Spatio-Temporal Modeling using Deep Learning: Methods and Applications in Real-World Systems

Abstract

Deep learning has emerged as a transformative force in scientific research, particularly in modeling complex spatio-temporal phenomena. This dissertation explores the application of advanced machine learning techniques to address pressing real-world prediction tasks in various domains, including climate science, urban planning, ecology, and economics. Specifically, it demonstrates how the integration of temporal dynamics, spatial dependencies, and multimodal data can significantly enhance predictive performance and support data-driven decision-making.

The first component of this work introduces a framework for predicting large-scale real estate trends. By combining Zillow's temporal data with demographic and socioeconomic indicators from the U.S. Census Bureau, we developed a Transformer-based model that captures long-term temporal dependencies and achieved over 90\% classification accuracy in multiple forecasting scenarios.

The second study presents a graph-enhanced deep learning model to forecast extreme weather events. Using spatially connected climate variables, the model improved predictions of heavy rainfall risks, contributing to early warning systems for flood mitigation and water resource management.

The third project uses a TCN (Temporal Convolutional Neural Network) to predict traffic accident-prone zones on national highways, achieving high accuracy in identifying accident-prone regions.

In the final component, we investigate the short-term prediction of the Leaf Area Index (LAI), a key indicator of vegetation health and ecosystem function. By evaluating CNN-based models across 11 types of global land cover, we show that deep learning can robustly generalize across diverse ecosystems, offering valuable insights for ecological modeling and agricultural planning.