Artificial intelligence applications for sustainable wildfire management

Abstract

Wildfire management in Chile requires an integrated framework spanning preventive planning, faithful landscape representation, and impact anticipation. This three-study dissertation addresses that methodological sequence: (i) comparing reinforcement-learning (RL) algorithms for efficient firebreak placement in simulated landscapes; (ii) multiband CNN-based fuel-type mapping in two regions; and (iii) predicting—later explaining—individual wildfire size with tabular and image-derived variables (ongoing work).

Chapter~\ref{chapter1:RL} benchmarks A2C, PPO, and Rainbow under increasing complexity. A2C is the only method that stabilizes, while PPO and Rainbow often converge to local optima and fail to surpass a Greedy-DPV baseline. Absent a clearly dominant solution, agents need many more episodes to find and maintain strong policies. Seed variability mirrors stabilization difficulty (higher for Rainbow; lower for PPO, yet stuck below baseline). A key limitation is the slow fire simulator ($\approx$8-20 s/episode), misaligned with RL’s need for high-throughput training (ideally $\le$1 s/episode), which precluded wider exploration of hyperparameters, reward designs, and approximator networks.

Chapter~\ref{chapter2:class} delivers regional fuel-type classifiers for Valparaíso (10×10×17, 20 classes, 5,268 samples) and Biobío (8×8×23, 30 classes, 15,169 samples); cross-region transfer was not pursued due to marked data differences. The best operational model achieved a Cohen's Kappa value of 0.5588 $\pm$ 0.0127 (Valparaíso) and 0.4935 $\pm$ 0.0221 (Biobío) for the test set. Beyond supporting prevention and response, these fuel maps improve fire simulations by territory-specific parameterization, thus strengthening the validity of model-based planning.

Chapter~\ref{chapter:3} models wildfire size with 838 cases and a skewed, heavy-tailed response (10-6,600 ha). Current models underperform, so explainability is deferred. We hypothesize a data-limited regime, limited predictor relevance, and phenomenon complexity, likely requiring dynamic meteorology, fine-scale fuel continuity, and operational signals.

Overall, the dissertation advances an integrated framework: plan (RL for firebreaks), represent the landscape (multiband classification), and anticipate impacts (size).

Future work focuses on accelerating and parallelizing the simulator—targeting $\ge$10× throughput and sub-second episodes—to enable substantive RL exploration of hyperparameters, reward designs, and network architectures; on coupling the regional fuel maps from Chapter 2 with the simulator and RL to increase realism and external validity of firebreak policies; on enriching size-prediction models with dynamic covariates (wind, evolving meteorology, fine-scale fuel continuity) to raise explanatory and predictive power; and on pursuing external validation with agency co-design to support operational usefulness and adoption.