Investigation of general indicators influencing on forest fire and its susceptibility modeling using different data mining techniques

• Published in: Ecological indicators Vol. 64; pp. 72 - 84
• Main Authors: Pourtaghi, Zohre Sadat, Pourghasemi, Hamid Reza, Aretano, Roberta, Semeraro, Teodoro
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.05.2016
• Subjects: algorithms; Annual rainfall; Boosted regression tree; Data mining; ecosystems; expert opinion; fire suppression; Forest fire susceptibility mapping; Forest fires; forests; Generalized additive model; GIS; Indicators; Iran; land use; Mathematical models; moderate resolution imaging spectroradiometer; normalized difference vegetation index; planning; rain; Random forest; regression analysis; rivers; Roads; soil texture; Surface layer; temperature; Texture; villages; wind; Iran; ISW, Iran; GIS; Random forest; Forest fire susceptibility mapping; Iran; Generalized additive model; Boosted regression tree
• ISSN: 1470-160X; 1872-7034
• DOI: 10.1016/j.ecolind.2015.12.030

•To identify and evaluate the importance of the general indicators involved.•To develop a model capable of predicting forest fire location using BRT, GAM, and RF data mining models in Iran.•To validate the models using receiver operating characteristics (ROC) curve.•To implement the model into a geographical information system to offer a data mining methodology for forest fire susceptibility mapping in Iran and other countries. Forests are living dynamic systems and these unique ecosystems are essential for life on earth. Forest fires are one of the major environmental concerns, economic, and social in the worldwide. The aim of current research is to identify general indicators influencing on forest fire and compare forest fire susceptibility maps based on the boosted regression tree (BRT), generalized additive model (GAM), and random forest (RF) data mining models in the Minudasht Township, Golestan Province, Iran. According to expert opinion and literature review, fifteen condition factors on forest fire have been selected in the study area. These are slope degree, slope aspect, elevation, topographic wetness index (TWI), topographic position index (TPI), plan curvature, wind effect, annual temperature and rainfall, soil texture, distance to roads, rivers, and villages, normalized difference vegetation index (NDVI), and land use. Forest fire locations were identified using MODIS images, historical records, and extensive field checking. 106 (≈70%) locations, out of 151 forest fires identified, were used for models building/training, while the remaining 45 (≈30%) cases were used for the models validation. BRT, GAM, and RF data mining models were used to distinguish between presence and absence of forest fires and its mapping. These algorithms were used to perform feature selection in order to reveal the variables that contribute more to forest fire occurrence. Finally, for validation of models, the area under the curve (AUC) for forest fire susceptibility maps was calculated. The validation of results showed that AUC for three mentioned models varies from 0.7279 to 0.8770 (AUCBRT=80.84%, AUCGAM=87.70%, and AUCRF=72.79%,). Results indicated that the main drivers of forest fire occurrence were annual rainfall, distance to roads, and land use factors. The results can be applied to primary warning, fire suppression resource planning, and allocation work.