Smart Prediction System for Territorial Resilience at the Large-Scale Level. Case Study of the Seasonal Forest Fires Risk in Northern Morocco

Predicting forest fire risks constitutes a significant component of territorial risk management and combat strategies. It plays a major role in resource allocation, in mitigation and recovery efforts as well as in anticipating landscape deterioration around urban areas that create an ecological balance at the territorial scale. The purpose of this study is to develop a smart predictive system of seasonal forest fire risk using a machine learning approach. To achieve this aim, data related to 2,130 forest fire events that occurred between 1997 and 2014 were used. Furthermore, biophysical characteristics over the study area were completely processed and retrieved from in-situ measurements; and from time series of MODIS and Landsat (TM, ETM+ and OLI/TIRS) satellite imagery. These data sources served to represent 5 groups of variables, namely Rainfall, Wind, Evapotranspiration, Normalized Difference Vegetation Index (NDVI) andWater Balance; in total, these variable groupings were structured into 39 elemental variables according to the month of the year. The Random Forests algorithm was used to find the best-fit link between theses predictors and the target variable of seasonal forest fire risk. The trained model exhibited a good predictive ability (83% of accuracy, p-value = 0.013). It showed that precipitations, mainly those of the wintry period, have a strong influence on fire occurrence and seasonal severity in the fire season. Accordingly, the developed model allows to predict seasonal risk according to the winter precipitation and to anticipate forest fire risks at a very early stage as well as their impact on improving socio-ecosystem services and territorial resilience.