CMF-SnowNet: Cross-Modal Fusion With Attent-Guided GAN and Temporal ConvNeXt for Long-Term Snow Cover Prediction

As a pivotal climate system indicator, snow regulates surface energy balance, hydrological cycles, and atmospheric dynamics, exerting a profound impact on global climate change. To overcome cloud occlusion in optical data and spectral deficiency in synthetic aperture radar (SAR) data, we propose CMF-SnowNet, a synergistic framework integrating two innovative modules: Cross-Modal Attention GAN (CMAtt-DDcGAN) for multimodal feature fusion and temporal-ConvNeXt for snow cover dynamic prediction. CMAtt-DDcGAN adopts dual-branch adversarial training and cross-modal attention to mine modal complementarity, while temporal-ConvNeXt augments ConvNeXt with temporal modeling to capture long-range spatiotemporal dependencies. Experiments verify our model outperforms state-of-the-art methods: cross-modal attention reduces MSE by 24.53%, dual-discriminator design decreases error by 26.36%, and dilated/deformable convolutions contribute 22.36% and 21.12% MSE optimization, respectively. This framework supports high-precision snow dynamic monitoring for climate research and water resource management.