Leveraging ML For Automated Urban Analysis Vision-Based CNN Approach Using Satellite Imagery

Urban planners traditionally use manual methods for analyzing urban architectural information such as typology, density, and usage. These analysis methods are time-consuming, error-prone, and may lead to misguided planning decisions. Modern planning tools such as Geographic Information Systems (GIS) contribute to better and faster urban analysis but still demand large time investment from the planners and are not available in many developing countries. We propose that vision-based Machine Learning (ML) models have the potential to contribute to the planning process by automatically mapping and analyzing some aspects of the urban architectural information. We trained Convolutional Neural Networks (CNNs) to test this hypothesis and analyze readily accessible Earth Observation (EO) data such as satellite images. We present a method to classify urban architectural properties such as building typology with a vision-based ML model (Faster R-CNN). The model is trained to identify building typology from publicly available satellite imagery in real time. The initial results of this research demonstrate a promising capability to accurately identify building typologies. The results illustrate the potential to develop methods and tools that allow for automatic mapping and analysis of architectural properties in cities in a way that is impossible with traditional, manual methods.