Image Segmentation for Environmental Data Analysis: Investigating Urban Environments through Google Street View and Virtual Reality
Phil Ho Combatir
Co-Presenters: Individual Presentation
College: The Dorothy and George Hennings College of Science, Mathematics and Technology
Major: Computer Science
Faculty Research Mentor: Feng Qi
Abstract:
Urban environments play a significant role in shaping public health, with features such as greenery, infrastructure, and city design directly impacting mental well-being and quality of life. This study explores how AI-based image segmentation can enhance the analysis of urban settings using 360-degree Google Street View (GSV) images. The research aims to provide actionable data for urban planners and public health professionals by identifying and quantifying key elements like roads, buildings, cars, and trees. Two segmentation models, DeepLabV3 with a ResNet-101 backbone and PaddleSeg, were used to analyze over 110 GSV images from various neighborhoods in New Jersey. These models were evaluated using scatterplot analysis to compare their effectiveness in detecting both natural and man-made features in urban environments. Results showed strong agreement between both models for static elements such as roads and buildings. DeepLabV3 demonstrated greater reliability in high-traffic areas and in segmenting the sky, especially when obscured by buildings or trees. However, both models struggled with the segmentation of trees due to complex textures, lighting variability, and seasonal changes. These findings highlight the need for improved training data and model refinement, particularly for natural elements. This research underscores the potential of AI-driven segmentation tools in real-world applications. Beyond urban planning, these models can support disaster response, smart city development, and autonomous vehicle navigation by offering real-time environmental data. Future work will focus on enhancing segmentation accuracy by integrating diverse datasets and improving model robustness across varying urban and environmental conditions.