Abstract: This work presents the culmination of the
doctoral research by the author in exploring
modern methods of Data Discovery in weather
radar data, improvements in the
cyberinfrastructure concerning multi-
dimensional gridded data, with a concentration
on real-time data streaming, and experimental
use cases involving real world datasets.
Included in this work is a successful method
for the classification of weather radar image
data using convo-
lutional neural networks, with inspiration
drawn from the subfield of Transfer Learning
in the Computer Vision community. Once this
model was developed, it was deployed on
single radar data from each of the radars in
the CASA DFW network to assign labels to
support a human-in-the-loop semi-supervised
method for data discovery in the weather
radar scans. This model has been furthermore
applied to the WSR-88D network of dual-
polarimetric
weather radars in the United States to
demonstrate the model’s generalizability, and
its utility in discovering phenomena of interest
in vast datasets. This work discusses the end-
to-end development of the data discovery
system, with special focus on initial data
labeling, choices and tradeoffs in model
architecture, and training concerns in the
machine learning model. This represents the
first published research known to the authors
on utilizing the power of transfer learning to
transfer the learning of high quality
convolutional neural networks trained on
photographic images to the weather radar
image domain.

Adviser: Dr. Chandra
Co-Adviser: N/A
Non-ECE Member: Dr. Jose Chavez
Member 3: Dr. Margaret Cheney
Addional Members: Dr. Sid Suryanaryanan

Publications:
N/A

Program of Study:
MATH550
ECE513
STAT525
ECE656
CS545
ECE742
GRAD544
N/A