dc.description.abstract | This dissertation emphasizes the contribution of expert knowledge in the development and assessment of machine learning (ML) models within the Earth sciences, specifically Meteorology. Despite the common focus on achieving high skill scores, conventional metrics may inadequately capture the nuanced patterns learned by these models. This dissertation underscores the importance of incorporating end-user feedback, demonstrating that with this feedback, tailored yet flexible ML models can effectively learn specific meteorological patterns while remaining applicable to broader contexts.
The first focus of ML development is in identification of above-anvil cirrus plumes (plumes). In satellite imagery, 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 and requires specialized expertise. To streamline this process, I develop a deep learning (DL) model trained on expert-annotated data to create skillful pixel-level plume classifications using remote sensing data that is available globally. This approach was tested on combinations of spectral data across the contiguous United States, showing above-average object correspondence with human-derived labels.
Another focus of this dissertation is leveraging ML models for severe hail prediction 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, I develop a novel localization technique that prioritizes storm object weighting without imposing substantial additional burdens on model developers. Results indicate that localized weighting of storm objects matches or outperforms existing ML approaches, while improving the physical relevance of the top predictors in the trained ML model.
Lastly, leveraging extensive satellite data archives, this dissertation addresses the challenge of efficiently creating training sets that accurately represent large-scale Earth science datasets. This work explores clustering approaches to capture regional nuances within a vast dataset of remote sensing data, focusing on a straightforward use case with an established baseline: land cover classification. By using surface reflectance bands in a random forest (RF) model, I compare classification outcomes between randomly sampled datasets of varying sizes and datasets created using clustering. The clustering approach produced a training sample that was 200% smaller than the largest sample studied, yet it achieved a 77% increase in F1 score. This suggests that clustering may offer an effective alternative (or addition) to increasing computing power when modeling "Big Data". | en_US |