A high-resolution fuel type mapping procedure based on satellite imagery and neural networks: Updating fuel maps for wildfire simulators

Abstract

A major limitation in the simulation of forest fires involves the proper characterization of the surface vegetation over the study area, based on land cover maps. Unfortunately, these maps may be outdated, with areas where vegetation is either not documented or inaccurately portrayed. These limitations may impair the predictions of wildfire simulators or the design of risk maps and prevention plans. This study proposes a complete procedure for fuel type classification using satellite imagery and fully-connected neural networks. Specifically, our work is based on pixel-based processing cells, generating high-resolution maps. The field study is located in the Northeast of Castilla y León, a central Spanish region, and the Rothermel criteria was followed for the fuel classification. The results record an accuracy of close to 78% on the test sets for the two studied settings, improving on the results reported in previous studies and ratifying the robustness of our approach. Additionally, the confusion matrix analysis and the per-class statistics computed confirm good reliability for all fuel types in a cross-validation framework. The predicted maps can be used on wildfire simulators through GIS tools.