Icing on power transmission equipment not only increases the mechanical load on insulator surfaces but can also lead to arc flashover and insulation failure, thereby posing significant threats to the reliability and safety of power delivery. Traditional approaches, such as manual visual inspection, edge detection-based image processing, and support vector machine (SVM)-based classification, are constrained by complex environmental conditions and unstable meteorological factors, making it difficult to achieve real-time monitoring and accurate classification. To address these challenges, a multimodal deep learning model based on an improved residual network (ResNet) is proposed. The model integrates three features: image features, texture features from icing images, and meteorological data, and enhances classification accuracy through feature-level fusion. First, an improved dehazing algorithm based on the dark channel prior (DCP) is employed to reduce haze interference, significantly enhancing image clarity and contrast. Subsequently, texture features are extracted from the dehazed images using the gray-level co-occurrence matrix (GLCM). These texture features are combined with image features processed using the improved ResNet to comprehensively capture fine structures and surface characteristics of icing. Next, a meteorological dataset comprising temperature, humidity, and wind speed is then constructed and integrated into the model. By fusing image, texture, and meteorological features, robust multimodal feature learning is achieved. Experimental results on real-world insulator icing samples show that the proposed model reaches an accuracy of 92.9% in icing type identification, demonstrating the effectiveness of the dehazing technique and multimodal deep learning framework in improving classification performance.