A New Perspective to Near-Infrared Reflectance Spectroscopy: A Wavelet Approach

Abstract

Near-infrared reflectance spectroscopy (NIRS) has been widely used in precision agriculture to quantitatively assess properties of many agricultural objects. The accuracy of NIRS depends largely on an adequate data analysis technique to preprocess raw spectroscopic data. A method for incorporating wavelet analysis as a preprocessing tool into NIRS was developed and evaluated with two datasets: (1) a synthetic spectroscopic reflectance dataset, and (2) an actual soil spectroscopic reflectance dataset. The resultant wavelet regression models were evaluated and compared with conventional regression models on their prediction accuracy and numbers of regressor variables. Physical interpretation of the wavelet regression model was attempted. For dataset 1, the results showed that both the wavelet and conventional (with band-averaging as the preprocessing tool) models yielded r2 values of 0.99 and root mean square errors (RMSE) of 1.5%. For dataset 2, the r2 value and RMSE of the wavelet model for total clay content were 0.83 and 57 g/kg, respectively, both of which were similar to those of the conventional model (with band-averaging, 1st derivative, and partial least squares). However, the wavelet model had fewer regressor variables and was thus better suited to sensor development. It was also shown that, because of the multi-resolution capability of wavelet analysis, the wavelet model had the potential to distinguish narrow and broad spectral absorptions by selecting wavelet regressors at different scales. This capability facilitated physical interpretation of the wavelet model in terms of relating wavelet regressors to true spectral absorptions of targeted objects.