Upcoming Events
Dissertation Defense - Department of Geography and Geoinformation Science
Apr 11, 2024, 9:00 - 11:00 AM
GMU Commerce Building 2006
or
Zoom: https://gmu.zoom.us/j/92537370843?pwd=Q3B0Mms4Q0pjNEhGUVppOTIyd2JTZz09
Meeting ID: 925 3737 0843
Passcode: 285618
PhD Candidate: Didarul Islam
PhD of Science, Earth Systems and Geoinformation Sciences
Department of Geography and Geoinformation Science
Title: A Decision-Rule & Spatial Transfer Learning-based Approach for Automated Local Climate Zones (LCZs) Mapping using Multi-source Geospatial and Remote Sensing Data
Dissertation Chair: Dr. Liping Di
Committee Members:
Dr. Ruixin Yang (GMU GGS)
Dr. John J. Qu (GMU GGS)
Dr. Daniel Tong (GMU AOES)
Abstract
Urbanization, industrialization, and population growth are driving rapid changes in global climate patterns, posing significant challenges to human health and environmental sustainability. Local Climate Zones (LCZs) classification offers a structured approach to understanding urban morphology and its relationship with climate, providing valuable insights for urban planning and policy-making. Leveraging remote sensing technologies, this study aims to advance LCZ mapping by addressing key limitations in current classification approaches and integrating spatial information into machine learning models. Using a combination of decision rules based on remote sensing and spatial parameters, this research automates the generation of training samples for LCZ classification on a global scale. By establishing universal decision rules, training samples are generated automatically, overcoming geographic and climatic variations. Additionally, a spatial transfer learning method is proposed to address the challenge where certain categories of training samples are scarce in one geographic location but plentiful in another. This model is designed to integrate local covariates, local spatial information, and global covariates. This integration enables the model to address spatial dependencies and transfer knowledge about scarce categories from locations where they are abundant. Consequently, this improves the precision, accuracy, and scalability of solving local classification problems. The study produces LCZ maps with the proposed method and compares them with existing products, demonstrating significant advancements in accuracy and detail. Statistical analyses confirm the promising performance of the proposed spatial transfer learning model, with overall accuracies consistently above 80%. Visual comparisons reveal discrepancies between LCZ maps generated by the proposed model and those from existing databases, highlighting the superiority of the spatial transfer learning approach. Additionally, the study identifies limitations in current classification approaches, including scale constraints, reliance on supervised methods, and inconsistencies in training data. Recommendations for future research include the refinement of decision rules, integration of more accurate building height data, and consideration of cloud cover in analysis. By addressing these limitations, LCZ mapping holds immense potential for informing urban planning, climate adaptation, and sustainable development efforts globally.