Ensemble Based Learning for Automated Safety Labeling of Prescribed Fires

Abstract

Prescribed fires are controlled burns of vegetation that follow a burn plan to reduce fuel build-up and mitigate unanticipated wildfire impacts. To understand the risks associated to a prescribed burn, modern fire simulation tools can be used to simulate the progression of a prescribed fire as a function of burn conditions that include ignition patterns, wind conditions, fuel moisture and terrain information. Although fire simulation tools help characterize fire behavior, the unknown non-linear interactions between burn conditions requires the need to run multiple fire simulations (ensembles) to formulate an allowable range on burn conditions for a burn plan. Processing the ensembles is often a labor intensive process run by user-domain experts that interpret the simulation results and carefully label the safety of the prescribed fire. The contribution of this paper is an algorithm of ensemble based learning that automates the safety labeling of ensembles created by a modern fire simulation tool. The automated safety labeling in this algorithm is done by first extracting important prescribed fire performance metrics from the ensembles and learn the allowable range of these metrics from a subset of manually labeled ensembles via a gradient free optimization. Subsequently, remaining ensembles can be labeled automatically based on the learned threshold values. The process of learning and automatic safety labeling is illustrated on 900 ensembles created by QUIC-Fire of a prescribed fire in the Yosemite, CA region. The results show a performance of over 80% matching of learned automated safety labels in comparison to manually generated safety labels created by fire domain experts.

Work is supported by WIFIRE Commons and funded by NSF 2040676 and NSF 2134904 under the Convergence Accelerator program.