Chemometric technique performances in predicting forest soil chemical and biological properties from UV-Vis-NIR reflectance spectra with small, high dimensional datasets

• Published in: IForest (Viterbo) 2016-02, Vol.9 (1), p.101-108
• Main Authors: Bellino, A, Colombo, C, Iovieno, P, Alfani, A, Palumbo, G, Baldantoni, D
• Format: Article
• Language: English
• Subjects: Altitude; Biological effects; Biological properties; Calibration; Carbon; Chemometrics; Datasets; Decomposition; Diffuse Reflectance Spectroscopy; Dimensional analysis; Elastic Net; Forest soils; Generalized linear models; National parks; Noise reduction; Organic carbon; Performance prediction; PLSR; Predictions; Reflectance; Sample Size; Soil chemistry; Soil properties; Soil sciences; SPC/LASSO; Spectra; Spectroscopy; Spectrum analysis; Total organic carbon; Variables; Wavelet transforms; Wavelets
• ISSN: 1971-7458; 1971-7458
• DOI: 10.3832/ifor1495-008

Chemometric analysis applied to diffuse reflectance spectroscopy is increasingly proposed as an effective and accurate methodology to predict soil physical, chemical and biological properties. Its effectiveness, however, largely varies in relation to the calibration techniques and the specific soil properties. In addition, the calibration of UV-Vis-NIR spectra usually requires large datasets, and the identification of techniques suitable to deal with small sample sizes and high dimensionality problems is a primary challenge. In order to investigate the predictability of many soil chemical and biological properties from a small dataset and to identify the most suitable techniques to deal with this type of problems, we analysed 20 top soil samples of three different forests (Fagus sylvatica, Quercus cerris and Quercus ilex) in southern Apennines (Italy). Diffuse reflectance spectra were recorded in the UV-Vis-NIR range (200-2500 nm) and 22 chemical and biological properties were analysed. Three different calibration techniques were tested, namely the Partial Least Square Regression (PLSR), the combinations wavelet transformation/Elastic net and wavelet transformation/Supervised Principal Component (SPC) regression/ Least Absolute Shrinkage and Selection Operator (LASSO), a kind of preconditioned LASSO. Calibration techniques were applied to both raw spectra and spectra subjected to wavelet shrinkage filtering, in order to evaluate the influence on predictions of spectra denoising. Overall, SPC/LASSO outperformed the other techniques with both raw and denoised spectra. Elastic net produced heterogeneous results, but outperformed SPC/LASSO for total organic carbon, whereas PLSR produced the worst results. Spectra denoising improved the prediction accuracy of many parameters, but worsen the predictions in some cases. Our approach highlighted that: (i) SPC/LASSO (and Elastic net in the case of total organic carbon) is especially suitable to calibrate spectra in the case of small, high dimensional datasets; and (ii) spectra denoising could be an effective technique to improve calibration results.