AI-edited green space images reveal how design elements shape human impressions

Researchers in Japan have developed a psychological evaluation framework that links the physical features of urban green spaces to the impressions they create, and then used that framework to guide AI-assisted image editing of landscape photographs. In a pilot study published in the journal City and Built Environment, Kenta Yuzawa and Yukihiro Masuda report that when physical elements such as water features and seasonal foliage were added to photographs of poorly rated green spaces, participants’ impressions of those spaces improved markedly.

The study was motivated by a long-standing gap in landscape design practice. Urban green spaces are widely recognized as psychological resources, not merely aesthetic amenities, and previous research has shown that the amount of visible greenery in urban environments influences outcomes such as comfort and a sense of richness. Yet, according to the authors, current design practice tends to focus on spatial configuration and quantitative indicators, while mechanisms for directly integrating psychological impressions and affective evaluations into the design process remain underdeveloped.

To build their framework, the researchers designed a questionnaire informed by adjective-based impression evaluation methods commonly used in landscape research, drawing on a psychological evaluation structure for urban green spaces proposed by Miura and Tobioka in 1993. The participants were seven undergraduate students from the authors’ laboratory. Each participant evaluated 23 photographs of green spaces, one at a time, using a five-point Likert scale. The evaluation covered 22 adjective items as well as five additional indicators: naturalness, scenic quality, comfort, design quality, and overall impression.

Alongside the impression ratings, the researchers classified the physical elements of each image using variables including viewing distance, green view index, vegetation density, tree species, viewing angle, tree height, paved surface, facilities, seasonal variation, users, and flowers. The green view index was calculated with an AI-based tool developed by Japan’s National Institute for Land and Infrastructure Management, which estimates the proportion of greenery visible in street images. Other elements that were difficult to quantify precisely were visually estimated from the photographs.

For the statistical analysis, mean evaluation scores were calculated for each image by averaging participant responses, so the regression analyses used the 23 images as analytical units rather than individual participant responses. Simple linear regression analyses treated the evaluation items as dependent variables and the physical elements as independent variables. The authors stress that, given the exploratory nature of the study, the limited participant group of seven people, and the image-based analytical structure, the regression results should be interpreted as preliminary, hypothesis-generating indicators rather than statistically confirmatory evidence.

Within those constraints, several patterns emerged. Perceived brightness showed a relatively strong negative association with vegetation density, suggesting that densely vegetated spaces tended to be perceived as darker. Viewing distance was positively related to perceptions of openness, indicating that broader visual fields contributed to impressions of spatial openness. Seasonal variation showed a positive association with perceptions of vibrancy. The analysis also identified two broader trends: impressions such as “spacious” and “open” were influenced by similar physical factors, while items such as “beautiful” correlated weakly with specific variables, implying that they are shaped more by overall spatial composition and atmosphere than by individual elements. The presence of water elements and seasonal characteristics emerged as particularly important for enhancing overall impressions.

By integrating the visual evaluation model of Miura and Tobioka with the regression results, the researchers grouped the evaluation items into seven categories based on similarity in influencing factors: openness; physical and atmospheric brightness; unique sense of place; water-rich; harmony between artificial and natural elements; naturalness of greenery; and depth and ambiance of greenery. These seven factors constitute the foundation of the psychological evaluation framework proposed in the study.

The second phase of the research tested whether the framework could function as a practical design guideline. Four of the original images that had received low scores were reused, and new versions were generated using the Generative Fill feature of Adobe Photoshop, powered by Adobe Firefly, a diffusion model-based generative AI system. Waterside elements were added to three of the edited images, autumn leaves were introduced into one to emphasize the sense of the season, the number of trees was increased in another, and the tree arrangement was changed to a more regular composition in one further image. All images were edited under standardized conditions, with consistent viewpoints, resolutions, and predefined modification categories derived from the framework, and generation was repeated iteratively until the intended modification was reflected while preserving the original characteristics unrelated to the target intervention.

A second questionnaire, using the same methodology as the first, evaluated the original and edited images side by side. The results showed that adding water elements generally increased the average scores for “water-rich,” “beautiful,” and “design quality” by more than one point on the five-point scale, and adding autumn foliage increased the average score for “beautiful” by more than one point. In one case, the mean score for “beautiful” rose from 2.86 to 4.29 after waterside elements were added, while the “water-rich” score rose from 2.00 to 4.71. In another image, the introduction of seasonal elements raised the “beautiful” score from 2.00 to 3.71, and in a third, the addition of water features increased the “design quality” score from 2.86 to 4.00.

The authors are careful to frame the study as a preliminary investigation rather than a definitive demonstration. The participants were all Japanese university students, and the framework may reflect culturally situated values, including aesthetic sensibilities associated with Japanese culture such as appreciation of seasonal change, water features, and natural harmony. The authors note that prior cross-cultural research in environmental psychology suggests certain biophilic preferences, including those for natural elements, water, and seasonal variation, show a degree of universality, but the extent to which the framework applies in non-Japanese urban contexts remains an open question. Other limitations include the small sample size, the use of photographs obtained from online sources whose impressions may differ from those of actual spaces, subjective bias in impression evaluation, and the inherent variability of generative AI outputs, which means different generative models or prompts could yield different outcomes. The image selection process also involved researcher judgment about visual plausibility, which may introduce further bias. The researchers plan to expand the participant pool, incorporate more diverse demographic groups, and eventually integrate the framework directly into AI models to enable automated spatial generation and adjustment, with the aim of establishing a methodology for creating green spaces that are psychologically comfortable for their users.

Subject of Research: A pilot study linking the physical elements of urban green spaces to user impressions and using AI-assisted image editing to test design interventions.

Article Title: Psychological evaluation framework for AI-assisted landscape design of green spaces

Article References: Yuzawa, K., & Masuda, Y. (2026). Psychological evaluation framework for AI-assisted landscape design of green spaces. City and Built Environment, 4(1), Article 16. https://doi.org/10.1007/s44213-026-00082-x

Image Credits: AI Generated

DOI: 10.1007/s44213-026-00082-x

Keywords: urban green spaces, landscape design, artificial intelligence, psychological evaluation, generative AI, Adobe Photoshop, image editing, questionnaire, regression analysis, water features, seasonal variation, Japan