Evaluating the efficiency of different machine learning models in extracting the map of erosion forms of arid watersheds (Case study: Mukhtaran plain watershed, South Khorasan, Iran)

Soil conservation is of paramount importance for sustainable development, food security, and environmental protection. Understanding soil erosion is considered as a critical practice for soil conservation programs worldwide. Currently, soil erosion is increasingly recognized as a serious concern for sustainable agriculture, water resources management, and modern civilization. Soil erosion poses a significant threat to soil, ecology, and humanity since the long-term productive capacity of soil is profoundly impacted by the degradation and washing away of topsoil and soil organic matter. In arid regions like Iran, soil erosion is a major crisis and can be considered as one of the critical challenges for agricultural development, natural resources, and the environment. The high sediment load entering rivers from upstream areas leads to increased water turbidity, reduces the lifespan of dams due to reservoir sedimentation, and negatively affects water quality and biological activity.

The Mukhtaran watershed is located in the South Khorasan province of Iran, covering an area of 2421 square kilometers. It lies on the southern slopes of the Bagheran Mountains and encompasses a diverse range of landforms, including mountainous and hilly terrain, vast plains, and playa devoid of vegetation. Annual precipitation in the Mukhtaran region varies between 150 millimeters in the lowlands and 220 millimeters in the highlands. The average annual temperature is 14.3 degrees Celsius; the average minimum temperature is 5.6 degrees Celsius, and the average maximum temperature is 22.7 degrees Celsius. In this work, at the first stage, various base maps including the drainage network, basin slope, geology, geomorphology, soil, and land use were extracted. Then, the map of work units was identified and classified. Subsequently, by interpreting the available aerial photographs, the extent of rill erosion forms was separated on the primary map. Through field visits, the locations of rill, gully, and streambank erosion were recorded using the GPS. The number of points for each erosion type and severity level were as follows: Rill erosion:  Low: 55 points, Medium: 90 points, Severe: 58 points, very severe: 66 points, Gully erosion: Low: 62 points, Medium: 69 points, Severe: 63 points, Streambank erosion: Low: 95 points, Medium: 107 points, Severe: 12 points. Based on a review of previous studies and considering the nature of water erosion and the available baseline information resources in the region, 25 important and influential variables on water erosion were identified and their maps were prepared using various sources. Employing the available information, 25 environmental variables were considered for model generation, including 5 physiographic variables, 2 climatic variables, 4 hydrologic variables, 8 soil variables, 4 land cover variables, and 2 geological variables. The data was divided into two groups for training and validation with a ratio of 70 to 30. The model was repeated five times to evaluate its stability. The performance of the model was evaluated using the metrics ROC, KAPPA, and TSS.

Based on the validation analysis results, the best model for slight to very severe rill erosion was the ensemble model (ESMs) with the accuracies of 97.0%, 85.0%, 90.0%, and 98.0%, respectively. For slight to severe gully erosion, the ensemble model (ESMs) also performed best simulation with the accuracies of 88.0%, 96.0%, and 96.0%. Finally, for slight and moderate streambank erosion, most models performed well, but the ensemble model (ESMs) had the highest accuracies with 94.0%, 98.0%, and 99.0%, respectively. In all forms of erosion, the ensemble model (ESMs) performed best simulation based on all three coefficients, at an excellent level.

Mukhtaran watershed is severely faced with the problem of land degradation in various forms of erosion such as channel erosion, rill erosion and stream bank erosion, and this is the reason that not only the economy of this area is affected, also the natural environment and ecosystem related to it. In this study, different machine learning approaches with random sample partitioning were applied to estimate the most accurate vulnerable areas with the maximum possible accuracy. An examination of the relative importance of all environmental factors in each of the three major types of water erosion in the study area showed that physiographic factors, geological factors, land cover, and soil had a significant impact on the geographical distribution of water erosion forms in the Mukhtaran watershed. These results are consistent with the studies that have used support vector machine as an effective tool for mapping erosion susceptibility in watersheds. Overall, it can be concluded that machine learning models are effective and novel approaches for land use planning and erosion risk management.