Here, a subset of Khuzestan province was selected as one of the areas affected by severe soil erosion to implement the proposed approach. In the first step, InSAR processes were applied to produce multi-temporal coherence images using a nearly 5-year time series (2015-2020) of Sentinel-1 satellite data with a temporal resolution of 24 days. The Small BAseline Subset (SBAS) algorithm was used to select SAR image pairs with short spatial and temporal baselines to generate coherence images. The second step was to obtain geospatial data on parameters that affect erosion, such as soil texture, elevation, slope, and vegetation. Landsat 8 multispectral imagery was used as the normalized difference vegetation index (NDVI) to incorporate the patterns of vegetation change, a highly effective factor in soil erosion, into the modeling. After preparing the input parameters of the model, in the third step, training and validation samples were prepared using the valid soil erosion map of the study area. In the fourth step, the support vector machine (SVM) classifier was implemented using input parameters and training samples based on different kernels. Finally, a soil erosion susceptibility map was generated and validated for each kernel.

By selecting image pairs with short spatial and temporal baselines, the effects of temporal decorrelation due to the atmosphere and spatial decorrelation due to satellite orbital displacement were minimized. Therefore, the patterns of change in the coherence images are most likely caused by changes in the physical properties of the Earth's surface, which can help model soil erosion. In addition to latitude, climatic parameters are largely a function of local topographic features. Highlands receive more rainfall than lowlands. In the highlands, on the other hand, the variety of slopes is greater. The slope factor can significantly accelerate and intensify the detachment of soil particles by water and wind. According to the valid soil erosion map of the region, a large part of the training samples corresponds to the medium to high soil erosion susceptibility classes, in accordance with the topographic pattern of the region. Because the areas covered by water and vegetation are highly dynamic, they show high variance and change in the coherence images. However, large changes do not necessarily mean severe soil erosion, and this issue has been addressed by incorporating NDVI time series into the modeling. To implement different SVM classifier kernels including linear, polynomial, radial basis function (RBF) and sigmoid, the number of 112 features including 95 coherence images, 12 average monthly NDVI images, 3 images of soil texture characteristics, and 2 images related to topographic characteristics were used. Soil erosion susceptibility maps generated by different kernels show different spatial patterns of erosion. These inconsistencies indicate the different effectiveness of the kernels according to their logic for modeling the complexity in the input data. Accordingly, RBF and Sigmoid kernels provided the best and worst performance in predicting different classes of soil erosion susceptibility with overall accuracies of 80.44% and 52%, respectively.

This research evaluated the effectiveness of InSAR coherence data combined with optical data for mapping areas susceptible to soil erosion based on machine learning capabilities. To improve the accuracy of the modeling, NDVI was also considered. In fact, by taking advantage of NDVI temporal patterns, signals of strong coherence variability caused by the phenological behavior of plants and water body changes have been modeled and detected from severe soil erosion signals. The results obtained from the validation of different kernels of SVM classifiers for mapping areas susceptible to soil erosion show the significant impact of the type of classification model on the accuracy of the outputs. Meanwhile, the RBF kernel achieved the highest performance compared to other kernels with an overall accuracy of 80.44%. Our results highlight the acceptable effectiveness of InSAR coherence data based on Sentinel-1 satellite imagery for modeling soil erosion.