To achieve efficient nonlinear regression analysis of near-infrared spectra，a near-infrared spectrum regression analysis algorithm based on an improved stacked autoencoder combined with LightGBM is proposed. The algorithm consists of the stacked autoencoder module and the LightGBM module. The hidden layer features obtained from the stacked autoencoder module are sent to the LightGBM module for regression analysis. The structure of the stacked autoencoder module is adaptively determined through a progressive strategy，and the hyperparameters of the LightGBM module are automatically optimized using the Optuna framework. To validate the effectiveness of the method，the reducing sugar，chlorine，potassium，and total nitrogen of tobacco were selected as the research objects，and a model was built using 1 911 tobacco samples. The results were compared with four other near-infrared spectral regression analysis algorithms. Experimental results showed that the average ,$R_{\mathrm{p}},\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}\mathrm{P},R_{\mathrm{p}}^{\mathrm{2}}$,https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=47889235&type=,https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=47889232&type=,25.56933212,4.57200003, of the predicted models for tobacco reducing sugar，chlorine，potassium，and total nitrogen were 0.911 0，0.056 8 and 0.832 8，respectively，and the prediction accuracy of the proposed method was the best among the five methods. Given equivalent performance on the training set，the proposed method achieves 1%-40% higher prediction accuracy on the test set compared to XGBoost. The improved stacked autoencoder combined with LightGBM demonstrates good component regression analysis ability in near-infrared spectral analysis and can be used to construct chemical composition prediction models for tobacco.