To address the issues of high dimensionality, redundancy, and nonlinearity in near-infrared spectroscopy, which lead to low prediction accuracy and poor interference resistance in established quantitative analysis models, this paper proposes an improved Deep Residual Shrinkage Network (CBAM-DRSN) for spectral feature selection. This method is based on a deep residual network and integrates the CBAM convolutional attention module into the network. Additionally, two adjustment factors are introduced to optimize the adaptive selection of noise thresholds in the residual shrinkage modules. Finally, the Guided-GradCAM is used for the selection and visualization of effective spectral segments. Using this method, the quantitative models built after selecting features related to total sugar and nicotine in tobacco leaves achieved root mean square errors of prediction (RMSEP) of 0.620 and 0.103, respectively, and correlation coefficients (R²) of 0.965 and 0.955, respectively. These results demonstrate higher model accuracy compared to other methods. The CBAM-DRSN feature selection method effectively extracts features relevant to modeling indicators, eliminates noise and redundant information from the spectrum, and improves model accuracy while reducing the model's complexity.