Keyword: Artificial Inteligence, Spatial Database, Spatial Modeling, Web Developement, Data Analysis
2023
Research
Remote Sensing
Deep Learning
Vacant Lot Detection: Deep Learning Segmentation Model of Philadelphia is a research project that applies computer vision to identify and map vacant land parcels across the city using satellite and aerial imagery. By combining deep learning architectures with geospatial data, the project seeks to provide planners, policymakers, and community organizations with a tool for tracking land use and identifying opportunities for urban greening and redevelopment.
Philadelphia’s 2021 Land Bank Strategic Plan estimated more than 40,000 vacant land parcels across the city, yet their identification and monitoring remain fragmented. Recognizing the potential of these underused spaces for community gardens, affordable housing, and environmental resilience, the project aimed to create an automated, scalable detection method that supports data-driven decision-making for sustainable urban planning.
High-resolution aerial imagery and labeled data from Philadelphia’s Vacant Property Indicators were processed into 512×512 tiles, cleaned, and augmented for model training. Three segmentation models were tested, including a baseline U-Net, a U-Net with MobileNetV2, and a U-Net with EfficientNetB0. Model performance was evaluated using Intersection over Union, Dice coefficient, and F1 score. The U-Net with MobileNetV2 encoder achieved the highest accuracy and generalization, effectively detecting vacant lots across the city.
Developed in Philadelphia, Pennsylvania, in 2023, as part of the Master of Urban Spatial Analytics (MUSA) program at the University of Pennsylvania.
The project was advised by Professor Guray Erus and conducted by Minwook Kang and Teresa Chang. Data analysis was performed in Python, and model development utilized U-Net, MobileNetV2, and EfficientNetB0 deep learning frameworks.
The MobileNetV2-based U-Net model achieved superior accuracy across all performance metrics, successfully identifying vacant parcels and detecting an overall increase in total vacant lot area between 2018 and 2021. Its adaptable framework can be applied to other cities to support land management, urban greening, and sustainable development. The project illustrates the growing potential of deep learning to uncover hidden spatial patterns and empower more informed, equitable urban planning.