عنوان مقاله [English]
Background and Objectives: Soil erodibility is one of the most important soil properties which investigation of its spatial variability is essential to crop management, land degradation and environmental studies. Therefore, information about spatial variability of soil erodibility has important role to modeling of water erosion. Investigation of variability of soil erodibility using traditional methods is expensive and time consuming. Therefore, one of the ways to solve this challenge is using digital soil mapping that digitally can predict soil characteristics using auxiliary data and data mining models. The aim of this research is using artificial neural network (ANN) and random forest (RF) models and auxiliary data to make soil erodibility map.
Materials and Methods: Using stratified random soil sampling method 100 soil samples in depths 0-30 cm of Dehgolan soils, Kurdistan Province (covers 48710 ha) were taken and soil texture, fin sand, infiltration, soil structure and soil erodibility (using Wischmeier and Smith equation) were measured and computed. Auxiliary data in this study were terrain attributes and Landsat ETM+ data. Terrain parameters (include 15 parameters) and clay index (SI) and normalized difference vegetative index (NDVI) were computed and extracted using SAGA and ArcGIS10.3 software, respectively. To make a relationship between soil erodibility and auxiliary data, RF and ANN models were applied and were validated using cross validation method. Finally, soil salinity map were made using better model.
Results: To prediction of soil erodibility, auxiliary variables include wetness index, Multi-resolution Valley Bottom Flatness (MrVBF), slope, clay index, NDVI index and B7 were the most important. The results of the study showed that two models (0.80, 0.003 and 0.021 for ANN and 0.76, 0.005 and 0.024 RF for determination of coefficient, mean error, and root mean square root, respectively) were closely matched to predict soil erodibility. Soil erodibility content ranged between 0 to 0.05 t h MJ-1mm-1 and the highest its contents were observed in southern high regions with high slope and low vegetation. In slope class > 10 % soil erodibility was higher than other slope classes. Slope class > 10 % also had the lowest contents of auxiliary data including Wetness index, MrVBF, clay index and band 7 and the highest content of NDVI index .
Conclusion: In this research to investigate spatial variability of soil erodibility ANN and RF models was used in Dehgolan region, Kurdistan province. Soil erodibility content was higher in slope class > 10 % compared to other slope classes. NDVI index was the most important auxiliary data to predict soil erodibility of the study area. ANN and RF also based on the results of statistics indices including determination of coefficient, mean error, and root mean square root (0.80, 0.003 and 0.021 for ANN and 0.76, 0.005 and 0.024 for RF) had accurate estimation of soil erodibility. It is suggested using pedometric techniques such as ANN model and auxiliary data of terrain attributes and satellite images to digital mapping of soil properties and updating old maps. It is suggested also direct measurement of soil erodibility and its result will be compared to this study.