High Resolution Fuel Type Mapping through Satellite Imagery and Neural Networks for Wildfire Simulations: A Case Study in Spain

Abstract

An important limitation in the simulation of forest fires is the correct characterisation of the surface vegetation documented in land cover maps. Unfortunately, these maps are not always available, or there is a lack of accuracy due to the dilated updating periods. These limitations can result in less-accurate predictions when wildfire models are applied to real-world situations employing information from these maps. New remote sensing technologies can provide up-to-date information on the state of the forest surfaces. On the other hand, in the last decade, we have also seen how artificial intelligence algorithms can efficiently process information to solve many different types of problems. Therefore, in this work we propose a complete procedure for fuel type mapping using satellite imagery and artificial deep neural networks. Specifically, our work is based on pixel-based processing cells, so the prediction of the fuel type is carried out by classifying isolated pixels, opening the door to generating high-resolution fuel-type maps. To test our technological solution, we studied an area located in Castile and León, a central Spanish region. The spectral information employed were collected from ETM+ sensor onboard Landsat 7 spacecraft and from ASTER sensor onboard Terra spacecraft. In addition, the classifier is also assisted with information about mean surface temperature and orography collected from MODIS device, and with several spectral indexes computed to enhance the spectral characteristics of the imagery. We have carried out classification of the surface vegetation for different fuel types, according to the Rothermel classification criteria adapted to the vegetation of the Iberian Peninsula. Results show an accuracy near 78%, improving some of the results reached in previous studies and demonstrating the robustness of our procedure.