Soil organic matter content prediction using Vis-NIRS based on different wavelength optimization algorithms and inversion models

Purpose Visible and near-infrared reflectance spectroscopy has been proven to be an efficient method for predicting soil properties, and the wavelength optimization can improve the simulation accuracy of SOM (soil organic matter), but the best combination of wavelength optimization algorithms and inversion model is unknown for alpine ecosystem soil. Methods In this study, 269 topsoil samples were collected in the Three-Rivers Source Region of China and were used to build the inversion model of SOM content. Four kinds of wavelength optimization algorithms were conducted, i.e., correlation analysis, uninformative variable elimination (UVE), successive projection algorithm, and competitive adaptive reweighted sampling (CARS) after spectral pre-treatments. Then, partial least squares regression, support vector machine, and random forest (RF) were used to develop the inversion model of SOM content. Various combinations of wavelength optimization algorithms and inversion models were constructed, and the accuracies were compared. Results The combination of the UVE-CARS-RF achieved the highest simulation accuracy (R-2 = 0.902, RPD = 3.218). For the single band selection method, the CARS algorithm has the highest simulation accuracy, especially the combination of CARS-RF (R-2 = 0.899, RPD = 3.133). Conclusion Appropriate combination of the wavelength optimization algorithm and inversion model not only can significantly reduce computational load but improve the prediction precision. In total, RF obtained the best predication effect.

Automated summary

This study revealed that UV/VIS spectroscopy can be used to improve the prediction accuracy of soil organic matter content. By combining four wavelength optimization algorithms and inversion models, and using 269 soil samples, the optimal prediction model was obtained.