Assessment of Spectral Transformations and Machine Learning Methods for Rapid Prediction of Soil Properties Through Vis-NIR Soil Spectroscopy

ABSTRACT This study investigates the efficacy of various spectral transformations and machine learning (ML) models in prediction of soil properties, namely Soil Reaction (pHe), Electrical Conductivity (ECe), Calcium Carbonate (CaCO3), and Soil Organic Carbon (SOC), using diffuse reflectance spectra. 540 model pipelines were employed for each property using a combination of filtering techniques, derivatives, polynomial orders, smoothening points, and machine learning algorithms. Large variations in 10-fold cross-validation Root Mean Square Error (RMSE) and coefficient of determination (R2-RSQ) were observed among the 540 model pipelines for each property. The RMSE and RSQ ranged from 0.43 to 0.60 and 0.23 to 0.61, respectively, for pHe. For EC, the values ranged from 5.39 to 10.32 (RMSE) and 0.21 to 0.71 (RSQ), respectively. For SOC, the RMSE ranged from 0.10 to 0.15 and RSQ from 0.16 to 0.60. For CaCO3, the values ranged from 1.79 to 3.07 and 0.58 to 0.85. The study advocates the use of a combination of meticulous spectral data preprocessing and diverse modeling approaches to improve predictive accuracy in soil spectroscopy.