Airport Visibility Classification Based on Multimodal Fusion of Image-Tabular Data

Low visibility is a major cause of flight delays and airport cancellations. Hence, providing accurate prediction of airport visibility is vital to prevent significant losses for airlines, accordingly to avoid catastrophic aviation accidents. Although many unimodal prediction methods have been developed, there is still room for improvement in airport visibility classification. Therefore, this study proposed an airport visibility classification model that employs a multimodal fusion of image and tabular data to improve classification performance. First, an enhanced image processing method on airport visibility was designed to extract more detailed features related to airport visibility. Next, EfficientNetB1 and Feature Tokenizer Transformer (FT-Transformer) were utilized to extract features from the images and tabular data, respectively. These features were then combined to classify airport visibility using a Multimodal Fusion Multilayer Perceptron with a focal loss function, validated through 5-fold stratified cross-validation. Experimental results on the 1864 pairs of image-tabular data from Nanjing Lukou International Airport showed that our model achieved an accuracy of 93.83%, and a Ma-F1 of 91.64%. Comparisons with various image extraction methods indicated that EfficientNetB1 with integrated images provides the best performance and shorter running time. The comparison of multimodal fusion and unimodal image classifications revealed that the accuracy and Ma-F1 of unimodal image classifications are lower than those for multimodal classification by 8.56% and 12.85%. In addition, ablation experiments also proved the effectiveness of the enhanced image processing and multimodal fusion modules.