Vegetation fuel characterization using machine learning approach over southern Portugal

Understanding the role of fire in the water and carbon cycles is crucial for understanding the Earth's system. Remote sensing is a valuable tool for this purpose as it covers large areas consistently over time. Furthermore, fire propagation models use vegetation parameters to gather information about wildfire conditions, thus reinforcing the need for vegetation dynamics comprehension. Hence, the study aims to improve vegetation representation of fuel load and moisture content, through remote sensing and in-situ data in Southern Portugal. In this study, three above-ground biomass (AGB) datasets and, for live fuel moisture content (LFMC), biweekly samples over two field sites (Herdade da Mitra and Serra de Ossa) were collected during the period between April and October 2022, counting 246 field samples. These samples combined with satellite data information derived from Sentinel-2 (spectral bands and indices) were used in a machine learning approach, the Random Forest (RF) classifier, considering 30 variables to predict the AGB and LFMC. Results showed reasonable agreement between predicted and observed values, with r2 and RSME values of 0.56 (0.69) and 17.56 ton ha-1 (6.47%) for AGB (LFMC). Finally, the RF model generated wall-to-wall dynamic AGB and LFMC maps. This study allowed a product derived from remote sensing data combined with an RF model to produce reliable information about vegetation conditions essential for wildfire risk assessment and atmosphere fire modelling in Southern Portugal.

Automated summary

This study improves the representation of fuel load and moisture content in vegetation through remote sensing and in-situ data, providing reliable information for wildfire risk assessment and atmosphere fire modeling in Southern Portugal.