Severe Weather Forecasting via the Sole Use of Satellite Data

(with the end goal of tornado prediction via satellite data alongside machine learning and fluid mechanics)

About this project:

This project involves the development of a pipeline for the automated construction of natively visualizable image-based SITS-derived datasets composed of AMVs which can be used as is in a modified version of the MS-RNN framework. The modified MS-RNN framework, as well as the pipeline, are both present in the existing repository, and should be usable as is (with customization as the user sees fit).

Custom Dataset Construction Pipeline:

• The Custom Dataset Construction Pipeline automates the process of generating AMVs (Atmospheric Motion Vectors) from raw GOES-16 ABI data via the use of several scripts designed to work in conjunction with one another. Further details regarding the pipeline, as well as how to use it, are detailed in the README.md file under the pipeline directory.

Modified MS-RNN framework

• The MS-RNN framework has been modified in order to work properly with a custom data loader developed for use with datasets as output from the custom dataset construction pipeline. Configuration for MS-RNN is done in config.py, and configuration for the custom data loader can be done in util/load_goes.py.

Further information regarding the project thus far

• Raw data source used for testing: PERiLS_2022 GOES-16 ABI Mesoscale Sector Data and PERiLS_2022 GOES GLM
  • The latter has yet to be fully implemented/integrated into the project
• Feature engineering and data handling:
  • Raw data automatically organized via Python scripts
  • AMVs (Atmospheric Motion Vectors) generated from 3 channels of the raw data via OCTANE
  • RGB Composition via MatPlotLib
  • Automation of the previous 3 bullet points via the Custom Dataset Construction Pipeline (check the pipeline directory for more info)
  • Not yet fully implemented:
    • GOES-16 GLM (Geostationary Lightning Mapper) data
    • SPOD (Spectral Proper Orthogonal Decomposition) analysis via PySPOD
• Prediction:
  • Several multi-scale neural network models via a modified version of the MS-RNN framework
• Classification:
  • TBD

BD 2023-24