Modelling of the factors affecting interrill erosion in pasture and forest landuses using artificial neural networks

Interrill erosion is one of the most important types of erosion, in which various factors such as soil, runoff, and rainfall influence its process and rate. Few studies have been conducted using artificial neural networks (ANNs) to determine the factors affecting interrill erosion in Iran. Furthermore, no research has been carried out in Jiroft on this matter. Therefore, this study was conducted to evaluate the factors influencing interrill erosion using ANNs in four different regions around Jiroft in Kerman province.

For this research, 100 soil samples were collected from two pastures and two forest land uses at depths of 0-10 cm using a random sampling method. Some physical and chemical properties were determined. The amount of interrill erosion was measured using Kamphorst rainfall simulator. Modelling was performed using feedforward multi-layer perceptrons (MLP) with the error backpropagation and Levenberg-Marquardt training algorithm along with 11 soil characteristics in two scenarios. Hill sensitivity analysis was used to investigate the significance of the input variables.

The results revealed that in the study areas, clay, silt, sand (0.05-2 mm), geometric standard deviation (σg), and geometric mean diameter (dg) of particles play a crucial role in interrill erosion while cementing agents such as organic matter (OM) and calcium carbonate equivalent (CCE) were less important. According to the results, the protected forest with high contents of sand as well as low amounts of silt, organic matter and calcium carbonate equivalent showed the lowest erosion rate (2.63 tons /ha). The R2 values for the test datasets in the scenario 1 (with input variables including soil acidity (pH), electrical conductivity (EC), bulk density (BD), organic matter, calcium carbonate equivalent, particulate organic matter (POM), sand, silt, and clay) were 0.81. Whereas the R2 values in the scenario 2 (with input variables such as pH, EC, BD, OM, CCE, POM, the dg and σg) were 0.72. In addition, root-mean-square error (RMSE) for the testing dataset in scenarios 1 and 2 were 0.77 and 1.14, respectively.

Both scenarios had almost the same accuracy in interrill erosion modeling. However, according to the values of R2 and RMSE of the data in scenario 1, this scenario showed better accuracy than scenario 2. In general, the results showed that the ANNs can estimate the amount of interrill erosion using appropriate input variables with high accuracy, and therefore it might be considered as a useful technique to estimate interrill erosion.