The prediction of available nitrogen content has attracted significant attention in soil nutrient diagnosis. The available nitrogen spectrum detection model prediction accuracy can be effectively improved using feature selection and regression prediction algorithms. This study selects 188 yellow-red loam samples in southern Anhui as objects, uses seven preprocessing methods to correct the spectral data. It combines the moving window method and five intelligent optimization algorithms for feature selection. Then, it establishes 36 regression calibration models for analysis and comparison based on different ensemble boosting (Boosting) algorithms. The experimental results show that 202 spectral features selected using the feature optimization algorithm based on particle swarm optimization (PSO) are concentrated in the range of 600--1000 nm. The Adaptive Boosting (AdaBoost) model developed using these features has the best performance, with the prediction accuracy of soil available nitrogen of 0.944. This study improves the prediction accuracy of soil available nitrogen and discusses the optimization algorithm of characteristic interval, which has a certain theoretical value.