INTERPRETABLE ENSEMBLE METHODS FOR GEOPHYSICAL VORTEX DETECTION AND TIME SERIES FORECASTING

Zayn Abou-Harb

Computational modeling of complex, real-world data, must balance effective scaling to large and noisy datasets, capture phenomena that often resist clean mathematical definition, and produce outputs that can be meaningfully interpreted and acted upon. Deep learning approaches have made remarkable strides on the first two of these challenges, but frequently do so at the cost of the third. This thesis argues that interpretable ensemble constructions, principled combinations of simpler, semantically grounded components, offer a compelling alternative, achieving competitive effectiveness while preserving the transparency necessary for informed decision-making, particularly in situations with noisy data or hard to define objectives, like geophysical vortex detection and time series forecasting. We present two such models, each addressing a different domain. First, we introduce SWIVVEL (Score-Weighted Identification and Visualization of Vortex Evolution and Location), a novel ensemble approach to vortex detection and tracking in large-scale geophysical wind fields. Rather than committing to a single binary definition of what constitutes a vortex, SWIVVEL employs a multi-criteria weighted scoring system that efficiently combines vorticity, aspect ratio, rotational consistency, and temporal persistence into a unified, configurable pipeline. We demonstrate how this holistic approach to vortex identification not only produces interpretable, effective outputs, with a median detection error of 159.78 km against real-world cyclone data, it also yields an information-rich output that is ideal for exploratory data analysis, interactive visualizations, and downstream applications, such as confidence score breakdowns and LLM-assisted semantic interpretation. Our second contribution is the F-MEM (Fourier-Markov Ensemble Model), a lightweight time series forecasting model meant to bridge the complementary strengths of two classical approaches. Signal processing methods excel at capturing long-term deterministic cycles but struggle with stochastic short-term movement, while Markov chain models handle local state transitions well but lose sight of longer-range periodic structure. The F-MEM combines these by using a Fourier transform to identify dominant cycles in a time series and constructing a separate, inspectable Markov chain to model stochastic behavior within each. Evaluated across three deliberately contrasting datasets the model demonstrates consistent improvements over the standalone Markov baseline wherever periodic and stochastic components genuinely co-exist, with the advantage growing at longer forecast horizons. More importantly, however, is the fact that because each sub-model corresponds to a concrete, identifiable timescale, the ensemble's reasoning can be inspected and interrogated in ways that black-box alternatives cannot support. Taken together, these contributions demonstrate that interpretability and effectiveness need not be in tension. By taking inspiration from machine learning's capacity for complexity while pairing it with domain-specific, semantically meaningful feature design, both SWIVVEL and the F-MEM offer a functional template for tackling ill-defined, high-complexity problems in a way that keeps the reasoning behind every output open to examination.

INTERPRETABLE ENSEMBLE METHODS FOR GEOPHYSICAL VORTEX DETECTION AND TIME SERIES FORECASTING

(1)