Coarse to superfine: can hyperspectral soil organic carbon models predict higher-resolution information?

Abstract

Introduction: Modeling and mapping of soil organic carbon concentration and distribution at the pedon scale is a current knowledge gap that can be addressed by laboratory-based hyperspectral imaging and chemometric analysis of soil cores. Despite the advancements in soil organic carbon models based on hyperspectral images, it is not clear how these models will perform upon input with images at higher resolutions than those of their training sets. This study aims to measure the generalizability of a soil organic carbon model based on a test set with higher resolution hyperspectral images than that of its training set.Methods: Organic carbon contents were measured at 10 cm intervals on eight soil cores for use as the training set and at 1 cm intervals on a single core for use as the test set. Three regression models, namely, multilayer perceptron, partial least-squares, and support vector regressions, were trained and tested with the median of each hyperspectral image for each of these intervals as the training and test predictors. Permutation importance analysis was performed to explain the models.Results: The results show that although all three models had the same validation R2 of 0.92 for cross-validation on the 10 cm data, multilayer perceptron regression allowed the best generalization with a test R2 of 0.96 compared to the partial least-squares regression (0.81) and support vector regression (0.86). It was demonstrated that the multilayer perceptron model is more robust to soil surface anomalies and that it predicts soil organic carbon on the test set by learning the spectral features related to soil organic matter chromophore activity in the 950–1,150 nm region along with clay mineralogy derived from peaks at 1,400, 1,900, 2,200, 2,250, and 2,350 nm.Conclusions: This study shows that while the regression models based on hyperspectral images perform well at the 10-cm-resolution cross validation, multilayer perceptron regression shows superior generalization and robustness for a higher 1-cm-resolution test set without much loss of prediction power.