Estimating soil texture from vis–NIR spectra

Summary Vis–NIR spectroscopy is a low‐cost method for proximal soil sensing, enabling rapid analysis of soil texture as an alternative to more laborious analytical methods. In this study we used partial least squares regression (PLSR) and random forest (RF) models trained on vis–NIR spectra of 173 soil samples from across Germany to estimate soil sand (63–2000 µm), silt (2–63 µm) and clay (< 2 µm) contents. Models were trained with different spectral processing methods. Texture was also estimated by averaging results across different models and by calculating each fraction as the difference from 100% of the sum of the other fractions. The PLSR models predicted clay and sand best, whereas RF performed better for predicting silt fractions. Spectral processing did not improve clay predictions, but standardized normal variate spectra predicted the silt fraction best, and log‐inverse spectra improved modelling of the sand fraction. The best models explained > 90% of variance in the evaluation samples of the textural fractions. Residual prediction deviations were > 3 for all fractions, which together with good accuracy indicated excellent model performance. Model averaging across the top three performing models improved the predictive performance of all fractions. Calculating each fraction from the sum of the other two fractions predicted only the sand fraction well, but increased model bias. In addition, we evaluated the models for their reliability at predicting the cumulative sum of the three fractions for samples of unknown textural composition (n = 1186). In a novel approach, we identified poorly predicted samples by propagating errors from the model uncertainties for the individual soil fractions, namely those samples whose cumulative sum lies outside the range covered by 100% ± propagated error from the three fractions. This new method provides the opportunity to optimize analyses efficiently because these samples can be prioritized and used to check and update the models. Highlights Partial least squares regression and random forest models of vis–NIR spectra predict soil texture reliably. Cumulative sums of independently predicted fractions enabled error estimation. Error propagation allowed robust classification of the whole soil texture. Error propagation identified poorly predicted samples, optimizing laboratory resources.