Improving the Accuracy of Soil Organic Carbon Estimation: CWT-Random Frog-XGBoost as a Prerequisite Technique for In Situ Hyperspectral Analysis

Abstract

Rapid and accurate measurement of the soil organic carbon (SOC) content is a pre-condition for sustainable grain production and land development, and contributes to carbon neutrality in the agricultural industry. To provide technical support for the development and utilization of land resources, the SOC content can be estimated using Vis-NIR diffuse reflectance spectroscopy. However, the spectral redundancy and co-linearity issues of Vis-NIR spectra pose extreme challenges for spectral analysis and model construction. This study compared the effects of different pre-processing methods and feature variable algorithms on the estimation of the SOC content. To this end, in situ hyperspectral data and soil samples were collected from the lakeside oasis of Bosten Lake in Xinjiang, China. The results showed that the combination of continuous wavelet transform (CWT)-random frog could rapidly estimate the SOC content with excellent estimation accuracy (R2 of 0.65–0.86). The feature variable selection algorithm effectively improved the estimation accuracy (average improvement of (0.30–0.48); based on their ability to improve model estimation on average, the algorithms can be ranked as follows: particle swarm optimization (PSO) > ant colony optimization (ACO) > random frog > Boruta > simulated annealing (SA) > successive projections algorithm (SPA). The CWT-XGBoost model based on random frog showed the best results, with R2 = 0.86, RMSE = 2.44, and RPD = 2.78. The feature bands accounted for only 0.57% of the Vis-NIR bands, and the most important sensitive bands were distributed at 755–1195 nm, 1602 nm, 1673 nm, and 2213 nm. These findings are of significance for the extraction of precise information on lakeside oases in arid areas, which would aid in achieving human–land sustainability.