Domain-Specific Machine Learning Approaches for Geospatial Problems

This dissertation explores novel algorithms for complex geospatial problems at the intersection of environmental, social, and computational sciences. Emphasizing the unique challenges of the geospatial domain, particularly the deviation from the independent and identical distribution (IID) assumption, the research spans various methodologies across different domains, demonstrating the benefits of specialized approaches in spatial analysis. First, we show that machine learning techniques can be effectively used in environmental modeling, which often has severe class imbalance challenges. Using artificial neural networks (ANN), support vector machines (SVM), and extreme gradient boosting (XGB) and techniques to address class imbalance provides insights into groundwater quality assessment, focusing on pesticide and nitrate contamination. Second, our research advances the prediction of complex social phenomena at high spatial resolutions and assesses the impact of geographic context using predictive performance. The development of both context-dependent and context-independent approaches, augmented with uncertainty quantification and feature importance analysis, enables a better understanding of social unrest drivers while offering reliable predictions. Third, our work addresses the fundamental challenge of geo-localization of non-georeferenced remotely-sensed imagery (NRSI) through an innovative two-stage deep learning framework. The integration of contrastive learning and regression, coupled with a specialized loss function, enables geographically-aware modeling that significantly improves geo-localization precision over alternate approaches. These advances in environmental science, social science, and geospatial artificial intelligence demonstrate that specialized machine learning approaches can effectively address complex geospatial problems. Advisors: Ashok Samal and Stephen Scott