A Memory-Based Learning Approach as Compared to Other Data Mining Algorithms for the Prediction of Soil Texture Using Diffuse Reflectance Spectra

Successful determination of soil texture using reflectance spectroscopy across Visible and Near-Infrared (VNIR, 400-1200 nm) and Short-Wave-Infrared (SWIR, 1200-2500 nm) ranges depends largely on the selection of a suitable data mining algorithm. The objective of this research was to explore whether the new Memory-Based Learning (MBL) method performs better than the other methods, namely: Partial Least Squares Regression (PLSR), Support Vector Machine Regression (SVMR) and Boosted Regression Trees (BRT). For this purpose, we chose soil texture (contents of clay, silt and sand) as testing attributes. A selected set of soil samples, classified as Technosols, were collected from brown coal mining dumpsites in the Czech Republic (a total of 264 samples). Spectral readings were taken in the laboratory with a fiber optic ASD FieldSpec III Pro FR spectroradiometer. Leave-one-out cross-validation was used to optimize and validate the models. Comparisons were made in terms of the coefficient of determination (R-cv(2)) and the Root Mean Square Error of Prediction of Cross-Validation (RMSEPcv). Predictions of the three soil properties by MBL outperformed the accuracy of the remaining algorithms. We found that the MBL performs better than the other three methods by about 10% (largest R-cv(2) and smallest RMSEPcv), followed by the SVMR. It should be pointed out that the other methods (PLSR and BRT) still provided reliable results. The study concluded that in this examined dataset, reflectance spectroscopy combined with the MBL algorithm is rapid and accurate, offers major efficiency and cost-saving possibilities in other datasets and can lead to better targeting of management interventions.