Objective Although computer vision has been widely used in urban landscape research, the overall understanding of the application thereof in the research on the relationship between landscape and health is still insufficient. In view of this, this research summarizes current research trends to demonstrate where computer vision is mainly applied and the strengths and weaknesses thereof, so as to provide a reference for future research.
Methods This research searches relevant literature included in CNKI databases and Web of Science core databases during the period from 2000 to 2021. The searching results cover both primary researches and systematic reviews, which are screened in a double-blind manner in line with the pre-defined inclusion criteria, with 20 Chinese and 90 English literature being finally identified. The research conducts a narrative synthesis of the literature identified to explore the current evidence for the application of computer vision in the research on the relationship between landscape and human health.
Results 1) In terms of technology application, semantic segmentation and image classification technologies are relatively mature and easy to operate, and are thus most widely used in the research on the relationship between landscape and health. At present, the semantic segmentation technology is mainly used to extract the proportion of landscape elements such as visible green index, sky visibility rate, enclosure degree, building and road, while the image classification technology is mainly used for streetscape image classification based on environmental perception. 2) In terms of application fields, previous researches mainly explore the association between landscape visual elements and human physiological, psychological and social activities based on the computer vision technology, which specifically includes the association between landscape elements and physical health involving such indicators as chronic non-communicable disease, infectious disease, air pollution, walking and cycling, and the association between landscape elements and mental health involving such indicators as happiness, depression, anxiety and environmental perception. For activities, existing researches mainly the target tracking and target detection technologies to investigate the connection between the spatiotemporal laws of crowd behavior and landscape elements. 3) Although most of the existing researches have the advantages of fast and convenient data acquisition, high operability and low cost, they still have defects in such aspects as algorithm and dataset. Although the computer vision technology may perform better than humans in processing simple images, the accuracy and stability thereof are prone to be affected by light and scale, and need to be further optimized for task processing in a complex environment. Meanwhile, due to the fact that most current researches only adopt a single algorithm, the landscape visual indicators extracted are relatively limited. Besides, in many researches, the data collected only describe the environment or health information at a certain moment, which is typically identified with mismatching in time series, and is different from the dynamic changes in the real world and people's health status. In addition, existing researches mainly adopt streetscape images from Tencent, Baidu or Google. However, these images are different from what is actually seen by pedestrians, and the street environment cannot represent the overall environmental characteristics of the city.
Conclusion Although evidence shows the diversity of computer vision algorithms and applications in the research on the relationship between landscape and human health, there still exist some challenges in the understanding of computer vision algorithms, expansion of research scope, and collection of reliable datasets. More efforts are entailed to further improve computer vision algorithms, promote the integration between social sciences and computer vision, and collect data incorporating different spatiotemporal observations. This research shows that the computer vision technology needs to be further developed in such aspects as the optimization of target detection algorithms, integration into design and post-evaluation processes, expansion of research on the relationship between landscape and social health, and construction of dynamic spatio-temporal datasets.