Learning Fire Domain Representations for Interpretable Wildfire Model Transferability

Abstract

     Large wildfires cause severe ecological and economic damage, making accurate burned area prediction essential. Wildfire models are trained on region-specific data, so transferring a model to different fire domains may not preserve accuracy. Existing feature-discrepancy methods use feature similarity between domains as a proxy for transfer quality, but these metrics are symmetric, treating source-to-target and target-to-source transfer as equivalent. Instead, we viewed transferability as a supervised learning target. A domain transfer matrix was built by training a separate model on each of 34 ecoregion clusters and scoring it on every other domain, giving a direct score per domain pair rather than an approximation. This matrix demonstrates asymmetric transfer: performance differs by transfer direction, which feature-discrepancy methods cannot capture. A model was trained on the transfer matrix, using a branched encoder structured by environmental subsystems (atmosphere, ground, fire behavior), to produce static, per-domain embeddings. The encoder is paired with a network that learns source and target bias terms. A deterministic (parameter-free) prediction head combines Euclidean embedding distance with source and target biases to predict transferability. This architecture outperforms feature-discrepancy baselines (Maximum Mean Discrepancy and Wasserstein distance) with a 0.284 Kendall tau improvement across the 34x34 matrix. While XGBoost reached higher performance, our model learns a shared embedding geometry that supports direct ecological inspection. Our model produces ecologically grounded representations: source and target biases correlate with corresponding transfer quality (Kendall tau 0.544 and 0.601, respectively, across 34 domains), and embeddings recover environmental structures like pyromes and climate zones. These results show transferability-supervised representations have potential for identifying viable source models for data-scarce wildfire domains.