Rapid Spatial Estimation of Soil pH using Machine Learning under Limited Covariate Conditions


Conventional soil mapping approaches require to spend long time in the field and laboratory, and are most of the time expensive; therefore, soil scientists continue to study producing reliable digital soil maps in a short time and at a less cost. The main aim of this study was to map the spatial distribution of soil pH at a field scale with fine resolution, and to assess the ability of two commonly used machine learning approaches to estimate soil pH at a scale. The machine learning models applied in this study were Multi-Layer Perception Artificial Neural Network (MLP-ANN) and Support Vector Regression (SVR). The study area covers an approximately 107.1 ha land, and is located in the orchards of fruit research station in Egirdir, Turkey. One hundred and three surface soil samples (0-30 cm) were collected from the corners of 50 x 50 m grid in the study area. The pH value ranged between 7.52 and 8.33 with a mean value of 7.95. The number of hidden node in the MLP-ANN architecture was 16 where the RMSE values in the validation (0.08), test (0.12) and training datasets (0.06) were the lowest. The RMSE, MAE and R2 values of SVR algorithm in the training and test datasets were 0.054, 0.043, 0.759, and 0.075, 0.060, 0.483, respectively. The accuracy of estimated soil pH map produced using MLP-ANN and SVR algorithms were 55.3 and 24.22% higher than the prediction map obtained by conventional ordinary kriging. The finding of the study revealed that machine learning algorithms can be used to produce spatial estimation maps of soil properties which are costly and require intensive time and labor.

MLP-ANN; SVR; Geostatistic; Ordinary Kriging; Soil pH; Machine Learning; Spatial Analysis