Wildfire susceptibility assessment: evaluation of the performance of different machine learning algorithms

Wildfires constitute a complex environmental disaster triggered by several interacting natural and human factors that can affect the biodiversity, species composition and ecosystems, but also human lives, regional economies and environmental health. Therefore, wildfires have become the focus on forestry and ecological research and are receiving considerable attention in forest management. Current advances in automated learning and simulation methods, like machine learning (ML) algorithms, recently aroused great interest in wildfires risk assessment and mapping. This quantitative evaluation is carried out by taking into account two factors: the location and spatial extension of past wildfires events and the geo-environmental and anthropogenic predisposing factors that favored their ignition and spreading. When dealing with risk assessment and predictive mapping for natural phenomena, it is crucial to ascertain the reliability and validity of collected data, as well as the prediction capability of the obtained results. In a previous study (Tonini et al. 2020) authors applied Random Forest (RF) to elaborate wildfire susceptibility mapping for Liguria region (Italy). In the present study, we address to the following outstanding issues, which are still unsolved: (1) the vegetation map included a class labeled “burned area” that masked to true burned vegetation; (2) the implemented model based on RF gave good results, but it needs to be compared with other ML based approaches; (3) to test the predictive capabilities of the model, the last three years of observations were taken, but these are not fully representative of different wildfires regimes, characterizing non-consecutives years. Thus, by improving the analyses, the following results were finally achieved. 1) the class “burned areas” has been reclassified based on expert knowledge, and the type of vegetation correctly assigned. This allowed correctly estimating the relative importance of each vegetation class belonging to this variable. (2) Two additional ML based approach, namely Multi-Layer Perceptron (MLP) and Support Vector Machine (SVM), were tested besides RF and the performance of each model was assessed, as well as the resulting variable ranking and the predicting outputs. This allowed comparing the three ML based approaches and evaluating the pros and cons of each one. (3) The training and testing dataset were selected by extracting the yearly-observations based on a clustering procedure, allowing accounting for the temporal variability of the burning seasons. As result, our models can perform on average better prediction in different situations, by taking into considering years experiencing more or less wildfires than usual. The three ML-based models (RF, SVM and MLP) were finally validated by means of two metrics: i) the Area Under the ROC Curve, selecting the validation dataset by using a 5-folds cross validation procedure; ii) the RMS errors, computed by evaluating the difference between the predicted probability outputs and the presence/absence of an observed event in the testing dataset.

Bibliography:

Tonini, M.; D’Andrea, M.; Biondi, G.; Degli Esposti, S.; Trucchia, A.; Fiorucci, P. A Machine Learning-Based Approach for Wildfire Susceptibility Mapping. The Case Study of the Liguria Region in Italy. Geosciences 2020, 10, 105. https://doi.org/10.3390/geosciences10030105