Presenter: Philippe Tissot (Texas A&M University-Corpus Christi)

Description:
Deep learning (DL) methods such as 3D Convolutional Neural Networks and Autoencoders are increasingly used in the environmental sciences to model complex problems while taking advantage of our ever-growing computational power and data availability. The presentation will describe the application of two deep learning approaches to the problem of predicting coastal fog. The models’ predictors consist of numerical weather predictions up to the lead time, satellite sea surface temperature measurements and coastal measurements. The models are calibrated to predict visibility at a coastal airport along the Texas coast for lead times of up to 24 hours. Performance is assessed based on several skills including Pierce's skill score and Heidke skill score and other performance metrics showing significant improvements over operational models such as HREF and SREF for the location. Explainable AI methods including permutations and partition SHAP are used to help understand the relative importance of the models’ architecture and predictors including the Physics based grouping and organization of input feature maps for the CNN and the use of 3D convolutional kernels as compared to 2D convolutional kernels.