Understanding Climate Change and Variability III: A Spatio-Temporal Data Mining Perspective

This thesis provides a computer science audience with an introduction to mining climate data with an emphasis on the singular characteristics of the datasets and research questions climate science attempts to address. We demonstrate some of the concepts discussed in the earlier parts of the thesis with two climate-related applications of relationship and pattern mining. In both instances, we show that insightfully mining the spatio-temporal context of climate datasets can yield significant improvements in the performance of learning algorithms. We focus on two spatio-temporal data mining applications one predicting Atlantic tropical cyclone (TC) activity and the other on mesoscale ocean eddy monitoring.;Tropical cyclones are among of the most devastating geophysical phenomena and predicting their occurrence has become a subject of intense scientific and societal interest. A large body of research focuses on using the large-scale environmental conditions to forecast cumulative TC activity on interannual scales. One of the known influencers of the large-scale conditions over the Atlantic ocean on interannual time-scales is the quasiperiodic warming and cooling cycle of the Pacific ocean, known as the El Niño Southern Oscillation (ENSO). Several research efforts have focused on capturing the ENSO cycle using empirical indices that average Pacific sea surface temperatures over fixed oceanic regions. These traditional indices have provided limited Atlantic TC forecasting insight, mainly because they have little predictive skill before the Northern Hemisphere spring (commonly known as the "spring predictability barrier"). We introduce the spatial ENSO index (S-ENSO) and show that it is better than traditional ENSO indices at forecasting Atlantic TC activity. Furthermore, S-ENSO is not susceptible to the ENSO spring predictability barrier traditional indices suffer from. Given the numerous global phenomena associated with ENSO, S-ENSO may be useful to researchers forecasting other phenomena associated with ENSO.;Our pattern mining application focuses on monitoring mesoscale ocean eddy dynamics. Mesoscale ocean eddies are coherent rotating structures that transport heat, salt, energy, and nutrients across oceans – also known as the "storms of the sea". As a result, accurately identifying and tracking such phenomena are crucial for understanding ocean dynamics and marine ecosystem sustainability. This thesis proposes several advances to traditional eddy identification and monitoring algorithms. First, we are able to leverage the spatio-temporal context of the data to identify more physically-consistent features than existing methods. Second, we introduce a novel eddy tracking algorithm that not only resolves eddy tracks better than traditional methods, especially in the presence of noise but is also able to take corrective measures autonomously on the independent detection step. Finally, we provide the community with different evaluation metrics to assess the performance of unsupervised learning algorithms in the physical sciences.