Prediction and Modeling of Soil Erosion Rates Using Machine Learning Algorithms in Maraveh Tappeh and Kalaleh Counties, Golestan Province

Introduction and Goal
Water erosion is one of the major environmental hazards worldwide, leading to soil degradation, reduced fertility, and increased sedimentation in rivers. Golestan Province, particularly its northern regions with sensitive loess soils and specific climatic conditions, is considered one of the main erosion hotspots in Iran. The investigated issue is the limitation of traditional methods in providing accurate, rapid, and reliable zoning of water erosion in the sensitive regions of Golestan Province, where complex and nonlinear relationships among environmental factors require more advanced approaches. Machine learning (ML) algorithms, due to their ability to model complex and non-linear relationships between variables, are powerful tools for the spatial prediction of phenomena such as erosion. The objective of this study was to model the spatial potential of water erosion in the Maraveh Tappeh and Kalaleh counties using advanced machine learning algorithms and to compare their performance with an ensemble model.
Materials and Methods
The study area covers the northern lands of Golestan Province, with an area of approximately 4944.5 km², which is susceptible to water erosion due to its steep topography, seasonal rainfall, and loess soils. Therefore, first, layers of informational of factors affecting erosion were first prepared. These factors included eight environmental and climatic indices: Precipitation Index (station data), Vegetation Index (NDVI), Topographic Grain Size Index (TGSI), Land Use layer, Normalized Difference Moisture Index (NDMI), Digital Elevation Model (DEM), Bare Soil Index (BI), and Normalized Difference Salinity Index (NDSI), all of which were extracted from Landsat 8 satellite imagery. A map of observed erosion locations was used as the dependent variable layer in the modeling. Accordingly, 100 points (50 points of water erosion and 50 points of its absence) were taken during the field visit. In this study, 70% of the data was used for training and 30% for testing. Three machine learning algorithms Random Forest (RF), Support Vector Machine (SVM), and Boosted Regression Trees (BRT) were employed to train the models. Finally, an ensemble model was developed based on the weighted average of the outputs of these three individual models to enhance prediction accuracy. The performance of the models was evaluated using statistical indices, including the Kappa coefficient, Receiver Operating Characteristic (ROC) curve, and True Skill Statistic (TSS).
Results and Discussion
The evaluation results indicated that the performance of each of the three individual models (RF, SVM, and BRT) in modeling water erosion was acceptable, however, the ensemble model (Kappa=0.90, ROC=0.93, and TSS=0.89), by leveraging the strengths of all three models, demonstrated superior performance in accurately distinguish erosion-prone from stable zones and in reducing classification errors. This superiority confirms that the ensemble approach, using weighted averaging, uncertainties can be managed more effectively than in any individual model, resulting in a more stable, reliable output. The final erosion potential map, generated by this model was clearly identified that the northern, western, and southwestern parts of the study area are the main hotspots for water erosion. This spatial pattern is not random but is rooted in the complex interaction of factors; in these areas have steep topography, which increases water flow velocity and its erosive power, receive higher rainfall, providing the necessary energy for the erosion process, and possess weaker vegetation cover, which acts as the soil's primary defense. Analysis of variable importance also revealed that precipitation as a driving factor and the Vegetation Index (NDVI) as the resisting factor, are respectively the most influential controls on erosion in the region. This finding indicates that any conservation measure that improve vegetation cover (such as planting, grazing management, and preventing degradation) can directly neutralize the destructive effects of rainfall. This approach is the most effective strategy for controlling erosion in sensitive areas. Therefore, the results of this study not only provide an accurate map but also offer a scientific framework for prioritizing and implementing evidence-based management actions.
Conclusion and Suggestions
The superiority of the ensemble model in this research demonstrates the high reliability of weighted averaging approaches in reducing error and enhancing prediction accuracy. The concentration of high-risk areas in steep slopes with low vegetation cover is fully consistent with the principles of erosion physics. The high importance of precipitation and vegetation cover emphasizes that sustainable land management, such as implementing biological operations (such as planting resistant and deep-rooted species) is the most effective strategy for erosion control in the region, and according to scientific reports, can reduce surface runoff and erosion by 30–50%. Decision-makers, natural resource managers, and urban and rural planners can use the erosion potential map produced in this study as a practical scientific tool for prioritizing conservation measures and optimize resource allocation. Based on the results of this study, it is recommended that future research, while considering existing limitations (such as the spatial resolution of some datasets), make use of more advanced ensemble models and higher-accuracy data.