Forest fires susceptibility mapping using machine learning algorithms (Case study: Marivan’s County forests)
2024
Identifying the effective factors in the occurrence of fire and zoning the sensitivity to its occurrence is one of the basic tools to achieve fire control and countermeasures. Forest fire modeling is very important to identify the distribution of forest fires based on scientific methods. In this research, the Information Gain Ratio (IGR) technique and the Average Merit index were used to evaluate the predictive power of factors affecting the occurrence of forest fires in Marivan County. The results of these methods showed that among the 14 effective factors considered at the beginning, 12 factors were involved in the occurrence of fire, which include: annual average wind speed, altitude, relative humidity, precipitation, average maximum temperature , distance from the road, land use, road density, distance from residential areas, NDVI, solar radiation and slope. Also, the results showed that the two factors of slope direction and topographic humidity index were removed from the final modeling due to the mean value of merit equal to zero. Meanwhile, the variables of average wind speed, altitude and relative humidity compared to other variables had the greatest impact on the occurrence of fire. Also, after training all three applied machine learning models, including random forest models, support vector machine and logistic regression, their performance in the field of finding the potential of fire occurrence was measured using statistical criteria. Therefore, in terms of training samples, the random forest model (0.98) had higher accuracy than the support vector machine model (0.931). The value of sensitivity index in random forest and support vector machine models was 0.982 and 0.934, respectively. This means that the random forest model is able to correctly classify 1.98% of fire pixels as fire-dominated areas, which has a higher predictive power than the support vector machine model. The prepared maps were classified into five classes of very low sensitivity, low sensitivity, medium sensitivity, high sensitivity and very high sensitivity based on the Natural Breaks classification method. Also, the area and the percentage of the area of the fire potential floors were extracted for all three models. The results of all three models for the potential of fire showed that the western and southwestern parts of Marivan have a higher fire risk potential than other parts of Marivan. ROC curve method was used to validate all three models. The results showed that among the random forest model, support vector machine and logistic regression, the highest accuracy was assigned to the support vector machine model (0.997).