In this paper, we propose a Near-Infrared (NIR) spectral data processing method (CDAE-DHA) based on the fusion of Convolutional Denoising Autoencoder and Dynamic Hybrid Attention. Which aims to obtain highly robust low-dimensional spatial features through denoising and feature extraction. The method employs CDAE to perform noise reduction on spectral data and combines Dynamic Hybrid Attention to adaptively focus on the key feature regions in the spectral data, to extract more representative features. The extracted features were applied to the Support Vector Regression (SVR) model for fitting to improve the predictive performance of the model. Through comparative experiments with traditional denoising methods and other modeling methods, the advantages of CDAE-DHA in denoising effect and feature extraction capability are verified. In addition, this study also analyzes the interpretability of the attention mechanism and reveals its action mechanism in the process of feature extraction, providing a new idea and method for data-driven NIR spectroscopy modeling.