Implementation of a fuel type classification system for Sardinia, Italy, with the integration of remotely sensed data

Forest fires are a keystone ecosystem process in the evolution and maintenance of the Mediterranean biome. However, the expected increase in fire activity under future climate projections poses a growing ecological and socio-economic threat to the Euro-Mediterranean region. The dynamic mapping of fuel types and models assumes relevance to wildland fire risk prevention and management across multiple spatial and temporal scales due to the tight dependence of fire ignition, spread, and growth on vegetation characteristics. Thematic maps of fuel types and models already exist at a global and continental scale, but the spatiotemporal variability of fuel characteristics in the Euro-Mediterranean region highlight the need for further improvements in the identification of techniques and methodologies based on the integration of field observations with remotely sensed data that allow for a periodic update of high resolution thematic maps. This research proposes a methodology for generating a fuel type map for a pilot site located in Sardinia, Italy, compliant with the first level scheme of the hierarchical classification system recently proposed within the context of the EU project FirEUrisk; the classification system has been chosen for its adaptability to different scales of investigation and to different geographical contexts in the Euro-Mediterranean regions. The adopted methodology applies supervised learning algorithms for the fuel type classification from time series of multispectral images of the Sentinel-2 (S2) missions of the Copernicus program. The classification methodology is applied to S2 time series of spectral indices for the period 2020-2021 after preliminary cloud cover masking and compositing. Remotely sensed information is integrated with auxiliary information derived from institutional datasets available on a continental scale, such as the CORINE Land Cover system, and on a local scale, such as the digital elevation model and the mosaic of high resolution orthophotos for the year 2018. Due to the lack of field data uniformly distributed in the pilot site, an ad hoc dataset is generated by photointerpretation for the supervised classification model training and testing. To this aim, a web application is designed and developed to support a consistent data collection within 1 ha regions of interest (ROIs). The application allows the labelling of each 10x10 meter S2 pixel within each ROI selected with a stratified random sampling strategy. A gradient boosting ensemble model is trained for a pixel-level classification that integrates spectral metrics, textural metrics, and other geo-environmental descriptive metrics. Finally, the classifier is applied to derive a 10 m resolution thematic fuel type map for the pilot site in Sardinia. Preliminary results show an overall accuracy greater than 0.9 and calculated as the number of pixels correctly classified by the model out of the total number of pixels included in the different cross-validation datasets. The classification specificity is greater than 0.7 for most of the fuel type classes. Future activities will be focused on a robust validation of the thematic fuel type map by selecting suitable sites based on the possibility of carrying out ad hoc field surveys. If the preliminary results are confirmed, the methodology could be extended to other pilot sites in the Euro-Mediterranean basin and the obtained thematic fuel type map could be processed to obtain a fuel model map to be used for fire management purposes, such as the simulation of fire spread and growth using simulators based on the Rothermel mathematical model.