Synergizing Multiscale Wavelet Decomposition and Machine Learning Approach for Improved Accuracy of Soil Organic Carbon Content Estimation in the Arid Lakeside Oases

ABSTRACT Soil organic carbon (SOC) content played a vital role in stabilizing oasis ecosystems and regulating carbon sequestration in arid lakeside regions. However, accurate estimation of SOC content using visible‐near‐infrared (Vis–NIR) spectroscopy was often hindered by spectral redundancy and high dimensionality. This study introduced a novel approach by integrating wavelet analysis with machine learning to improve SOC content estimation accuracy in the lakeside oasis of Bosten Lake, Xinjiang. A total of 82 topsoil (0–20 cm) samples were collected, and their SOC content and corresponding Vis–NIR spectra were measured. The hyperspectral data were processed using continuous wavelet transform (CWT) and discrete wavelet transform (DWT). Key spectral features were selected by three algorithms—successive projections algorithm (SPA), Boruta, and competitive adaptive reweighted sampling (CARS)—and used to develop SOC content estimation models based on partial least squares regression (PLSR), back propagation neural network (BPNN), and random forest (RF). Results indicated that CWT outperformed DWT in noise reduction, especially at low decomposition scales (1–5), with a 19.21% improvement. The best CWT‐based model yielded a 23.20% increase in residual prediction deviation (RPD) over the best DWT‐based model. Feature selection further enhanced model accuracy, improving the determination coefficient ( R 2 ) by up to 49.04% and RPD by 58.23%. Among the algorithms, CARS provided the highest improvement, followed by SPA and Boruta. Thus, the combination of CWT‐1‐CARS and the RF algorithm showed the strongest nonlinear modeling performance. The RF (CWT‐1‐CARS) configuration achieved calibration metrics of R 2 = 0.79, root mean square error (RMSE) = 2.57, and RPD = 2.23 to outperform the original spectral models, with an improvement of 63.3% over PLSR (RPD = 1.84) and BPNN (RPD = 1.91). The spatial interpolation analysis showed 91.3% consistency with field‐measured SOC content values, validating the model's practical reliability. The most sensitive spectral response bands for SOC content were primarily located in the visible range (401–504 nm) and the near‐infrared range (1638–2369 nm). This study established a robust technical foundation for accurate estimation of SOC content, for precise ecological monitoring, and sustainable management of arid, lakeside oases.