UROP Proceedings 2022-23

School of Science Department of Mathematics 43 A Machine Learning Approach to Study the Relationship Between Urban Morphology and Urban Heat Island Supervisor: FUNG, Jimmy Chi Hung / MATH Student: WANG, Yuanfu / MAEC Course: UROP1100, Fall UROP2100, Spring UROP3100, Summer The study focuses on the urban heat island (UHI) effect in the Greater Bay Area (GBA). In the study, the realtime air temperature and land surface temperature (LST) are compared and analyzed in both daytime and nighttime, which show a higher correlation in the nighttime. Then, the impact of urbanization on the temperature itself and the correlation of air temperature and LST are examined, proving the significance of urban morphology in the UHI effect. Finally, LST models are built with real-time data, which have relatively good predictions but can still be improved by adding more parameters such as date and time. Retrieving of Urban Morphology from Satellite Images and Machine Learning Supervisor: FUNG, Jimmy Chi Hung / MATH Student: HE, Yang / DSCT Course: UROP1000, Summer This article is focused on the application of U-Net 3+ architecture to satellite imagery segmentation. To test the performance (accuracy and computation efficiency) of U-Net 3+, we trained and validated a model with the code and dataset provided by previous students. To study the versatility of the model and to further generate a segmentation of Hong Kong and the Greater Bay Area for practical use, the model will also be tested with the dataset of Hong Kong (in progress). The following introduces the method and the data used during the whole process. It discusses the accuracy, efficiency and some common problems found during the tests. It also includes the limitation of the testing and provides some directions for the future. Retrieving of Urban Morphology from Satellite Images and Machine Learning Supervisor: FUNG, Jimmy Chi Hung / MATH Student: KAN, Wan On / DSCT Course: UROP1100, Fall UROP2100, Spring This article presents a novel pipeline for autonomously extracting urban morphology information from satellite imagery. The pipeline utilizes a customized DeTR model, which includes a prediction head for estimating building height, in addition to the standard bounding box and classification losses. By combining the strengths of the DeTR model for object detection and a preliminary U-Net model for image segmentation, the pipeline achieves improved accuracy, efficiency, and comprehensibility in building detection and 3D building reconstruction. The DeTR model also enables fast-response, memory-efficient prediction on scaled geospatial data, making it an ideal frontend model for the same purpose. Overall, this pipeline offers a sophisticated solution for automated image segmentation and object detection tasks in urban morphology analysis.