Comparing Forest Understory Fuel Classification in Portugal Using Discrete Airborne Laser Scanning Data and Satellite Multi-Source Remote Sensing Data

Wildfires burn millions of hectares of forest worldwide every year, and this trend is expected to continue growing under current and future climate scenarios. As a result, accurate knowledge of fuel conditions and fuel type mapping are important for assessing fire hazards and predicting fire behavior. In this study, 499 plots in six different areas in Portugal were surveyed by ALS and multisource RS, and the data thus obtained were used to evaluate a nationwide fuel classification. Random Forest (RF) and CART models were used to evaluate fuel models based on ALS (5 and 10 pulse/m2), Sentinel Imagery (Multispectral Sentinel 2 (S2) and SAR (Synthetic Aperture RaDaR) data (C-band (Sentinel 1 (S1)) and Phased Array L-band data (PALSAR-2/ALOS-2 Satellite) metrics. The specific goals of the study were as follows: (1) to develop simple CART and RF models to classify the four main fuel types in Portugal in terms of horizontal and vertical structure based on field-acquired ALS data; (2) to analyze the effect of canopy cover on fuel type classification; (3) to investigate the use of different ALS pulse densities to classify the fuel types; (4) to map a more complex classification of fuel using a multi-sensor approach and the RF method. The results indicate that use of ALS metrics (only) was a powerful way of accurately classifying the main four fuel types, with OA = 0.68. In terms of canopy cover, the best results were estimated in sparse forest, with an OA = 0.84. The effect of ALS pulse density on fuel classification indicates that 10 points m−2 data yielded better results than 5 points m−2 data, with OA = 0.78 and 0.71, respectively. Finally, the multi-sensor approach with RF successfully classified 13 fuel models in Portugal, with moderate OA = 0.44. Fuel mapping studies could be improved by generating more homogenous fuel models (in terms of structure and composition), increasing the number of sample plots and also by increasing the representativeness of each fuel model.