To address the inefficiencies of traditional strict-dense methods for insulator contamination classification, and the poor accuracy of machine learning methods that ignore spatial features, this paper draws on the field of hyperspectral image classification and innovatively uses local regions of the original image as input. This paper proposes a dual-attention multi-scale enhancement network to achieve pixel-level classification of insulator contamination levels based on hyperspectral images. This model utilizes a dual-branch spectral enhancement module, employing max and average pooling to generate channel-attention feature representations. It also uses 1D convolution to capture cross-channel features, enabling nonlinear modeling of spectral information. Furthermore, a multi-scale spatial-spectral feature extraction module is designed to model local spatial information. Using a dual-branch architecture, the spatial features extracted by multi-scale convolution are fused with channel-attention features to extract the spatial feature distribution of contaminated insulators. Experiments were conducted on samples of ceramic insulators with five different contamination levels. Our model has been verified to be effective in classifying the contamination levels of insulators, with an overall accuracy of 95.81%, an average accuracy of 94.01%, and a kappa coefficient of 0.9381. Ablation experiments also confirmed the effectiveness of our proposed module, enabling high-precision pixel-level classification of the contamination levels of ceramic insulators.