Background and aimsDetermining soil organic carbon (SOC) content rapidly and precisely was a prerequisite for conducting land quality assessment and realizing precision agriculture with the help of visible-near-infrared (Vis-NIR) reflectance spectroscopy. However, spectral redundancy and information overlap reduced the spectral identification ability. It was hypothesized that the feature variable and optimal band combination algorithms could efficiently solve the problem of information redundancy.Methods102 soil samples were collected from the lakeside oasis of Lake Bosten in northwestern China for this study. The soil hyperspectral data were pre-processed using the Continuous Wavelet Transform (CWT) method. The characteristic variables were extracted using the Successive Projections Algorithm (SPA), Boruta and Optimal Band Combination (OBC) algorithms. Subsequently, the Random Forest (RF) estimation model for SOC content was constructed.ResultsThe CWT-OBC-RF model using Vis-NIR bands with a scale of 3 demonstrated the best results among all models (i.e., validation R2 = 0.84). The results indicated that estimating the SOC content by combining the CWT with the OBC algorithm achieved better accuracy (R2 of 0.60-0.84, RPD of 1.44-2.44). The screened feature bands represented merely 0.39% of the Vis-NIR bands, with the most significant sensitivity bands being distributed at 406 nm, 764-765 nm, 903 nm, 1125 nm, 1195 nm, 1200 nm, and 1688 nm.ConclusionThe study's findings confirmed that reducing the redundancy of information could improve the potential for estimating organic carbon through combination of CWT and OBC. The findings will provide technical support for SOC estimation in lakeside oasis in arid regions.