Meteorological Visibility Estimation Through Multi-Modal Feature Fusion with Convolutional and Frequency Domain Representations

Traditional computer vision approaches for visibility estimation often rely exclusively on spatial features extracted through convolutional neural networks (CNNs), potentially overlooking valuable frequency-domain information available in compressed images that could enhance robustness to environmental variations. In this paper, we present a novel dual-branch deep learning architecture that integrates Convolutional Block Attention Module (CBAM) mechanisms within the CNN branch and employs a cross-attention fusion mechanism to enable non-linear interactions between spatial and frequency features. Additionally, our progressive training strategy prevents feature suppression and ensures balanced contributions from both branches to the final prediction. Experimental evaluation on a comprehensive visibility dataset demonstrates that our multi-modal feature fusion approach outperforms traditional single-modality methods, achieving a 4.7% improvement in MAE and an 8.9% improvement in MSE compared to existing approaches. © 2025 IEEE.