A Visibility Level Evaluation and Prediction Algorithm for Airports Based on Multimodal Information

To address the issue that many existing visibility prediction methods rely primarily on single numerical data and insufficiently consider the integration of multimodal information, leading to suboptimal accuracy, this paper proposes a visibility prediction model based on multimodal information. By analyzing collected image and atmospheric data, various correlation analysis methods are employed to extract key influencing factors. The influencing factors in this paper include atmospheric factors such as humidity, temperature, and air pressure. A multimodal feature extraction network is constructed, utilizing an improved ResNet50 and BP neural network for feature extraction. A feature fusion layer is designed to facilitate the interaction of modal information through early fusion, and to evaluate the visibility level at the current time. Considering the temporal characteristics of visibility, the Informer encoder-decoder network is introduced to predict visibility trends. Ablation and comparative experiments demonstrate that the proposed multimodal network outperforms unimodal models and other temporal networks in robustness and accuracy across various metrics in visibility prediction and evaluation tasks.