Date of Award
Spring 2025
Language
English
Embargo Period
1-27-2025
Document Type
Master's Thesis
Degree Name
Master of Science (MS)
College/School/Department
Department of Atmospheric and Environmental Sciences
Program
Atmospheric Science
First Advisor
Oliver Elison Timm
Second Advisor
Zheng Wu
Subject Categories
Atmospheric Sciences
Abstract
Extreme events pose challenges in prediction on account of their low frequency and erratic nature. Among these events, Kona lows are rare subtropical cyclones that significantly impact the weather pattern of Hawaii, sometimes causing heavy rainfall, strong winds, and coastal flooding. In the past, studies have used traditional statistical methods such as synoptic climatology, regression analysis, and manual classification that rely on expert judgment to identify Kona lows; however, they often struggle to capture the complex spatial patterns of the Kona storms. Convolutional Neural Networks (CNNs) exhibit a property called translation invariance, which allows them to learn and recognize Kona lows regardless of any changes in the spatial patterns. The goal of this study was to assess the CNN’s ability in correctly identifying extreme events. It uses the zonal wind data of Hawaii from the ERA5 reanalysis between the years 1990 and 2010 as an input. Several CNNs are trained using techniques like undersampling, oversampling, and manual weight distribution to overcome data imbalance, which occurs due to the high frequency of non-extreme events compared to the rare Kona lows. The best model achieved an Extreme Dependency Score (EDS) of 0.84, where 1 represents the best possible score reflecting perfect prediction with no false positives and a frequency bias of 0.90, which suggested slight underprediction. Combining EDS with the bias provides a more comprehensive picture where the scores indicate that the model is well calibrated and non-biased. The CNN model outperforms the reference random prediction and a simple artificial neural network model that uses the numerical PC1 ( First Principal Component ) and PC2 (Second Principal Component ) values derived from the Kona lows as an input. This is followed by a physical interpretation of the forecast using one of the Explainable AI (XAI) approaches. Randomized Input Sampling for Explanation (RISE) explains how the CNN model makes decisions by randomly masking a part of the input image to see how it affects the model output. The composite of the resulting heat map of the true positive events indicates the features of the zonal wind that are important for prediction .The heat map is consistent with the observed trend that Kona storms typically form to the southwest of Hawaii, as this area shows the highest importance, followed by the region surrounding the Hawaiian islands. These results demonstrate the potential of using deep learning techniques in forecasting rare events, promising avenues of future research in climate risk assessment.
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Sharma, Anadhi, "USING CONVOLUTIONAL NEURAL NETWORKS TO IDENTIFY RARE WEATHER EVENTS: APPLICATION TO KONA LOW CLASSIFICATIONS WITH LARGE-SCALE WIND PATTERN" (2025). Electronic Theses & Dissertations (2024 - present). 107.
https://scholarsarchive.library.albany.edu/etd/107
