Using data mining to model and interpret soil diffuse reflectance spectra

Abstract

The aims of this paper are: to compare different data mining algorithms for modelling soil visible–near infrared (vis–NIR: 350–2500 nm) diffuse reflectance spectra and to assess the interpretability of the results. We compared multiple linear regression (MLR), partial least squares regression (PLSR), multivariate adaptive regression splines (MARS), support vector machines (SVM), random forests (RF), boosted trees (BT) and artificial neural networks (ANN) to estimate soil organic carbon (SOC), clay content (CC) and pH measured in water (pH). The comparisons were also performed using a selected set of wavelet coefficients from a discrete wavelet transform (DWT). Feature selection techniques to reduce model complexity and to interpret and evaluate the models were tested. The dataset consists of 1104 samples from Australia. Comparisons were made in terms of the root mean square error (RMSE), the corresponding R 2 and the Akaike Information Criterion (AIC). Ten-fold-leave-group out cross validation was used to optimise and validate the models. Predictions of the three soil properties by SVM using all vis–NIR wavelengths produced the smallest RMSE values, followed by MARS and PLSR. RF and especially BT were out-performed by all other approaches. For all techniques, implementing them on a reduced number of wavelet coefficients, between 72 and 137 coefficients, produced better results. Feature selection (FS) using the variable importance for projection (FS VIP) returned 29–31 selected features, while FS MARS returned between 11 and 14 features. DWT–ANN produced the smallest RMSE of all techniques tested followed by FS VIP–ANN and FS MARS–ANN. However, both the FS VIP–ANN and FS MARS–ANN models used a smaller number of features for the predictions than DWT–ANN. This is reflected in their AIC, which suggests that, when both the accuracy and parsimony of the model are taken into consideration, the best SOC model was the FS MARS–ANN, and the best CC and pH models were those from FS VIP–ANN. Analysis of the selected bands shows that: (i) SOC is related to wavelengths indicating C―O, C═O, and N―H compounds, (ii) CC is related to wavelengths indicating minerals, and (iii) pH is related to wavelengths indicating both minerals and organic material. Thus, the results are sensible and can be used for comparison to other soils. A systematic comparison like the one presented here is important as the nature of the target function has a strong influence on the performance of the different algorithms.