Spectra transfer based learning for predicting and classifying soil texture with short-ranged Vis-NIRS sensor

Abstract

Classifying soil texture is critical to investigate soil processes and functions influencing agronomic decisions and environmental stewardship. Despite the multiple benefits together with accurate estimation potential of visible and near-infrared reflectance spectroscopy (Vis-NIRS) over the traditional method of texture analysis, the prediction accuracy of the Vis-NIRS decreases due to the negative effect of soil moisture content (MC) on soil spectra. This study evaluated the performance of piecewise direct standardization (PDS) and direct standardization (DS) for eliminating the soil MC influence from fresh (non-processed wet) soil spectra to improve the accuracy of soil texture classification using a short-ranged (305–1700 nm) Vis-NIRS (Tec5 Technology, Germany). A set of 96 composite soil samples was collected from 18 fields, and the second set of 294 non-composite samples was collected from 10 of these 18 fields. All soil samples were scanned using a Vis-NIRS sensor after drying, grinding, and sieving, while the non-composite samples were additionally scanned in wet conditions. After analysing the composite samples for percent sand, silt, and clay, partial least squares regression (PLSR) models were calibrated (72 samples in calibration set) and validated (23 samples in validation set). Then, the PLSR model was used to predict the textural fractions for the wet, air-dried, DS and PDS transferred spectra of non-composite soil samples followed by their texture classification. Validation results indicated that the Vis-NIRS sensor predicted sand, silt, and clay fractions with excellent accuracies [coefficient of determination (R2) = 0.88–0.93; residual prediction deviation (RPD) = 2.91–3.74; ratio of the performance to interquartile range (RPIQ) = 2.12–5.02; root means square error (RMSE) = 3.62–8.06 %). Spectra transferred by the PDS and DS improved the texture classification accuracy up to 70 % and 90 %, respectively, while the wet soil spectra misclassified most fields (accuracy = 20 %). Unlike PDS, the DS predicted particles’ distribution seemed approximately similar to laboratory observations. Therefore, it was concluded that transferring the wet Vis-NIRS spectra by the DS algorithm can provide a rapid and accurate prediction of particle fractions and classification of soil texture classes with minimized effort, time, and cost.