Soil organic carbon content estimation with laboratory-based visible-near-infrared reflectance spectroscopy: feature selection.

This study, with Yixing (Jiangsu Province, China) and Honghu (Hubei Province, China) as study areas, aimed to compare the successive projection algorithm (SPA) and the genetic algorithm (GA) in spectral feature selection for estimating soil organic carbon (SOC) contents with visible-near-infrared (Vis-NIR) reflectance spectroscopy and further to assess whether the spectral features selected from one site could be applied to another site. The SOC content and Vis-NIR reflectance spectra of soil samples were measured in the laboratory. Savitzky-Golay smoothing and log10(1/R) (R is reflectance) were used for spectral preprocessing. The reflectance spectra were resampled using different spacing intervals ranging from 2 to 10 nm. Then, SPA and GA were conducted for selecting the spectral features of SOC. Partial least square regression (PLSR) with full-spectrum PLSR and the spectral features selected by SPA (SPA-PLSR) and GA (GA-PLSR) were calibrated and validated using independent datasets, respectively. Moreover, the spectral features selected from one study area were applied to another area. Study results showed that, for the two study areas, the SPA-PLSR and GA-PLSR improved estimation accuracies and reduced spectral variables compared with the full spectrum PLSR in estimating SOC contents; GA-PLSR obtained better estimation results than SPA-PLSR, whereas SPA was simpler than GA, and the spectral features selected from Yixing could be well applied to Honghu, but not the reverse. These results indicated that the SPA and GA could reduce the spectral variables and improve the performance of PLSR model and that GA performed better than SPA in estimating SOC contents. However, SPA is simpler and time-saving compared with GA in selecting the spectral features of SOC. The spectral features selected from one dataset could be applied to a target dataset when the dataset contains sufficient information adequately describing the variability of samples of the target dataset.