Assessing urban landscape diversity: deep learning approach for street view images

Abstract

Assessing urban landscape diversity is crucial for evaluating urban landscape quality and urban viability. As street view images become more accessible and computer vision technologies, especially semantic segmentation, advance, it has become feasible to measure urban landscape diversity using street view images. This study presents a deep learning approach for assessing urban landscape diversity based on street view images.

The aim can be summarized into three objectives. Firstly, to assess the performance of pre-trained models in the task of segmenting street view images. Secondly, to develop formulas utilizing the segmentation results to measure the urban landscape diversity index. Lastly, to interpret the index and unveil the distinctive characteristics of the urban landscape in a specific city.

To address these objectives, a systematic research framework comprising three components was proposed: data sources, modelling, and index calculation. The pre-trained models trained on the ADE20K dataset were selected and evaluated using the Mapillary Vistas dataset. Based on metrics including object accuracy, pixel accuracy, and mean Intersection-Over-Union (mIoU) score, the PSPNet model was identified as the most robust and utilized for semantic segmentation on Google Street View (GSV) images of Helsinki. Outcomes from the segmentation process were subsequently employed to calculate different view indices (sky, building, tree, grass, road, and car) and the Shannon Diversity Index, providing insights into urban density, vegetation distribution, traffic patterns, and overall landscape quality in Helsinki.

The study is methodology-centered, demonstrates the potential of employing deep learning techniques on street view images for assessing urban landscape diversity, and can be applied to other cities, making it a scalable approach for urban studies.