چکیده:
Soil texture is variable through space and controls most of the soil’s Physico-chemical, biological and hydrological characteristics and governs agricultural production and yield. Therefore, determining its variability and generating accurate soil texture maps have a key role in soil management and sustainable agriculture. The purpose of this study is to introduce a numerical algorithm named Least Square Support Vector Machine for Regression (LS-SVR) as a predictive model in Digital Soil Mapping (DSM) of soil texture fractions and evaluating its performances based on modeling evaluation criteria. In this study, the soil texture data of 49 soil profiles in Tabriz plain, Iran, was used. The important covariates were selected using Genetic Algorithm (GA). The model evaluation results based on ME, MAE, RMSE, and R2 indicate the high performance of LS-SVR in predicting soil texture components. The prediction RMSE for sand, silt, and clay was 6.82, 5.08 and 6.06, respectively. Silt prediction had the highest ME and the lowest MAE and RSME values. The algorithm simulated the complex spatial patterns of soil texture fractions and provided high accuracy predictions and maps. Therefore, the LS-SVR algorithm has the capability to be used as predictive models in soil texture digital mapping. This study highlighted the potential of the LS-SVR algorithm in high precision soil mapping. The generated maps can be used as basic information for environmental management and modeling.
خلاصه ماشینی:
The purpose of this study is to introduce a numerical algorithm named Least Square Support Vector Machine for Regression (LS-SVR) as a predictive model in Digital Soil Mapping (DSM) of soil texture fractions and evaluating its performances based on modeling evaluation criteria.
The model evaluation results based on ME, MAE, RMSE, and R2 indicate the high performance of LS-SVR in predicting soil texture components.
Therefore, the LS-SVR algorithm has the capability to be used as predictive models in soil texture digital mapping.
DSM employs field observations, laboratory measurements, digital elevation model (DEM), and satellite imagery derivates as inputs for building mathematical/statistical (quantitative) models to map spatial patterns of soil properties through the area (Minasny and McBratney, 2016).
Methodology In the present study, the RBF based LS-SVR algorithm employed as a predictive model for estimating soil texture fractions (sand, silt, and clay) using 49 samples, DEM, and Lansat5 image derivates.
During the feature selection phase; Band 5, Band7, wetness, and X coordinate were selected as important covariates for predicting all soil texture fraction (sand, silt, and clay).
Based on the models’ R2 values, the algorithm was modelled the soil texture components with R2 of 0086, 0083, and 0084 for sand, silt, and clay, respectively.
The results indicate that the LS-SVR has high performance and could model the complex spatial patterns of soil texture fractions by providing high accuracy predictions and maps.
Accordingly, the LS-SVR algorithm has the potential to be used as predictive models in DSM for soil texture fractions mapping.