|
|
:
|
Deforestation, as a major environmental issue, is recognized as the most serious threat to biodiversity and a key driver of land-use change. This study investigates the spatial distribution of deforestation using machine learning techniques and Geographic Information Systems (GIS) within the Loveh Forest Management Plan, covering an area of over 10,000 hectares in Golestan Province. To achieve this, deforested areas within the Loveh forest compartments were identified and their geographic locations recorded. A binary classification approach was employed, where deforested areas were assigned a value of "1" and non-deforested areas were assigned a value of "0." Fourteen explanatory variables—including slope aspect, slope gradient, elevation, landform, plan curvature, slope length, wind effect, minimum temperature, mean temperature, maximum temperature, precipitation, distance to roads, distance to residential area, and distance to agricultural land—were extracted from various data sources and used to predict deforestation probability in the study area. Three ensemble models—Generalized Additive Model (GAM), Classification and Regression Tree (CART), and Random Forest (RF)—were applied to assess deforestation probability. The evaluation results of the area under the curve (AUC) revealed that the random forest model, with an AUC of 0.959 and a Kappa coefficient of 0.88, demonstrated higher accuracy compared to other models. The relative importance analysis of explanatory variables revealed that distance to agricultural land (28.905), wind effect (28.868), slope aspect (22.910), and mean temperature (21.251) had the highest contributions in identifying deforestation-prone areas. Among anthropogenic factors, distance to residential area and roads ranked next in importance after distance to agricultural lands. Overall, deforestation probability was highest in areas close to agricultural lands. The wind effect trend indicated that deforestation probability increased with wind intensity, reaching a minimum around 1.1. The southwestern, southeastern, and eastern aspects showed the highest partial dependence, indicating a greater likelihood of deforestation. The mean temperature analysis revealed an increasing deforestation probability up to approximately 17°C. Deforestation susceptibility maps generated by the three models were classified into four categories using the natural breaks method. According to the Random Forest model results, approximately 15% of the study area falls within the high-risk deforestation zone. The "very low probability" class, with a numerical value of approximately 39%, covered the largest area extent. The findings of this study demonstrate that integrating machine learning techniques with GIS provides a practical and effective approach for identifying deforestation-prone areas. These results suggest that policymakers and forest resource managers can utilize these models to pinpoint critical areas, optimize conservation efforts, and develop proactive deforestation mitigation strategies, particularly in regions experiencing high human pressure. Implementing this approach in forest monitoring and management programs can help reduce deforestation rates and enhance the efficiency of conservation measures.
|