Forest fire disaster is currently the subject of intense research worldwide. The development of accurate strategies to prevent potential impacts and minimize the occurrence of disastrous events as much as possible requires modeling and forecasting severe conditions. In this study, we developed five new hybrid machine learning algorithms namely, Frequency Ratio-Multilayer Perceptron (FR-MLP), Frequency Ratio-Logistic Regression (FR-LR), Frequency Ratio-Classification and Regression Tree (FR-CART), Frequency Ratio-Support Vector Machine (FR-SVM), and Frequency Ratio-Random Forest (FR-RF), for mapping forest fire susceptibility in the north of Morocco. To this end, a total of 510 points of historic forest fires as the forest fire inventory map and 10 independent causal factors including elevation, slope, aspect, distance to roads, distance to residential areas, land use, normalized difference vegetation index (NDVI), rainfall, temperature, and wind speed were used. The area under the receiver operating characteristics (ROC) curves (AUC) was computed to assess the effectiveness of the models. The results of conducting proposed models indicated that RF-FR achieved the highest performance (AUC = 0.989), followed by SVM-FR (AUC = 0.959), MLP-FR (AUC = 0.858), CART-FR (AUC = 0.847), LR-FR (AUC = 0.809) in the forecasting of the forest fire. The outcome of this research as a prediction map of forest fire risk areas can provide crucial support for the management of Mediterranean forest ecosystems. Moreover, the results demonstrate that these novel developed hybrid models can increase the accuracy and performance of forest fire susceptibility studies and the approach can be applied to other areas.
