A Variational Autoencoder for Coastal Visibility Predictions: Architecture, Performance and R2X Potential

Fog has a substantial economic and safety impact on coastal communities through disruptions of land, air, and sea transportation. This is the case for the northwest Gulf of Mexico, home of the majority of the top ten US ports by tonnage. A recent AI coastal fog prediction model, FogNet (Kamangir et al. 2021), showed improved performance over present operational models while using a 3D CNN architecture and fusing numerical weather output with satellite imagery of sea surface temperatures. The application of eXplainable AI (XAI) to FogNet (Kamangir et al., 2022) provided information as to the impact on performance and the sensitivity of the model to model predictors and geographic locations. The XAI analysis identified the location of the prediction as a dominant factor for most predictors. This suggests that a simpler model solely focused on predictions at the coastal location of interest may approach the performance of a 3D CNN model. The present work describes the architecture and performance of a variational autoencoder (VAE) developed for predicting visibility at 14 locations along the Texas coast as far inland as Victoria and Houston Intercontinental Airport. The hourly predictors are selected from the output of NOAA’s High Resolution Rapid Refresh (HRRR) model, which is compatible with potential operational application of the VAE predictions. The VAE models are trained to predict visibility with lead times up to 12 hours. The model architecture and hyperparameters are optimized using a Tree-Parzen Estimator and the relative importance of the selected predictors is estimated using XAI methods. The geographic generalization potential of the model is tested by training the models at some locations and evaluating performance at other locations held out from model training and validation. The performance of the VAE model is evaluated against that of the HRRR model and IBM’s Forecast on Demand product at 1, 3, and 5 mile thresholds, while the potential for further simplification, tests of spatial coherence, and eventual operational deployment is discussed.

Kamangir, H., Collins, W., Tissot, P., King, S. A., Dinh, H. T. H., Durham, N., & Rizzo, J. (2021). FogNet: A multiscale 3D CNN with double-branch dense block and attention mechanism for fog prediction. Machine Learning with Applications, 5, 100038.

Kamangir, H., Krell, E., Collins, W., King, S. A., & Tissot, P. (2022). Importance of 3D convolution and physics on a deep learning coastal fog model. Environmental Modelling & Software, 154, 105424.