Identification of Streetscape Compositions: A Deep Learning Approach

• Autor(a) principal: Favarão Leão, Ana Luiza
• Data de Publicação: 2020
• Outros Autores: Queiroz Abonizio, Hugo, Barbon Júnior, Sylvio, Kanashiro, Milena
• Tipo de documento: Artigo
• Idioma: por; eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://doi.org/10.47235/rmu.v8i1.140
• topic: morfologia urbana; ambiente construído; aprendizado profundo; classificação de imagens; google street view; urban morphology; built environment; deep learning; image classification; google street view

The environment’s composition can have an impact on human behavior, however, this relationship remains uncertain until the cities' qualities and landscape can be analyzed empirically. Images obtained through Google Street View (GSV) enable a large volume of data for automated assessment of environmental characteristics. Deep learning techniques have advanced in the identification of compositional elements of the built environment. In this sense, this study seeks to investigate and test a procedure for identifying the configuration and composition of the urban landscape, classifying images obtained from GSV through a deep learning approach. From an image dataset of three different neighborhoods in Londrina-PR, a deep learning model for image classification was proposed. The model had a good performance, correctly attributing 87.6% of the samples to the corresponding neighborhoods in the case study. Compositional characteristics were empirically identified, considering the distribution of the samples in the obtained search space. The proposed model contributes to the definition of spatial units as well as in the measurement of environmental qualities, optimizing data collection, expanding sample sizes, and providing objectivity to results. This approach contributes to the expansion of city's analytical scales, identifying compositional and relational patterns in the understanding of elements influent in human behavior.