Coarse to Super-Fine: Can Hyperspectral Soil Organic Carbon Models Predict Higher Resolution Information? Provisionally Accepted

Shayan Kabiri^1* Sharon O'Rourke¹

• ¹University College Dublin, Ireland

Modelling and mapping 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 advances in soil organic carbon models based on hyperspectral images, it's not clear how these models will perform on input with higher resolution than their training set. This study aims to measure generalization power of a soil organic carbon model based on hyperspectral images for a test set that has higher resolution than the model's training set. To do this, for eight soil cores, organic carbon content was measured at 10cm intervals to be used as the training response and for one core, organic carbon content was measured at 1cm intervals to be used as the test response. Three regression models, multilayer perceptron, partial least square regression and support vector regression were trained and tested with the corresponding median of hyperspectral images for each of these intervals as the training and test predictors. In addition, permutation importance analysis was performed to explain the models. It has been shown that although all of the three models had the same validation R 2 of 0.92 for the cross-validation on 10cm data, multilayer perceptron regression had the best generalization with test R 2 of 0.96 compared to partial least square regression and support vector regression with R 2 s of 0.81 and 0.86. Moreover, it was demonstrated that multilayer perceptron model is more robust to soil surface anomalies. In addition, it was shown that the multilayer perceptron predicts soil organic carbon on the test set by learning spectral features related to soil organic matter chromophore activity in 950-1150 nm region and clay mineralogy derived from peaks at 1400, 1900, 2200, 2250 and 2350 nm. This study shows that while all the regression models based on hyperspectral images perform well at a 10cm resolution cross-validation, the multilayer perceptron regression shows superior generalization and robustness when tested at a higher 1cm resolution test set without losing much prediction power.