Deep Learning Approach to Floor Area and Building Material Stocks Estimation Using Aerial & Street View Image

Abstract

Urban centers contribute substantially to global greenhouse gas emissions, and with ongoing urbanization, the demand for construction materials is set to rise. This paper addresses the challenge of quantifying building material stocks (MSs) in urban landscapes, a critical step in mitigating the environmental footprint of urban development. Traditional methods of estimating MSs often falter due to the lack of granular building data. We propose a novel solution by employing deep learning to derive MSs estimates from readily available aerial and street-view imagery. Our methodology involves the development of two deep learning models that adeptly classify building types and predict floor areas, respectively. The models demonstrate exceptional performance, with building type classification accuracy reaching 84.71% and floor area predictions achieving a mere 1.86% error. These predictions facilitate an MSs estimation of concrete and total building materials with errors as low as 2.07% and 0.29%, respectively. The successful application of these models illustrates a scalable and effective approach to MSs estimation, thereby aiding numerous cities in planning for a sustainable, circular economy where conventional methods are impractical.