Start
July 2, 2024 - 12:00 pm
End
Categories
School of Meteorology (Defense)PhD Defense
Expert-Guided Machine Learning for Meteorological Predictions Across Spatio-Temporal Scales
Amanda Burke
Tuesday, July 2nd, 2024
Zoom / 12:00 pm
https://oklahoma.zoom.us/j/94120953482?pwd=w0hj5XwYsWb5mTsxuoeE6FrysEbKO8.1
Abstract:
In satellite imagery, above-anvil cirrus plumes (plumes) serve as critical indicators of impending severe weather, often appearing 30 minutes before reported events. Their real-time identification is particularly valuable in radar-deficient regions, where they offer insights into the convective environment. However, manually labeling plumes is labor-intensive due to their small size and requires specialized expertise. To streamline this process, deep learning (DL) models like Unet have been trained on expert annotated data, enabling rapid, pixel-level classification using diverse satellite inputs. This approach is showcased through a study analyzing plume classification with different spectral data combinations over the contiguous United States in 2020.
Another focus area is advancing machine learning (ML) methods for predicting severe hail events on localized scales. Existing ML models have demonstrated proficiency across the United States during spring and summer but have struggled to capture the nuanced spatio-temporal dynamics of thunderstorm development in local contexts. Addressing this gap, a novel localization technique prioritizes storm object weighting within the High-Resolution Ensemble Forecast system version 2. This strategy enhances the accuracy of hail forecasts without imposing substantial additional burdens on model developers. Results indicate that localized weighting of storm objects not only matches but often surpasses the performance of traditional ML approaches, effectively pinpointing regions at highest risk of hailstorms based on relevant seasonal conditions.
Lastly, leveraging the expansive archives of MODIS satellite data, this research tackles the challenge of accurate global land cover classification. Unlike conventional methods reliant on expert-selected data, the study explores clustering approaches to extract regional nuances amidst the vast dataset’s strong global signal. Employing MODIS surface reflectance bands in a random forest (RF) model, the study compares classification outcomes between expert-chosen and randomly sampled global datasets. Despite visually superior results from clustering-based RF models, quantitative evaluations reveal comparable classification skills across different sampling methodologies, underscoring the importance of holistic evaluation strategies in land cover assessment.