Thursday, 1 February 2024
Hall E (The Baltimore Convention Center)
Weather radar is a vital resource for detecting and quantifying the intensity of precipitation,
hail, thunderstorms, and other weather hazards. Much of the world outside of developed
nations lack access to any form of weather radar which leaves a critical data gap for a
number of applications that require real-time, rapidly updating weather information.
Recently, deep learning (DL) has shown to be a powerful tool for creating global weather
radar analyses that fill these gaps, however training a model that adequately generalizes to
all parts of the globe is challenging. This talk will describe recent work aimed at improving
the performance of the US Air Force's Global Synthetic Weather Radar (GSWR)
capability, which uses a DL model to estimate radar-like depictions from combinations of
geostationary satellite, lightning detection and numerical model fields. Previous versions
of GSWR curated region-specific datasets that were used to train several independent DL
models to cover the entire globe. This methodology did not allow for cross-learning
between different regions. To improve overall performance, a global dataset was
constructed that spans 5 years and contains aligned patches of geostationary satellite,
lightning detection, selected numerical model fields, and a set of 2D fields computed from
the Global Precipitation Measurement Mission's (GPM) dual frequency precipitation radar.
This data was used to estimate composite reflectivity, echo tops, and vertically integrated
liquid. Layers of the deep learning architecture use a technique called weight superposition
to better account for differences in geographic regions and sensor types. It is shown that
weight superposition also allows for more efficient generalization to new datasets. To
improve the quality of output storm depictions over previous versions of GSWR, the loss
function of this architecture also incorporates a conditional generative adversarial network
(cGAN) loss term. We show the impact of adding the cGAN loss term measured using a
range of quality metrics.
hail, thunderstorms, and other weather hazards. Much of the world outside of developed
nations lack access to any form of weather radar which leaves a critical data gap for a
number of applications that require real-time, rapidly updating weather information.
Recently, deep learning (DL) has shown to be a powerful tool for creating global weather
radar analyses that fill these gaps, however training a model that adequately generalizes to
all parts of the globe is challenging. This talk will describe recent work aimed at improving
the performance of the US Air Force's Global Synthetic Weather Radar (GSWR)
capability, which uses a DL model to estimate radar-like depictions from combinations of
geostationary satellite, lightning detection and numerical model fields. Previous versions
of GSWR curated region-specific datasets that were used to train several independent DL
models to cover the entire globe. This methodology did not allow for cross-learning
between different regions. To improve overall performance, a global dataset was
constructed that spans 5 years and contains aligned patches of geostationary satellite,
lightning detection, selected numerical model fields, and a set of 2D fields computed from
the Global Precipitation Measurement Mission's (GPM) dual frequency precipitation radar.
This data was used to estimate composite reflectivity, echo tops, and vertically integrated
liquid. Layers of the deep learning architecture use a technique called weight superposition
to better account for differences in geographic regions and sensor types. It is shown that
weight superposition also allows for more efficient generalization to new datasets. To
improve the quality of output storm depictions over previous versions of GSWR, the loss
function of this architecture also incorporates a conditional generative adversarial network
(cGAN) loss term. We show the impact of adding the cGAN loss term measured using a
range of quality metrics.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner