Spatial Modeling for Soil Properties Prediction in Mountainous Areas Using Partial Least Squares Regression

Abstract

Soil properties are one of the most important categories of information for land management and environmental modeling. Unfortunately, soil properties in mountainous areas with slopes of more than 35% are rarely investigated in Thailand due to the complexity of their landscapes and the cost and time requirements. The main objective was to predict soil properties in mountainous areas relating to soil forming factors using partial least squares regression (PLSR). The combination of topographic position index values from two different scales and criteria sets was firstly used to classify landform for in situ soil survey. Then, analyzed soil properties of the topsoil and subsoil (sand, silt, clay, pH, organic matter, total N, available P, exchangeable K, cation exchange capacity (CEC) and base saturation) and soil forming factors (rainfall, normalized difference vegetation index, elevation, slope, aspect, plan curvature, profile curvature, curvature, topographic wetness index and Al/Si ratio) were used to construct soil-landscape models using PLSR. It was found that the best predictive model for topsoil prediction was sand (R2 = 0.92) and the worst was silt (R2 = 0.52) while the best predictive model for subsoil property prediction was CEC (R2 = 0.85) and the worst was total N and available P (R2 = 0.59). Accuracy assessment for the topsoil and subsoil properties prediction models using normalized root mean square error varied between 0.18 to 0.25 and 0.18 to 0.36, respectively. In addition, the selected predictive soil properties were used for soil texture classification and soil fertility assessment. In conclusion, it is suggested that soil-landscape modeling using PLSR can be efficiently used as a tool for spatial soil property prediction in mountainous areas where soil characteristics and properties are not available.