Tolulope Ale Successfully Defends His PhD Dissertation
Congratulations Dr. Tolulope Ale
Tolulope (Tolu) successfully defended on Tuesday, November 18, 2025.
Tolu's dedicated contributions have made him an invaluable member of the iHARP and UMBC community. Following 3 successful internships with NASA GESTAR II, Amazon, and Microsoft, Tolu has secured a full-time Microsoft Data Scientist position commencing in 2026., This significant achievement is a testament to his hard work and talent.
Congratulations, Tolu, on this incredible accomplishment! We look forward to your continued growth and success. The entire iHARP community wishes you great success in your next adventure!
_____________________________
Dissertation Title
MULTIVARIATE EXPLAINABLE ANOMALY DETECTION WITH UNCERTAINTY ESTIMATION IN CLIMATE DATA
Committee
- Dr Vandana Janeja, UMBC, Advisor/Chair
- Dr. Jianwu Wang, UMBC
- Dr. Patricia (Patti) Ordóñez, UMBC
- Dr. Nicole Schlegel, NOAA
- Dr. Sudip Chakraborty, iHARP/ UMBC,
- Dr. Ratnaksha Lele, iHARP/ UMBC
Abstract
The multivariate time-series analysis of climate data represents a
crucial yet underexplored field. This is particularly relevant when
examining extreme climate events, such as snow melting in polar regions,
which require consideration of multiple variables to accurately capture
climate extremes. Anomalies in climate data often result from the
interplay of several variables, meaning that what appears anomalous
under univariate analysis may in fact, align with expected patterns once
contextualized within a multivariate framework. This approach more
accurately reflects the interconnected nature of real-world phenomena,
where events seldom occur in isolation. Despite advances in deep
learning for anomaly detection, very few efforts have focused on
analyzing multivariate climate data; this may be due to the lack of
comprehensive annotations and the complexity of climate variables.
Additionally, a significant limitation of existing anomaly detection
algorithms is their lack of explainability, especially in climate data,
where it is crucial to pinpoint which variable most significantly
influences an anomaly score. Beyond merely identifying anomalies, it is
vital to determine the primary variables driving them, enabling targeted
strategies to mitigate such occurrences in the climate domain.
We first propose a Variational Autoencoder (VAE)-based anomaly detection framework called Cluster-LSTM-VAE (CLV) that leveraged correlation-based feature clustering and dynamic thresholding, to capture localized dependencies and complex variable interactions across time. To provide explainability, we develop an unsupervised attribution framework grounded in a counterfactual explanation method to determine variables contributing most to detected anomalies. This approach identifies which climate drivers significantly contribute to anomalous melt events. We further extend our framework to include a comprehensive uncertainty-aware anomaly-detection module. By integrating Three-Cornered-Hat (3CH) error-variance, we estimate data uncertainty and propagate it through the detection pipeline to learn from uncertain data while maintaining reliability.
We performed a comparative evaluation across multiple climate model to demonstrate the performance of the end-to-end pipeline. The results provide robust insights into the simulation of ice-sheet surface melt dynamics, highlighting the reliability of the climate models in representing snow-melt evolution.
Overall, this dissertation delivers a unified framework for detecting, explaining, and quantifying uncertainty in climate anomalies, providing a scalable, interpretable approach for Earth system monitoring. The proposed methods not only offer methodological innovations for machine learning in environmental science but also hold practical implications for policymakers and stakeholders in climate analysis and adaptation planning.
Tags:
Posted: November 20, 2025, 10:10 AM
