Document Type : Research Article
Authors
Department of Soil Science, College of Agriculture, Isfahan University of Technology, Isfahan, Iran
Abstract
Introduction
Soil erosion is the most important cause of land degradation in the world, which threatens the sustainable use of the world's soil resources, especially in semi-arid regions. Despite the important implications of soil erosion in sustainable land use, there is limited information about soil erosion in Iran’s watersheds. The lack of this information is related to the complexity of erosion processes, which makes predicting soil erosion costly, time-consuming and difficult. This difficulty has led to the development of various models and tools that seek to simplify soil erosion models and improve our understanding of soil erosion patterns and processes. In recent years, with the development of computer technologies, the evaluation of different machine learning algorithms used in creating predictive models has become the focus of researchers. Machine learning models have been used in numerous studies and have proven as a helpful tool for assessing and mapping various types of water-induced soil erosion. Accordingly, machine learning methods could be used to study the K-factor to identify the areas with higher soil erodibility potential. This research aimed to model and predict the soil erodibility coefficient using various machine learning methods, introduce the most important parameters affecting the prediction of this factor, as well as predict soil erosion using the RUSLE model, and ultimately present a soil erosion risk map in part of the lands of central Iran.
Materials and Methods
The Revised Soil Loss Equation (RUSLE) model, which estimates annual soil erosion rate and evaluates erosion risk by considering five factors rain erosivity (R), soil erodibility (K), slope length and percentage (LS), vegetation cover (C) and conservative operations (P) was employed in this study. Across the study area, 100 points were sampled using the cLHS method to calculate the spatial variations of the K factor. Some of the soil properties including soil organic matter, primary particle size distribution, soil structure stability, and saturated hydraulic conductivity were measured in the laboratory to aid in the K factor calculation. Subsequently, the K factor was estimated using three machine learning methods (SVM, Cubist, and RF) and a map of its spatial variations was predicted. R factor estimated using data from five climatic stations located within and outside the watershed. The LS factor derived from a digital elevation model (DEM) with 12.5 × 12.5 m resolution. The P factor was considered as 1, while the C factor calculated from the NDVI map. Finally, all factors were integrated into the RUSLE model using ArcGIS to estimate the average annual soil loss.
Results and Discussion
The results of the effective factors in soil erosion showed that the western parts of the studied basin have more rainfall than the eastern part of the region. The value of soil erodibility factor in agricultural lands and also in highlands due to the topographical conditions had higher values than the pastures of the study area. The length and percentage of the slope are other effective factors in the calculation of soil erosion. Moreover, the findings confirmed that the amount of soil erosion increased significantly with the increase of the slope. The vegetation factor exhibited high diversity across the region, with low values in areas of proper vegetation like agricultural lands near the river, and high values in degraded pastures. In the study area, no conservation measures, including cultivation on contour lines, strip cultivation, and traces or implementation of banquets, had been carried out. The soil erosion in the central parts of the study area, which includes plains and lands with topographical changes and low slope, had a lower amount than the high lands on both sides of the watershed. In addition to topography, issues such as the loss of vegetation in pastures due to over-grazing and poor management are the main drivers of soil degradation and high erosion in these areas. Various inappropriate management practices, unprincipled cultivation and tillage, the conversion of pastures to low-yielding rainfed agriculture, and excessive grazing are primary drivers of soil degradation in the study area. Factor K modeling results indicated that the random forest model generally exhibited higher performance compared to Cubist and SVM models in almost all machine learning modeling. In all cases, the combination of all environmental variables, including remote sensing indicators, topographic features, and thematic maps (third scenario), resulted in the highest spatial modeling performance.
The results of the implementation of RUSLE model showed that the soil erosion rate for the study area was 8.27 ton/ha/yr. The erosion risk map was prepared using the RUSLE model for the study area and revealed that the critical parts of soil erosion are located at the northeastern part of the basin, which should be given special attention in the executive operations related to watershed management. According to this map, about half of the study area has very low and low erosion, but about 21% of the study area is in severe and very severe soil erosion classes, which require urgent measures to prevent sediment production in the sub-basin.
Conclusion
In predicting soil erodibility coefficient with DSM method, RF model had better results than Cubist and SVM models. Implementing RUSLE model estimated soil erosion rate of the study area to be 8.27 ton/ha/yr. Soil erosion classes were weak, moderate, severe and very severe with about 33, 20, 26, 18 and 3% respectively. About 21% of the basin was classified as severe and very severe, which require urgent soil conservation measures.
Keywords
Subjects
