This study introduces a depth auto-encoding network into spectral feature learning and proposes an improved feature extraction method based on a convolution auto-encoding network (1D-BCAE) to address the impacts of high dimension, nonlinearity, and a lot of noise in the near-infrared spectrum on quantitative modeling. The feature extraction method is applied to the quantitative modeling of key indexes of tobacco using near-infrared spectroscopy, which improves the accuracy and robustness of the model. First, this method uses a one-dimensional convolution kernel and a pooling window suitable for spectral data feature extraction. Second, in the coding process, the BasicBlock module and batch normalization (BN) structure are added to optimize the network structure, which reduce the number of parameters and computation, reduces the noise and nonlinear characteristics of the spectrum, and optimizes the training efficiency of the network. By designing a corresponding connected structure, the parameters of each module in the encoder are passed to the corresponding decoder, which reduces the loss of detailed features in the network training process. The effectiveness of the proposed method is verified by comparing the reconstruction error and root mean square error in experiments. The quantitative models about nicotine and total sugar in tobacco leaves are established using the features extracted by the full spectrum segment and principal component analysis (PCA), convolutional auto-encoding (CAE) network, and 1D-BCAE combined with the partial least squares (PLS) method, respectively. The results reveal that 1D-BCAE can effectively learn the internal structure and nonlinear relationships in high-dimensional data, and the established model performs better. The proposed method can effectively extract the spectral information of the components to be measured, which is critical for developing a robust correction model and reducing model complexity.