Gridded Hail Nowcasting using U-Nets, Lightning Observations, and the Warn-on-Forecast System

Aug 2, 2026

Upload by Anda

This thesis develops gridded hail nowcasting using U-Nets combined with lightning observations within the Warn-on-Forecast framework. It assembles predictors and labels from GridRad reflectivity and GridRad MESH hail size targets, and derives lightning information via Vaisala detection. Data mining and preprocessing steps, including storm-day clustering, cross-validation, and normalization, enable robust training and evaluation. Multiple experiments test data distribution effects, label-and-predictor configurations, lightning ablations, and U-Net architectures using the Critical Success Index. Results include overall performance and case studies highlighting practical storm behavior.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages119

Reading Time5 h 0 min

AccessFree to download

Introduction

Background

Hail Forecasting

Machine Learning, U-Nets, and Segmentation

The Warn-on-Forecast System

Methods

Datasets

Data Mining and Pre-processing

Machine Learning Architecture (U-Nets)

Hyperparameter Searches and Model Training

Max Critical Success Index Metric

Results

Critical Importance of Data Distribution

Predictor and Label Experiments

Architecture Experiments

Overall Performance

Case Studies

Discussion and Conclusions

Reference List

Appendices

UNIVERSITY OF OKLAHOMA GRADUATE COLLEGE GRIDDED HAIL NOWCASTING USING UNETS, LIGHTNING OBSERVATIONS, AND THE WARN-ON-FORECAST SYSTEM A THESIS SUBMITTED TO THE GRADUATE FACULTY in partial fulfillment of the requirements for the Degree of MASTER OF SCIENCE By TOBIAS GARRICK SCHMIDT Norman, Oklahoma 2023 GRIDDED HAIL NOWCASTING USING UNETS, LIGHTNING OBSERVATIONS, AND THE WARN-ON-FORECAST SYSTEM A THESIS APPROVED FOR THE SCHOOL OF METEOROLOGY BY THE COMMITTEE CONSISTING OF Dr. Amy McGovern, Chair Dr. Corey Potvin, Co-Chair Dr. John Allen Dr. Cameron Homeyer ©Copyright by TOBIAS GARRICK SCHMIDT 2023 All Rights Reserved. Dedication To my dearest parents and brother, who cared for and loved me throughout my life. Their love helped nurture a young scientist’s desire to find some understanding of God’s creation... Acknowledgements Firstly, I would like to thank my advisors Dr. Amy McGovern and Dr. Corey Potvin for taking me on as their student and for supporting me throughout my time at OU. The expertise and knowledge I have gained from them will prove invaluable in future academic or industry endeavours. The many lighthearted moments (and stormshelters) we shared together will always be a source of fond memories for me. I also thank Dr. John Allen of Central Michigan University for generously donating his time each week to discuss thesis topics with a student he wasn’t formally advising. I would also like to thank my mother Susan, my father Garry, and my brother Benjamin for all the support they have given me throughout my academic career. I must also thank my friends Joshua, Conner, Jarod, Alexander, and Nolan back in Canada who kept me encouraged through my degrees. Our many nights spent gaming while I was in the US and our travels when we could unite kept me pushing through. I specifically thank Nolan for the many machine learning discussions we had, which offered unique perspectives on my work. I of course must extend my thanks to Chad for his friendship and expertise on American culture during my time at OU. Thank you for joining me on numerous hair raising adventures, I look forward to the next ones. Studying severe weather at OU was a childhood dream and I must thank you all for making such a dream possible. This material is based upon work supported by the National Science Foundation under Grant No. ICER-2019758 . Funding was provided by NOAA/Office of Oceanic and Atmospheric Research under NOAA–University of Oklahoma Cooperative Agreement NA21OAR4320204, U.S. Department of Commerce. The computing for this project was performed at the OU Supercomputing Center for Education and Research (OSCER) at the University of Oklahoma (OU) . The statements, findings, conclusions, and recommendations are those of the author(s) and do not necessarily reflect the views of NOAA or the U.S. Department of Commerce. Table of Contents Dedication iv Acknowledgements v List Of Tables viii List Of Figures ix Glossary xiii Abstract xv 1 Introduction 1 2 Background 5 2.1 Hail Forecasting ............................... 5 2.2 Machine Learning, U-Nets, and Segmentation .............. 9 2.3 The Warn-on-Forecast System ....................... 14 3 Methods 17 3.1 Datasets ................................... 17 3.1.1 Predictors .............................. 18 3.1.1.1 Warn-on-Forecast System ................ 19 3.1.1.2 Vaisala Lightning Detection ............... 22 3.1.2 Labels (GridRad MESH) ...................... 24 3.1.3 Other Data (GridRad Reflectivity) ................ 27 3.2 Data Mining and Pre-processing ...................... 27 3.2.1 Data Mining (AutoPatcher) .................... 28 3.2.2 Preprocessing ............................ 29 3.2.2.1 Test Set Partition and Storm Day Clustering ..... 30 3.2.2.2 Minima/Maxima and Architecture Data Slicing .... 31 3.2.2.3 Cross-Validation ..................... 33 3.2.2.4 Normalization and Remaining Preprocessing

FAQs

What data sources are used for gridded hail nowcasting in this thesis?

The work uses GridRad reflectivity as additional data, and GridRad MESH as labels for hail size targets. It also incorporates lightning observations from Vaisala lightning detection, along with Warn-on-Forecast system information as predictors.

How are the training and evaluation datasets prepared?

The thesis applies data mining and preprocessing including storm-day clustering, a test set partition strategy, cross-validation, and normalization. It also slices architecture data using minima/maxima-based methods to structure model inputs.

Which experiments assess the value of lightning observations and different modeling choices?

The results include predictor and label experiments such as Gaussian expansion and predicted hail size tests, plus a lightning observations ablation experiment. Architecture experiments compare U-Net configurations, and overall performance is evaluated using the Critical Success Index.

Aug 2, 2026

Upload by Anda