Document Type : Research Paper
Authors
1 Department of Soil Science, Faculty of Agriculture, Lorestan University, Lorestan, Khorramabad, Iran
2 Department of Soil Science and Engineering, Faculty of Agricalture, University of Tehran, Karaj, Iran
Abstract
The scale of environmental variables is one of the most important features to consider when selecting data. The aim of this study is to improve the accuracy of digital mapping by selecting the optimal scale for predicting six soil properties, For this purpose, 100 surface soil samples (0-30 cm depth) were collected based on a random sampling pattern. Environmental variables related to topography and remote sensing were extracted from the digital elevation model (DEM) and Landsat-8 satellite. The optimal environmental variables were selected using the recursive feature elimination method in the Silakhor Plain region. Soil property modeling was conducted using machine learning models such as random forest (RF), Support Vector Regression (SVR), Cubist (CB), and hybrid modeling. The modeling results showed that the RF model performed best for predicting CCE, pH, sand, and silt with R² values of 0.64, 0.65, 0.59, and 0.70, respectively. Additionally, the SVR model showed the highest accuracy for predicting SOC with an R² of 0.62, while the CB model had the highest accuracy for predicting clay with an R² of 0.66. The most suitable cell sizes for CCE, pH, sand, and silt were identified as 30*30 m, for SOC as 60*60m, and for clay as 90*90m. The most important environmental variables for SOC, pH, silt, sand, and clay were valley depth, differential vegetation index, and modified vegetation index, respectively. Overall, the results indicated that in the study areas, the use of intermediate scales (cell sizes of 30 to 90 m) led to higher accuracy in predicting soil properties. This is because using larger cell sizes introduces noise that hinders accuracy.
Keywords
)