Soil Organic Carbon (SOC) is the carbon component of Soil Organic Matter (SOM) and is crucial for maintaining the balance and stability of forest ecosystems. Traditional methods for analyzing the organic matter content in soil involve chemical analysis, which is time-consuming and labor-intensive, and generates chemical wastewater that pollutes the environment. Hyperspectral technology offers a non-contact, efficient means of detecting soil nutrient information. Addressing the limitations in the accuracy and computational efficiency of existing machine learning models for soil organic matter prediction, this study uses soil samples from Guangxi State-owned Huangmian Forest Farm and State-owned Yachang Forest Farm. Using full-spectrum data, Principal Component Analysis (PCA) was employed to select the optimal wavelength number for feature bands. Fractional-order differentiation, which processes data more precisely than first-order differentiation and balances spectral noise and resolution, was used as one of the preprocessing methods to transform the spectral data. Finally, the Dung Beetle Optimizer (DBO), known for its higher robustness and fault tolerance compared to traditional centralized algorithms, was used to optimize the parameter combination of the Gaussian kernel function in Support Vector Regression (SVR). The results indicated that the PCA-DBO-SVR model effectively improved the coefficient of determination (R2) for soil organic matter prediction and reduced the Root Mean Square Error (RMSE). The PCA-DBO-SVR model demonstrated the best generalization performance and accuracy among the compared prediction models, with a validation set R² of 0.942 and an RMSE of 2.989 g·kg^-1, showcasing excellent accuracy.