This tutorial will introduce participants to neural networks, Keras, convolutional neural networks, and ways to interpret what the neural network has learned. Participants will use a neural network to predict whether a storm will produce strong rotation near the surface based on radar reflectivity and wind information. After training a convolutional neural network, techniques like permutation feature importance, saliency maps and backwards optimization will be used to identify different storm structures associated with weak and strong rotation.
David John Gagne is a Machine Learning Scientist in the Computational and Information Systems and Research Applications Labs at NCAR. He received his PhD in Meteorology from the University of Oklahoma in 2016 and served for 2 years as an ASP Postdoc. His research focuses on developing machine learning modeling systems for high impact weather and physics parameterizations.
Notebooks, instructions and additional material
https://github.com/djgagne/ams-ml-python-course
