FireMPC: A Multi-Source Pan-Canadian Wildfire Benchmark Revealing Spatiotemporal Generalization Gaps

Under Climate Change, wildfire activity is intensifying worldwide, with Canada experiencing increasingly severe events that now affect every province and territory. However, existing machine-learning (ML) datasets remain constrained by geographic bias, narrow driver sets, opaque sample selection, and limited exploration of model failure modes. We introduce \textbf{FireMPC} dataset, the first pan-Canadian multisource wildfire risk benchmark, covering approximately one billion hectares, or $9{,}840 \times 4{,}639$ km$^2$, across the entire Canada at $1$ km daily resolution from $2000$ to $2025$, and integrating $55$ drivers across fuel, topography, meteorology, and human activity. FireMPC is also the first large-scale machine-learning benchmark to embed the six knowledge-driven indicators of the Canadian Forest Fire Danger Rating System (CFFDRS), bridging knowledge-based and data-driven modeling. Building on this dataset, we propose \textbf{FWI-guided hard negative mining (FWI-HNM)} strategy. We score each non-fire candidate using a calibrated CFFDRS composite, and then incorporate these hard negatives during model training. FWI-HNM mitigates the trivial-negative failure mode of random sampling, yielding F1/PR-AUC gains of $1\%$-$2\%$ and ECE reductions of up to $82\%$ across thirteen benchmark ML models. Stress tests under horizon-extension and spatial-generalization protocols expose substantial generalization gaps. Mean F1 falls from $70.2\%$ at the next-day horizon ($\Delta t = 1$) to $54.2\%$ at the 30-day horizon ($\Delta t = 30$). All six ecological, land-cover, and regional spatial-shift scenarios yield a $25\%$-$29\%$ relative F1 drop. These results motivate the development of large-scale, driver-diverse, spatially comprehensive datasets. The dataset, code, and sampling artifacts are available at \url{https://anonymous.4open.science/r/FireMPC}.