Comparisons of the prediction results of soil properties based on fuzzy c-means clustering and expert knowledge from laboratory Visible – Near-Infrared reflectance spectroscopy data

Abstract

Soil organic matter and soil particle composition play extremely important roles in soil fertility, environmental protection, and sustainable agricultural development. Visible – near-infrared reflectance (Vis–NIR) spectroscopy is a rapid, effective, and low-cost analytical method to predict soil properties. In this study, laboratory Vis–NIR spectroscopy data were used to compare the differences among partial least squares regression (PLSR), artificial neural network (ANN) and multivariate adaptive regression splines (MARSplines) based on fuzzy c-means spectral clustering and expert knowledge classification methods for soil prediction. The results showed that (1) the sand content (R2 = 0.69–0.77) had the best prediction, followed by the silt (R2 = 0.56–0.71) and organic matter (R2 = 0.54–0.69) contents, whereas the clay content (R2 = 0.29–0.65) had the poorest prediction, (2) the performance of the models followed the order of PLSR > ANN > MARSplines, and (3) the accuracies of the organic matter and sand contents were higher when applying expert knowledge classification, whereas the prediction of the clay and silt contents was better when applying spectral clustering. However, the overall accuracy of the spectral clustering method was slightly better than that of expert classification. Our findings showed that the spectral cluster-based models produced effective and interpretable prediction results for estimating soil properties. Therefore, this approach should be considered when dealing with large and heterogeneous soil samples.