Machine learning prediction of particle-size distribution from infrared spectra, methodologies and soil features

Abstract Accuracy of infrared (IR) models to measure soil particle-size distribution (PSD) depends on soil preparation, methodology (sedimentation, laser), settling times and relevant soil features. Compositional soil data may require log ratio ( ilr ) transformation to avoid numerical biases. Machine learning can relate numerous independent variables that may impact on NIR spectra to assess particle-size distribution. Our objective was to reach high IRS prediction accuracy across a large range of PSD methods and soil properties. A total of 1298 soil samples from eastern Canada were IR-scanned. Spectra were processed by Stochastic Gradient Boosting (SGB) to predict sand, silt, clay and carbon. Slope and intercept of the log-log relationships between settling time and suspension density function (SDF) (R 2 = 0.84-0.92) performed similarly to NIR spectra using either ilr -transformed (R 2 = 0.81-0.93) or raw percentages (R 2 = 0.76-0.94). Settling times of 0.67-min and 2-h were the most accurate for NIR predictions (R 2 = 0.49-0.79). The NIR prediction of sand sieving method (R 2 = 0.66) was more accurate than Bouyoucos (R 2 = 0.53). The NIR 2X gain was less accurate (R 2 = 0.69-0.92) than 4X (R 2 = 0.87-0.95). The MIR (R 2 = 0.45-0.80) performed better than NIR (R 2 = 0.40-0.71) spectra. Adding soil carbon, reconstituted bulk density, pH, red-green-blue color, oxalate and Mehlich3 extracts returned R 2 value of 0.86-0.91 for texture prediction. In addition to slope and intercept of the SDF, 4X gain, method and pre-treatment classes, soil carbon and color appeared to be promising features for routine SGB-processed NIR particle-size analysis. Soil Classification ( Soil Taxonomy ): Inceptisols, Spodosols