Importance of 3D convolution and physics on a deep learning coastal fog model

The forecasting of hazardous atmospheric phenomena is often challenging. Artificial intelligence (AI) models have been applied to atmospheric science problems. Model complexity provides a motivation to quantify the importance of model architecture components. We studied the relative importance of the components of the FogNet model that was designed for big atmospheric data: 1) 3D versus 2D convolution, 2) physics-based grouping and ordering of meteorological input features, 3) different auxiliary CNN-based feature learning modules and 4) parallel versus sequential spatial-variable-wise feature learning. We investigate the relative importance of these CNN architectural features by predicting coastal fog,a complex spatiotemporal dynamical process. We use four explainable AI techniques to better understand input feature contributions. The results of the experiments demonstrate that 3D-CNN based models better capture the complexity of the fog prediction process than the 2D-CNNs. We also show that physics-based feature grouping, and the order in which they are fed into the CNNs, significantly impacts performance. • •Evaluating FogNet v1.0 on fused numerical model output and satellite imagery.• •Comparing 3D vs 2D kernel learning ability of the atmospheric vertical structure.• •Evaluating physics-based grouping and ordering of atmospheric input data.• •Using four XAI techniques for discovering input atmospheric variable importance.• •Evaluating using auxiliary modules such as attention to improve the FogNet v1.0.