Fog prediction using deep learning models

The occurrence of fog has adverse impacts to human activities, aviation and water transportation operations. These events can cause postponement and cancellation of flights and accidents between ships and vessels leading to economic costs. The design of accurate models is required to forecast the low visibility events caused by fog. However, the prediction of fog remains a challenge due to the rare occurrence of these events. In this study, a deep learning networks (DNN) was proposed using the output from Numerical Weather Prediction (NWP) to predict 6-hour, 12-hour, and 24-hour lead time low visibility levels in the Corpus Christi Area. The model based on the autoencoder architecture was applied as a post-processing of deterministic NWP model and sea surface temperature (SST) output. The autoencoder was utilized to reduce the dimension of the input features to select a higher order of representation from raw data. By converting data from high dimensional space into a lower dimension, autoencoder models preserve the meaningful properties of original features in an unsupervised learning fashion. A logistic regression was also added to solve the classification problem of visibility level. Additionally, the under sampling and oversampling were also examined to solve the class imbalance problem caused by the less positive cases (fog cases). A 11-year database of NWP and SST was used to develop, train, validate, and test the proposed models to predict the occurrence of fog. The target of the models was categorized into three overlapping classes, including ≤ 1600m, ≤ 3200m, and ≤ 6400m. The prediction skill of these models was evaluated by relative operating characteristic curves and seven different skill scores. The results indicate that the DNN models are able to generate good discrimination for all lead times and visibility categories. The DNN models consistently outperformed an operational NWP model ensemble used by National Weather Service with respect to five skill scores (HSS, PSS, POD, CSI, and ORSS). The performance of the proposed models in dimensional reduction exceeds that of the combination between principal component analysis and logistic regression.