Strategies for the efficient estimation of soil organic matter in salt-affected soils through Vis-NIR spectroscopy: Optimal band combination algorithm and spectral degradation

Visible and near-infrared (Vis-NIR) spectroscopy is a cost-effective technique for alternative soil physical and chemical analyses for estimating soil properties. The optimal band combination algorithm is an effective method of extracting spectral variables by considering the interaction information between wavebands, but for laboratory Vis-NIR spectral data, this method is susceptible to the "dimensional curse". Here, we hypothesized that properly degrading the spectral configuration (i.e., decreasing the number of spectral bands and coarsening the spectral resolution) can improve the computational efficiency without affecting the prediction accuracy. To test this hypothesis, we constructed six degraded spectral configurations from an initial spectral database (i.e., consisting of 2001 spectral bands acquired with a portable ASD spectroradiometer) with a reduction in the number of spectral bands from 2001 to 19, a coarsened spectral resolution from 3 to 100 nm, and a spectral sampling interval equal to the spectral resolution (i.e., uniform interval sampling). In this study, the databases consisted of 255 soil samples collected from the Ebinur Lake area in Northwest China. The relationship between the soil organic matter (SOM) and the spectra was established using a partial least-squares-support vector machine (PLS-SVM) through two strategies: one is in accordance with the different salinity levels, and the other involves applying the optimal band combination algorithm from each spectral configuration. The results indicated that the soil salinity had a strong negative influence on the performance of the SOM models (R-cv(2), 0.46-0.81). However, the optimal band combination algorithm can improve the sensitivity (R-pre(2), 0.36-0.65) of spectral information and the SOM. Overall, the prediction accuracy obtained through the optimal band combination algorithm was generally superior to that from full-spectrum data. The prediction performance of the optimal band combination algorithm was accurate (R-pre(2) >= 0.85) and stable (RPIQ(pre), similar to 3.20), with a spectral resolution between 3 and 20 nm (i.e., the number of spectral bands decreased from 2001 to 99). Considering the accuracy and time-consuming nature of this approach, the combination of a 20 nm spectral resolution and an optimal band combination algorithm was the most effective method. In summary, this research will guide future studies in transforming hyperspectral datasets into parsimonious representations and uses the optimal band combination algorithm efficiently to determine the informative variable. Furthermore, the optimal band combination algorithm has broad application prospects in soil Vis-NIR spectroscopy and other fields of spectroscopy.