Study design

To answer the two guiding research questions R1 and R2, we applied different analysis steps based on landscape photographs (Fig 1, see following sections). The original photographs (A) had different degrees of cloud cover, and a cloud-free version (B) was created for each photograph. People’s preferences were collected in an online survey. Predictor variables were derived from three methodological approaches, including eye-tracking simulation, photo content analysis, and Geographic Information System (GIS)-based analysis [53]. Finally, statistical analyses were carried out to test for differences between picture pairs (R1) and to explain variations in preferences scores (R2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Analysis steps to assess the effect of clouds in landscape photographs on people’s preferences.

For each photograph (A), a manipulated version without clouds was created (B). People’s preferences were collected in an online survey and the picture pairs were tested for differences in preferences between A and B to answer research question R1. For all pictures, various predictor variables were calculated using three different approaches [53]. The variables were related to the preference scores via regression analysis to explain variations in preference scores and answer research question R2.

https://doi.org/10.1371/journal.pone.0288424.g001

Photo-based survey

To evaluate the influence of clouds on landscape preferences, we used landscape photographs from two previous studies [25, 54]. The photographs were taken in various locations of the Central European Alps and depicted different mountain landscapes that were typical for this area, including rural settlement areas, agricultural areas, alpine pastures, and high mountain landscapes. All photographs showed a 360° panorama and were taken at normal eye level (approx. 1.5–1.7 m) during sunny summer days. When taking the pictures, attention was paid that the horizon did not extend over the center of the picture [55]. Despite a blue sky, the photographs contained variations in cloud patterns, ranging from almost no clouds to a sky largely covered by clouds, but no threatening cloud atmosphere.

From this pool of 49 landscape photographs [25, 54], we selected 29 photographs with varying cloud cover (S1 Fig) to not overload the respondents with a too high number of pictures. The sky of all original photographs (A) was manipulated by removing the clouds and recoloring these areas to derive a version of the photograph without clouds (B), using Adobe Photoshop image processing software (Adobe Photoshop CS6, Adobe Systems Inc, CA, USA). The manipulation only altered the appearance of the sky, and no landscape features were changed (S1 Fig).

All original (A) and manipulated (B) photographs (n = 58) were arranged in an online questionnaire using LimeSurvey (LimeSurvey GmbH, Hamburg, Germany; www.limesurvey.org/). The questionnaire was prepared in German and contained three sections. In the first section, participants were informed about the scope of the survey and that the participation in the survey was voluntary and anonymous (see ethics statement below). In the second section, we asked the respondents to express their preference for each photograph on a 10-point Likert scale (1 = ‘I don’t like it at all’ to 10 = ‘I like it very much’). There is no consensus among landscape preference studies on the most appropriate rating scale. We used an even scale, because we wanted the respondents to decide whether their preference for an image is more positive or negative. Moreover, response scales from 7 to 10 are most reliable, and respondents prefer the 10-point scale the most [56]. All participants assessed all 58 photographs, but the order of the pictures was randomly arranged for each respondent to avoid priming effects [57]. In the third section, participants could also indicate socio-demographic information such as gender, age group, language group, place of residence (city, country), school education and field of work.

The online survey was carried out between March and July 2022, using a convenience sampling approach. We invited students in ecology of the University of Innsbruck (Austria) and students in physical geography of the Ludwig Maximilian University in Munich (Germany) during lectures on applied ecology and on environmental management to participate in the survey by providing a link to the questionnaire. We also asked them to forward the link to friends and relatives (particularly parents/grandparents, but no minors). The participants received neither a compensation nor other incentives. They completed the questionnaire in an unsupervised setting using a digital device and the responses were saved in a database. We estimated that the time to fill it out was about 10 minutes. There were no time restrictions in viewing and rating the photographs. The participants could also pause the survey by saving their responses and continue at a later time.

To ensure a minimum sample size for statistical analysis, i.e., to set up a significant regression model, we determined the minimum sample size through statistical power analysis [58]. With a determination coefficient of R² = 0.7, a statistical power of 0.9 and a significance level of α = 0.01, we needed a sample size of at least 68 valid responses. In total, 176 people opened the questionnaire and 124 participants completed it. The participants were mostly female (70.3%), under 25 years old (66.4%), and residents in North Tyrol (69.2%). Most of them had a highschool degree (69.7%) and were not (yet) employed/still in eduction (63.3%). For full demographic details, see S1 Table.

Due to the convenience sampling approach, we first evaluated the representativeness of the sample, i.e., whether our results were specific for this group of respondents or whether they matched the preferences of the general public. For all pictures, we first calculated mean preference scores and then compared those of the original photographs in our study to the results of the two former surveys, from which we selected the photographs [25, 54]. Both studies gathered landscape preferences in a comparable way between 2016 and 2019 in five different study areas in the Central Alps. Both studies used a stratified sampling approach based on the demographic distribution in the study areas, i.e., considering key socio-demographic characteristics such as gender, age, origin (residents, tourists), place of living (village, city), and native language (German, Italian). The two surveys, comprising 967 respondents [25] and 384 respondents [54], can therefore be considered to be representative for the general public of the Central Alps. In both surveys, differences between socio-demographic groups in preference scores were generally very small, although some significant differences occurred, mostly between people of different language, age, and origin [25]. After aligning the preference scores among the three surveys using the linear stretch method [59], we calculated the degree of correlation applying linear regression analysis. Since we obtained a high correlation (R² = 0.842; p<0.001; S2 Fig), we can assume that our sample also represents well enough general preferences.

Collection of variables

To explain variations in preferences, we prepared a set of predictor variables for each picture (Tables 1 and S2–S4), using three different methodological approaches [53] as described in the following.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Variables derived for all pictures using three different analysis approaches [53].

https://doi.org/10.1371/journal.pone.0288424.t001

Cloud types and spatial arrangement

In addition to the predictor variables, we determined the cloud type of each original photograph. Furthermore, the photo composition in terms of spatial arrangement, e.g., position of horizon, can influence people’s preferences [55, 68]. Based on the Rule of Thirds, as a simplification of the golden section or golden ratio (Phi, ϕ), we divided the picture into thirds both horizontally and vertically [69, 70]. This grid with 3x3 cells (S6 Fig) was overlaid over the hotspots to determine a spatial shift between pictures for which the preference scores significantly differed between the original picture with clouds and the manipulated picture without clouds.

Statistical analyses

To assess differences in the preference scores between the original landscape photograph with clouds (A) and the manipulated picture without clouds (B), we performed statistical comparisons using a bootstrap paired samples t-tests with 10,000 iterations. Significant differences were set as p < 0.05. We calculated mean values of the variables sky and clouds across picture pairs with decreased (A>B), not changed (A = B), and increased (A<B) preference score. Statistically significant differences were tested using the least significant difference (LSD) post hoc test.

To explain variations in landscape preferences between the picture pairs, we applied backward stepwise linear regression. The backward stepwise approach starts from a full model with all predictors and successively reduces them to obtain a model that best explains the data, while solving problems of multicollinearity and overfitting [71]. We used the difference in preference score (B-A) between the manipulated picture without clouds (B) and the original landscape photograph with clouds (A) as dependent variable and all variables from the eye-tracking simulation, photo content analysis, and GIS-based analysis as independent predictor variables (S6 Table). We removed collinear variables to avoid overfitting and to support the interpretability of the model by first screening for multicollinearity among the predictor variables. We calculated the variance inflation factor (VIF) values for each variable and removed variables with a VIF <10. Then, we applied a backward stepwise linear regression routine to further diminish the number of variables [72] and iteratively excluded variables with low predicting performance until no further improvement was possible (Table 1). To assess the level of variability explained by the model, we used the difference between R² and the adjusted R², as the adjusted R² is less influenced by overfitting than R² [73]. Through analysis of variance (ANOVA) or t-tests, validity, quality, and significance of the coefficients were evaluated. The reliability of the regression was determined using the Shapiro-Wilk test (residual normality), the Durbin-Watson test (autocorrelation of the variables), the Breusch-Pagan test (homoscedasticity) and the VIF statistics (multicollinearity effect). All preconditions for the reliability of the regression using were fulfilled (Shapiro-Wilk test for residual normality: p = 0.503; Durbin-Watson test for autocorrelation of the variables: 2.143), Breusch-Pagan test for homoscedasticity: p = 0.326) and the VIF statistics (multicollinearity effect, see Table 2). All statistical analyses were performed in SPSS Statistics (IBM SPSS 27, NY, USA).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Result of the multiple linear regression, using the difference in preference score (B-A) between the manipulated picture without clouds (B) and the original landscape photograph with clouds (A) as dependent variable.

Positive differences indicate that the photo without clouds is preferred over the photo with clouds, negative values indicate the opposite result. Accordingly, a negative regression coefficient B indicates that this variable has a positive effect on the preference score of photos with clouds. Only predictors with variance inflation factor (VIF) <10 during collinearity diagnostics are included.

https://doi.org/10.1371/journal.pone.0288424.t002

Perception of sky and clouds

In this study, the hypothesis that clouds in landscape photographs can influence the stated preferences due to color contrasts between clouds and the blue sky was partially confirmed. Significant differences in preferences were dependent on the percentage of cloud cover and proportion of the sky. Our findings indicate that a high proportion of clouds leads to lower preference scores compared to a sky without clouds. In these pictures, the percentage of clouds in the sky was on average 44% and the sky sometimes appeared darker, for example, due to nimbostratus clouds. This aligns with studies showing that inclement weather conditions or ephemeral phenomena such as thunderstorms lead to lower landscape preferences [41, 43]. In contrast, our results also indicate that single, smaller clouds (e.g., cumulus clouds) positively influenced preference scores. The visibility of the fractal outline of clouds [74] in these cases, helped by its clear contrast to the sky, might be part of the explanation for this. Such a finding fits well with our current understanding of the preferential perception of geometrical structures based on a hierarchy of scales, as in the nested symmetries of a fern leaf, for example, see Taylor et al. [75] as a review. Functional magnetic resonance imaging (fMRI) work has revealed the potential neural correlates of this phenomenon, as the processing of recursive forms has been shown to recruit different resources to the processing of non-recursive forms. Remarkably, it has been shown to recruit the default mode network, a functional brain network known to be involved in the processing of internal information [76, 77], thus implying a less computation-intensive processing of such shapes. In the pictures with higher preference, the eye-tracking hotspots were also often located on such small but high-contrast clouds or smaller cloud groups. Similarly, the study by Smalley and White [43] demonstrates that the appearance of the sky during sunrise or sunset increases perceived beauty of landscapes.

Regarding the composition of the pictures, painting and photography indicate clear rules, which can be related to our findings. The golden section or golden ratio (Phi, ϕ), referring to the positioning of the main subject in the picture or photograph, is considered the ideal ratio value determining aesthetic preferences, although controversially discussed [55, 78]. It is based on the fact that the human brain perceives images as harmonious if this rule is respected. As a simplification of the golden ratio, the Rule of Thirds is also central in design, films, and paintings, as well as in photography [69, 70]. Visual preferences have been found for photographs that depict about one third sky and two thirds foreground, if the foreground is interesting and striking [55]. In contrast, if the foreground is inconspicuous and/or there is a beautiful sky with clouds, then the rule of thirds is reversed, i.e., about one third foreground and two thirds sky [55]. Regarding our results, this means that landscape preferences may be higher if the proportion of sky reaches up to one third of the picture compared to another photograph that has considerably less proportion of sky. The vertical alignment was less important in our study, as the landscape did not change between the picture pairs. Nevertheless, the hotspots in the more attractive pair were located more towards the sides due to the lack or presence of clouds.

Implications for landscape research

Our findings also provide insights for landscape research. Many landscape preference studies using photographs as stimuli included pictures with blue or clear sky [24, 25, 30, 79], which is still recommendable in the light of our findings. There are also no constraints to include pictures with a proportion of sky around 22% and a cloud cover of about 39%. Nevertheless, for specific applications aiming at supporting real-world decisions, for example, evaluating aesthetic impacts of wind power farms, different weather conditions are often included [39, 80]. For this purpose, eye-tracking simulation may be a useful tool to detect changes in contrasts and to test different appearance of the infrastructures to reduce their visibility and thus, their impact on landscape aesthetic values. More generally, eye-tracking simulation may support visual impact assessment in an efficient way. For example, it could improve saliency mapping, which has been recently applied to evaluate visual impacts for wind and solar infrastructure in vineyard landscapes [81]. However, while there is a high correlation between eye-tracking data and saliency maps, it still needs be tested whether eye-tracking simulation provides similar results.

With regard to the composition of the picture (i.e., percentage of sky), our findings corroborate those of Svobodova et al. [55]. This aspect has been considered less often in landscape research, although other studies describe the influence of the characteristics of the horizon line on people’s perceptions, e.g., referring to the presence of mountains [29], in urban settings [82], and in relation with fractal characteristics [83, 84]. To reduce the influence of the sky on landscape preferences, our findings suggest that studies need to pay attention to the percentage of sky. Avoiding variations in percentage of sky will therefore produce a better comparability across different landscapes. For photographs taken in landscapes with complex horizon lines (e.g., mountains, differing vegetation height, and tall buildings) this may be more difficult but can be estimated, as done in this study. Using manipulated or computer-generated pictures, the percentage of sky can and should be rigorously controlled.

Limitations and future research

In general, our findings contribute to a deeper understanding of the influence of clouds in photographs used for eliciting landscape preferences in mountain regions. However, these results may not be transferrable to other landscapes. For example, the importance of sky and clouds may be differently evaluated in urban environments [43, 83, 85], in flat or hilly landscapes [86, 87]. Similarly, our results suggest that the influence of clouds was less relevant in pictures with diverse landscape patterns. It is therefore important to include different types of landscapes in future research. Moreover, the manipulations of the sky in some pictures did not completely reproduce natural conditions, as the sky sometimes had little texture and the lighting of the foreground landscape did not always match the lighting provided by a blue sky. It should therefore paid attention in future studies that the manipulations are not noticeable by the participants to avoid a potential negative influence on the preference scores of the manipulated pictures.

While we can assume that our results represented well enough general preferences, as suggested by the high correlation of our results with those of two former representative surveys [25, 54], future studies should use a stratified sampling approach, requiring also a higher number of respondents, to assess potential differences across socio-cultural groups. Regarding the composition of the landscape, differences in preferences were found, for example, between lay people and experts [17, 26, 27], and between residents and tourists [25, 26]. Preferences may also depend on the educational level [27], environmental values [30], or the connection of respondents with the landscape [88]. However, it still needs to be verified whether such differences also occur in terms of clouds.

While the above indicated recommendations apply for research focusing on general differences in perceptions across different landscapes, a less studied research field opens up in examining temporary changes, such as seasonal shifts [22, 89, 90] and diurnal changes [43, 85]. For example, different weather and light conditions can moderate preferences for a specific landscape [43]. The influence of such temporary states of landscapes on cultural ecosystem services is still less explored, despite studies addressing specific ‘events,’ such as the cherry blossoms [91], wildflower blooms [92], or the occurrence of rainbows [43, 93]. It may therefore not enough to use ‘blue-sky’ pictures to evaluate the overall aesthetic value of a location but to account for typical weather conditions and seasonal changes. A deeper understanding of the relationships between landscapes and human well-being may also allow decision-making and planning to better scope with global change pressures and integrate cultural ecosystem services into management plans and policies [15].