Performance and Efficiency of Machine Learning Based Approaches for Wildfire Susceptibility Mapping ^†

Estimating the probability of wildfire occurrence in certain areas, under particular environmental and anthropogenic conditions, is a powerful tool to support forest protection and management plans. In this context, the implementation of Wildfire Susceptibility Maps (WSM) and the investigation of the main driving factors (e.g., land cover class, type of vegetation, topography) are fundamental.

Susceptibility maps indicate areas with the potential to experience a particular hazard in the future based on the intrinsic local properties of a site, as well as the observed past events. Machine Learning (ML) based approaches lend themselves well to this purpose. ML is essentially based on algorithms capable of learning from and making predictions on data by modelling the hidden/non-linear relationships between a set of input variables (driving factors) and output observations (Figure 1).

In the present work, we discuss three case studies for WSM dealing with (i) areas mapped at different scales, and (ii) characterized by a different degree of accuracy/quality of the input datasets. They consist of: (1) comparison between deterministic methods (assuming a priori knowledge of driving factors) versus stochastic approaches (based on artificial neural network and decision trees) for WSM in Dão-Lafões region (Portugal) [1]; (2) implementation of an ensemble ML algorithm based on decision trees (Random Forest) for WSM in Liguria Region (Italy) [2]; and (3) in Santa Cruz (Bolivia) [3].

The first case study emphasizes the advantage of using ML compared to deterministic/linear methods. The second introduces a well-structured and easily replicable application of Random Forest for WSM, including model validation (avoiding spatial autocorrelation and overfitting) and the use of categorical variables. Finally, the third case study successfully demonstrates that it is possible to implement a simple, but powerful, model even for a country such as Bolivia, with poor resources in terms of data availability and informatisation.