Predicting and preparing a risk map of rangeland fires using random forest algorithms and support vector machine (Case study: Arak rangelands)

Abstract Background and objectives: Rangeland fires have devastating effects on the landscape, performance and services of rangeland ecosystems. Despite the efforts of experts, decision makers, stakeholders and government agencies in recent decades to reduce the effects of fire, its number and related economic and human losses are increasing worldwide. One of the most important measures to reduce the damage caused by fire is to predict and prevent the occurrence of fire, which is based on determining the danger zone or areas prone and potential for fire. The purpose of this study is to identify and determine areas sensitive to fire in the winter rangelands of Mohammad Gholi in Arak city of Markazi province. Methodology: The studied rangelands with an area of 3100 hectares with arid to semi-arid climate are located 15 km southeast of Arak city in Markazi province. The altitude of the region is 1900 to 2500 meters (a.s.l) and the average annual rainfall is 225 mm. The thermal extremes of the region are -11 (February) to 35 degrees Celsius (August). In order to zoning the fire risk in the region, 9 factors of slope, direction of slope, altitude, geology, land use, distance from the road, distance from the waterway network, soil science and vegetation percentage were used. Fire events were considered as a basis for predicting future fires. Non-fire areas were also identified and selected. For backup zoning and fire prediction, two support machine models (SVM) and random forest model (RF) were used.  In order to evaluate the results of these models, the statistical indices of coefficient of explanation (R2: correlation between observational and estimated data), the square root of the mean squared error (RMSE: deviation of predicted values ​​from observed values) and efficiency coefficient (CE: coefficient of efficiency: between infinite negative and 1, the closer to one indicates the higher the performance of the model in forecasting) was used. The output of RF and SVM models is between 1 and zero, which is divided into 5 (floors) of the area with very low to very high fire hazards. The models will identify the most important variables affecting the past fire and then zoning the fire risk in the region. Results: Vegetation, direction, slope and altitude variables had the greatest impact on fire, respectively, and the variables of geology, land use, distance from the road, distance from the waterway network and soil science were removed from the modeling process due to inappropriate and insignificant coefficients. Were. The most sensitive slope floor in case of fire is 12-25% and above and floor 8-12 has the lowest fire incidence. Also, the highest fire occurred at an altitude of 2100-1900 meters and the lowest at an altitude of 2500-2400. In terms of direction, the southwestern and southern slopes had the most and the northern slopes and the non-sloping direction had the least fire events. Vegetation, by providing the necessary fuel, has shown the highest incidence of fire in the coverage of 50-75% and the lowest in the coverage below 25%. According to the results of the implementation of the models, the support vector machine model with a coefficient of efficiency of 0.86 and an error of 3.55 in the test phase is a more accurate model in this study. The results also showed that in terms of fire risk, 11% of the rangelands were in the very low category, 16% in the low category, 35% in the medium risk category, 17% in the high risk category and 21% in the very high risk category are located. Conclusion: High slopes and heights with maximum vegetation (suitable fuel source) in the area and lower grazing intensity have the highest incidence of fire. While in low cover due to insufficient fuel and in low slopes due to change of Rangelands to agriculture, fire is less likely to occur. The area to the south also provides suitable fuel for the fire by receiving more solar heat, dominant cover of Astragalus and dense cover of annual grasses. Among the selected models, the vector support machine model had better performance in zoning and fire risk prediction than the random forest model, which is due to its ability to integrate many input variables without changing them, and Establishing nonlinear relationships between variables identifies effective factors and can provide valuable information for fire control and prevention to rangeland managers.