Multimodal data integration plays a critical role in accurately mapping snow avalanche susceptibility, yet existing models struggle to capture the complex nonlinear interactions among optical imagery, synthetic aperture radar (SAR) backscatter, snowpack measurements, and meteorological variables. Traditional statistical and shallow machine learning (ML) approaches often rely on handcrafted features and linear assumptions, while recent deep learning (DL) methods employ simplistic fusion strategies that overlook higher order dependencies and cross-modal discrepancies. To address these limitations, this study introduces a novel multimodal DL framework combining a covariance-correlation differential transformer (CCDT) for second-order feature extraction with an avalanche differential attention network (ADA-Net) for targeted cross-modal fusion. CCDT constructs differential covariance correlation matrices to reveal subtle interdependencies across 17 topographic, snowpack, and meteorological inputs, and ADA-Net's Siamese transformer architecture amplifies critical discrepancies between modalities. Evaluated on a curated dataset of 150 avalanche events along China's G217 and G218 highways in the central Tianshan mountains (CTM), our CCDT-ADA-Net framework achieves superior predictive performance (ROC-AUC = 0.9903, F1=0.96 ) and spatial accuracy, successfully classifying 97.4% of historical avalanche release areas within its high-susceptibility zone. It significantly outperforms benchmark models, including vision transformer (ViT), convolutional long short-term memory (ConvLSTM), recurrent neural network (RNN), HybridMLP, and TabNet. Permutation importance analysis confirms snow density, elevation, and snow depth as dominant risk factors. This scalable, interpretable approach offers practical utility for operational avalanche forecasting and infrastructure protection in high-mountain regions.