Satellite-based wildfire monitoring has a coverage problem. Cloud cover, smoke, and sensor artifacts create systematic gaps in fire observations. A satellite pass might show fire at points A, C, and E along a fireline, with B and D hidden by smoke. The next pass, hours later, shows the fire has advanced — but you don't know what happened in the gaps.
Robust wildfire forecasting under partial observability (arXiv:2603.09042) treats this as a two-stage problem: first reconstruct the missing fire perimeter from incomplete observations, then forecast spread from the reconstruction. The critical insight is that reconstruction and prediction are different tasks requiring different architectures — but most prior work conflated them, trying to predict future spread directly from gapped observations.
The counterintuitive result from the broader literature: more complex transformer architectures (Swin Transformers) don't outperform simpler temporal attention models (UTAE) for next-day spread prediction. The bottleneck isn't model capacity — it's observation completeness. A simpler model working from better-reconstructed inputs outperforms a powerful model working from raw, gapped data.
This inverts the usual deep learning scaling narrative. In most domains, bigger models compensate for noisier data. In wildfire prediction, the spatial correlation structure of fire spread means gaps aren't random — they cluster around the most active parts of the fire (where smoke is densest). The hardest-to-observe regions are precisely the most informative regions. Model capacity can't compensate for systematically missing the signal.
The fix: separate reconstruction from prediction. Invest modeling effort where the information deficit is worst (filling gaps), not where the data is already richest (predicting from complete observations).