The aim of this study is to investigate the problem of redundant soil selenium content spectral data and high model complexity. Several selenium-containing soil samples were collected, and the selenium content and spectral information of the samples were obtained, The raw spectra were preprocessed using Savizkg-Golag multivariate scatter correction first-order differential (SG-MSC-FD), and the feature wavelengths were screened using stability competitive adaptive reweighted sampling (sCARS) and other algorithms to establish the partial least squares regression (PLSR), support vector machine (SVM), random forest (RF), soil selenium-content seagull optimization algorithm (SOA)-RF prediction models. The coefficient of regression (R²), root mean square error (RMSE) and relative predictive deviation (RPD) values of the models under different feature screenings were compared to determine the best combination model. The results show that the accuracy of the models under different feature filtering is improved. The sCARS algorithm extracts the least number of variables, accounting for only 0.49% of the full band, and the algorithm combined variable combination cluster analysis and genetic algorithm has the highest accuracy. The RF model exhibits better robustness than the SVM and PLSR models, and the inversion accuracy of the models significantly improves with parameter optimization of SOA-RF. In summary, the SOA-RF model with VCPA-GA feature extraction is the best prediction model (R²=0.92, RMSE is 0.08, RPD is 2.911), and it can achieve rapid and efficient inversion of soil selenium content.