Machine Learning for Coastal Fog Predictions and the AI2ES National AI Institute

Machine learning methods have shown to be a powerful approach to modeling coastal systems
including combining the nonlinear combinations of atmosphere, ocean, and land forcings. One of the
foci of the NSF AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal
Oceanography (AI2ES) is the development of ML based methods to predict visibility and onset and
duration of coastal fog events. Improving such predictions is of great importance for coastal area
management particularly for air and sea transportation. The reliable prediction of fog with machine
learning is however challenging due to the infrequency of the target event, and the spatiotemporal and
variable inter-dependency of the inputs, along with data non-stationarity. 3D CNN-based models are
able to learn not only 2D spatial patterns and correlations between groups of pixels and a target but also
learn spectral correlations between bands or temporal correlations between input variables. FogNet is a
3D CNN-based model taking advantage of the combination of an atmospheric numerical model output,
sea surface temperature satellite imagery and derived air-sea interaction features. The input to FogNet
consists of up to 384 ordered variable maps organized in five data cubes organized based on the physics
of the problem (1) wind, (2) turbulence, kinetic energy and humidity, (3) lower atmospheric
thermodynamic profile, (4) surface atmospheric moisture and microphysics and (5) sea surface inputs. A
more granular physics grouping will also be tested. The model predicts fog and mist visibility categories
below 1600m, 3200m and 6400m for 6-, 12− and 24−hr lead times with performance comparable or
superior to existing operational models.
As performance continues to improve, often through the use of novel and/or more complex models, it is
important to study and quantify the relative importance of the components and the inputs of these
models. Along with the development of the model, explainable AI (XAI) methods were applied and
adapted for FogNet. Results show that the 3D architecture indeed outperforms several 2D kernels, that
the physics-based grouping of input meteorological variables leads to better performance. XAI also
allowed to evaluate the benefits of different auxiliary modules, the multiscale feature learning and the
parallel and separate spatial-variable-wise feature learning.

Ongoing further developments include a new version of FogNet based on a Vision-Transformer
architecture. The new architecture introduces a multi-view attention method to model more explicitly
nonlinearly correlated inputs and help better understand their interactions, particularly in the spatial,
temporal, and variable dimensions. While the 3D CNN FogNet model and its new transformer-based
architecture have shown significant improvements over present operational models, they would be
challenging to implement operationally. Based on XAI results emphasizing the importance of
atmospheric predictions for the target location, a Variational AutoEncoder (VAE) that would be easier to
implement broadly is being developed. The model inputs are the High-Resolution Rapid Refresh (HRRR)
predictions for the target location and the same calibration for 14 locations along the Texas coast.
As many organizations are working on harnessing the power of Artificial Intelligence to better predict
and manage our environment, we need many more young scientists ready to contribute to this
discovery and operational implementation process. AI2ES trains dozens of scientists from Community
Colleges, to 4-year universities, MS, PhD and a robust cohort of Post Doctoral students. Thanks to
partners at universities, national laboratories, including a Naval Research Laboratory, private industry an
ecosystem is being developed to give a chance for young ML geoscientists to develop withing diverse
environments and be ready to tackle important and pressing challenges as part of convergent teams.