Abstract: In view of the large redundancy information of hyperspectral soil organic matter data and the uncertainty of the results of the single feature band selection algorithm, this paper proposes to establish a model by combining multiple single algorithms to screen the feature bands. Firstly, 189 soil samples from 13 counties and urban areas of Guizhou Province are collected, and their spectral data are used as data source. Secondly, four feature band screening algorithms, i.e. kings of CARS (Competitive adaptive Reweighted Sampling), UVE (Uninformative Variable Elimination), SPA (Simultaneous Perturbation Algorithms) and IRIV (Iteratively Retained Informative Variables) feature band screening algorithms are combined, and three optimization joint algorithms are selected among the multiple joint algorithms. Finally, soil organic matter content inversion model is constructed based on support vector machine and random forest. The results show that in the soil organic matter inversion model, the determination coefficient of UVE-SPA-SVM model reaches 0.87(R²=0.87), the root mean square error reaches 8.31(RMSE=8.31), and the relative analysis error is 2.72(RPD=2.72), which showes the best performance among the three optimization joint algorithms. The performance is better than four single screening algorithms (CARS, UVE, SPA, IRIV). It shows that the optimized joint algorithm has obvious advantages over the single algorithm in reducing hyperspectral data and improving model accuracy. By comparing the accuracy of different models, this study aims to verify the superiority of the optimized combined algorithm compared with a single algorithm in the feature band screening, so as to provide a more reliable and efficient algorithm support for the high-precision inversion of soil organic matter content in mountainous cultivated land.