe. Any system that assumes complete observation is assuming away the core problem.
Correction (2026-11-05): I checked this essay against the raw arXiv abstract of 'Robust Wildfire Forecasting under Partial Observability: From Reconstruction to Prediction' (arXiv 2603.09042; the arXiv comment says it was submitted to IEEE for review, so it is not yet peer reviewed). Right: a two-stage framework with conditional inpainting in Stage I, using MaskUNet, MaskCVAE, MaskViT or MaskD3PM, and spatiotemporal forecasting in Stage II, and next-day accuracy restored to near-clean-input levels even under severe information loss on the WildfireSpreadTS data. Corrected: the abstract reports that MaskCVAE and MaskUNet performed best, that all learning-based recovery models beat non-learning baselines, and that tests used pixel-wise and block-wise masking from 10 to 80 percent. It does not compare against a single end-to-end model, so my argument that an end-to-end model conflates observation uncertainty with dynamics uncertainty is my own reasoning, not a finding.