Correlating fluency theory-based visual aesthetic liking of landscape with landscape types and features

ABSTRACT

People inherently assess landscapes by creating spontaneous aesthetic liking judgments based on the surrounding stimuli. To understand these judgements objectively, use may be made of the fluency theory of aesthetic pleasure (the psychological processes through which people experience beauty). This study aims to predict people’s visual aesthetic preferences based on fluency theory and to correlate these preferences with landscape types and features. An ordinary least squares (OLS) regression model was developed to predict visual aesthetic liking, using image statistics as explanatory variables. We determined types of landscape using Landscape Character Assessment (LCA) and applied viewshed analyses distinguishing between near, medium, and far zones. We identified landscape features by content analysis making use of machine learning-based image recognition supplied by Google Cloud Vision API. The results show that vegetation and geological forms were the most significant features for people’s visual aesthetic liking, followed by waterscapes and built structures/human settlements. The viewshed analyses indicated that ‘medium-altitude, low-gradient artificial areas’ were visible in photographs with high aesthetic visual liking in all zones (i.e., at all distances). When the photographs showing this type of landscape are examined, the artificial areas in the photographs turn out to consist mostly of historical buildings or remains. This finding suggests that historical sites are not just important for their cultural value, but for their visual aesthetic value as well.

KEYWORDS:

• Fluency theory
• landscape character assessment
• machine learning
• visual aesthetic preference
• viewshed analysis
• Turkey

1. Introduction

1.1. Aesthetic preferences in landscape

Analyses of the perception of landscape (i.e. of awareness of the elements of the environment acquired through the physical senses) have been conducted with a view to determining the natural, cultural, and visual values of landscapes (Zube, Sell, and Taylor 1982; Skřivanová and Kalivoda 2010; Tveit, Ode Sang, and Hagerhall 2018; Yao et al. 2019; Khachatryan et al. 2020; Qi et al. 2020). Determining these values can play a key role in efforts to shape future landscapes, and is therefore linked to landscape management practices (Tribot, Deter, and Mouquet 2018). As a cultural ecosystem service, the aesthetic value of landscapes provides various health and social benefits (van Berkel and Verburg 2014; Guzmán 2017; Cheng et al. 2019). These considerations have also served to increase awareness about the preservation of natural settings (Assandri et al. 2018; Kerebel et al. 2019).

Landscape aesthetics can be defined as the enjoyment and appreciation experienced as a result of the observation of any type of landscape (Swaffield and McWilliam 2014; Tribot, Deter, and Mouquet 2018). There are two contrasting approaches to the aesthetics of landscape: the subjective and objective paradigms (Keshtkaran, Habibi, and Sharif 2017; Karasov et al. 2020). The former is a human-focused theory in which beauty depends on the subjective visual perception of the beholder; the latter focuses on the intrinsic qualities of the object, i.e. the attributes of the landscape (Dearden 1987; Lothian 1999; Kerebel et al. 2019). The objective paradigm has been influenced by environmental management practices (Daniel 2001). Some researchers suggest that a combination of the two approaches is crucial for a more comprehensive understanding (de la Fuente de Val, Atauri, and de Lucio 2006; Frank et al. 2013; Atik et al. 2017).

Hong, Lee, and Seung Lee (2016) defined the term “aesthetic preferences” as the degree to which people enjoy a particular visual stimulus, or how they rate its beauty. Aesthetic preference depends on a number of factors, such as the type of landscape, the memorized images of individuals, and cultural differences (Kalivoda et al. 2014; Wang, Zhao, and Liu 2016). In order to sharpen and improve perception-based theories of aesthetic preference, it seems relevant to explore people’s understanding of attributes of the landscape (Wang, Zhao, and Liu 2016; Sugimoto 2018). Most researchers have concentrated on subjective techniques when assessing the aesthetic value of landscape. Fewer researchers concerned with aesthetic preference analysis have employed objective techniques (Ode, Tveit, and Fry 2008; Saeidi et al. 2017).

Previous studies have shown that visual aesthetic preferences regarding landscape are affected by three factors: the attributes of the landscape and their spatial distribution, the organization of the landscape elements (spatial pattern of discernible objects), and variance across the landscape (Ode, Hagerhall, and Sang 2010; Ode and Miller 2011). Types of landscape can be used to represent these factors. Attributes of the landscape can be analyzed using Landscape Character Assessment (LCA) techniques, and types of landscape can then be associated with visual aesthetic preference (Chesnokova et al. 2019). It has been reported that aesthetic preferences vary between types of landscape, and that types of landscape therefore have a significant influence on aesthetic judgment (Sevenant and Antrop 2009; Kalivoda et al. 2014; Wang, Zhao, and Liu 2016; Martín et al. 2016). However, no consensus has been reached on the precise relationship between types of landscape and aesthetic preferences. Although LCA is a robust method for determining landscape character, different conclusions may be reached when associating landscape character with visual aesthetic preference, because variations in the spatial arrangement of types of landscape can be site-specific (Wang, Zhao, and Liu 2016).

1.2. Image fluency theory and visual aesthetic preferences

The concept of image fluency, or processing fluency, may help to quantify visual aesthetic preferences. In this context, fluency can be defined as the facility with which people process images. Image fluency theory proposes that perceptions of visual symmetry, self-similarity, simplicity, and contrast can enhance aesthetic associations (Mayer and Landwehr 2018). In image fluency theory, visual simplicity refers to amounts of information that can be processed using relatively little capacity (Barton and Barton 1987; van der Helm 2000). Visual symmetry is the similarity of the parts of an image when transformed through reflection or translation, especially through the Euclidean plane (Makin, Pecchinenda, and Bertamini 2012; Bertamini and Makin 2014). Visual contrast defines the variation in luminance between the features of a picture (Sheikh and Bovik 2006; Ponomarenko et al. 2011). More specifically, contrast implies the difference between the different components of an image that portrays an object which is unique or stands out prominently from its background (Whittle 1994; Mayer and Landwehr 2018). Self-similarity refers to repetitive visual patterns, such as lack of variation in scale, that are highlighted by zooming out (Amirshahi et al. 2012; Boeing 2016). All these concepts refer to low-level stimuli that help to unveil predictive aesthetic preferences (Graf, Mayer, and Landwehr 2018; Mayer and Landwehr 2018). According to processing fluency theory, the aesthetic liking of a landscape is influenced by how easily the landscape can be perceived, since higher fluency and ease of perception are positively experienced in a subjective effect (Mayer and Landwehr 2018).

Photographs obtained from social media platforms can be used as material (Hur, Lim, and Lyu 2020) for the application of image fluency theory. Platforms such as Flickr, Twitter, Instagram, and Facebook have huge databases of photographs that are publicly associated with users of different backgrounds (Hu et al. 2015; Heikinheimo et al. 2017). Geotagged photographs are known to be useful for identifying geographical mapping patterns as well as for exploring perceptions of different places by estimating the flow of visitors or the number of times people have visited them (Figueroa-Alfaro and Tang 2017; Oteros-Rozas et al. 2018; Wartmann et al. 2019; Vivek, Griess, and Keena 2020; Shao et al. 2021). Recently, there has also been an increase in the use of geotagged photographs for evaluating aesthetic perceptions of landscapes (Philipps, Zerr, and Herder 2017; Tieskens et al. 2018; Do and Yoon Kim 2020; Gosal and Ziv 2020).

Meanwhile, image recognition software has improved in recent years, and has been found particularly useful for analyzing big data (Zerdoumi et al. 2018; Dold and Groopman 2017; Lock and Pettit 2020). Accessible websites such as Google’s Cloud Vision and Microsoft’s Computer Vision provide Application Programming Interfaces (APIs) for image recognition (Google Cloud Vision 2017; Microsoft Computer Vision 2017). APIs make it possible for users to classify images according to their content. The Google Cloud Vision API uses a machine learning algorithm (an application of artificial intelligence that is able to learn and improve from experience without being explicitly programmed) to assign keywords to images. APIs are currently being used for research purposes, including for understanding the landscape elements and spatial configurations that inform people’s choices (Jorgensen 2011; Richards and Tunçer 2018; Callau et al. 2019).

Artificial intelligence has recently been put to use in landscape planning for making comprehensive evaluations of natural environments (Jahani, Kalantary, and Alitavoli 2021; Jahani et al. 2021; Qi et al. 2020; Yan, Schultz, and Zipf 2019). Most studies have focused mainly on visual aspects of the landscape and found that artificial intelligence technology has great potential for modeling and predicting natural phenomena (Jahani et al. 2021; Jahani 2019). Since this technology can recognize patterns in various types of images with a high degree of accuracy, the predictions made are reliable and practical (Zhang et al. 2018; Jahani, Kalantary, and Alitavoli 2021).

Landscape character types, as defines by Atik et al. (2015), are distinct types of landscape that are relatively homogeneous in character and share broadly similar combinations of landscape properties. The research reported in this paper – which was conducted in the environs of Lake Bafa, a site of both natural and historical importance – aimed to investigate the relationship between landscape types and visual aesthetic preferences using social media photographs, machine learning, and geographic information systems (GIS). We used an ordinary least squares (OLS) regression model to predict visual aesthetic liking of landscape, determining types of landscape using LCA. In addition, we identified landscape features by means of content analysis, classifying the landscape features in the photographs using machine learning-based image recognition supplied by the Google Cloud Vision API. In this way, we combined diverse analytical methods into a methodological approach for quantifying visual aesthetic preferences related to landscapes with different characteristics.

In particular, we addressed the following research questions:

1. How can people’s visual aesthetic liking of landscape be predicted?

2. How does visual aesthetic liking of landscape vary by type of landscape?

3. Which landscape features dominate in landscapes of high visual aesthetic value?

Understanding the visual aesthetic value which people attach to landscapes is of considerable importance, especially in the case of landscapes that have become endangered. By correlating visual aesthetic liking of landscape with the type of landscape and landscape features, inputs can be generated for designing future landscapes and determining potential conservation strategies (Krause 2001; Ha and Yang 2019). We go on to discuss how the determination and analysis of visual aesthetic liking of landscapes with different typologies (natural areas, historical sites, etc.) might be incorporated into the practice of landscape planning and development.

2. Materials and methods

2.1. Study area and data

The study area (Figure 1) extends over 55,366.33 ha between the district of Milas in the province of Muğla and the district of Söke in the province of Aydın, in southwest Turkey. It centers on Lake Bafa and the Lake Bafa Natural Park.

Figure 1. The geographical location of the study area and the historical sites.

In pre-classical times, Lake Bafa formed part of a gulf of the Aegean Sea. The lake was separated from the Aegean by alluvial sediments from the River Meander and became an alluvial dam lake (Müllenhoff et al. 2004). Lake Bafa and its environs (areas shaded in Figure 1) were officially declared a nature reserve, named Lake Bafa Natural Park, in 1994 (DKMP 2021). The park provides a breeding and wintering environment for a large number of bird species (Hetemoğlu 2019). It is therefore considered one of the most important wetlands in Turkey and has been declared an important bird area and potential Ramsar (Convention on Wetlands of International Importance) site by BirdLife International (2001). Besides birdwatching, the park has great potential for activities like angling, camping and walking (Deniz et al. 2011).

The study area contains several layers of vegetation, predominantly olive, pine, and Eastern Mediterranean maquis. Sixteen endemic species have been identified (DKMP 2021). The native distribution of the stone pine (Pinus pinea L.) is significant for the ecology and economy of the area (Müllenhoff et al. 2004).

The area boasts a variety of landscapes that feature ecological, architectural, cultural, geological, recreational, and historical assets (Herda et al. 2019). In and around the village of Kapıkırı, at the foot of the Beşparmak Mountains – known as Latmos in ancient times – are to be found the ancient cities of Herakleia and Latmos, and the Byzantine Yediler Monastery. There are five islands in the lake, shaped by rock outcrops: Kahve Asar Island, the Ikiz Islands (Küçük Ikiz Island and Büyük Ikiz Island), Menet Island, and Kapıkırı Island. Byzantine archeological remains are to be found both on the islands and in Kapıkırı Village (Figure 1). These sites exemplify the general characteristics of Byzantine monastic complexes, such as the need for protection and access to water sources, as well as the architectural and spatial features and typical building techniques and materials of the time (Hetemoğlu 2019). The fortress of Ikiz Island on the edge of Lake Bafa was built to protect the monastery (Thonemann 2011). Latmos is also a unique cultural site on account of the ancient rock paintings that have been found there, dating back to the 6th-5th millennia BCE (Peschlow-Bindokat and Gerber 2012).

Intensive agriculture and grazing, human settlement and exploitation for timber have affected the landscape over the past decade or so. According to one case study, endemic biota, archeological sites and geomorphological features are under threat from mining activities (Gül, Zorlu, and Gül 2019).

The materials of this study consist of geotagged photographs collected from Flickr and Google Earth, a roadmap downloaded from OpenStreetMap, the Digital Elevation Model (DEM) accessed via NASA Earthdata (USGS 2020), a land use map obtained from the Geographical Information Systems Office of Muğla Municipality, and gradient and elevation maps produced from the DEM (Table 1).

Table 1. Materials used in the study.

Materials	Source	Format
Geotagged photographs	Flickr and Google Earth	Point shapefile and.jgp files
Road data	OpenStreetMap	Polyline data
Digital Elevation Model (DEM)	NASA Earthdata	Raster data
Land uses map	Muğla Municipality	Polygon shapefile
Slope map	Generated from DEM	Raster data
Elevation map	Generated from DEM	Raster data

2.2. Work flow

The methodology of the study comprises seven steps as follows: 1) data collection and pre-processing of the data; 2) accessibility analysis; 3) computation of fluency metrics and the development of an ordinary least squares (OLS) regression model to predict visual aesthetic liking and categorize visual aesthetic liking values (low, medium, high); 4) viewshed analyses based on different distance zones; 5) determination of types of landscape by Landscape Character Assessment (LCA); 6) association of types of landscape with visual aesthetic liking levels under three viewshed computations, and 7) identification of landscape features from geotagged photographs that have high levels of visual aesthetic liking using the Google Cloud Vision API (Figure 2).

Figure 2. Methodology of the study.

In the first step, the photographs taken in the study area were analyzed, together with their dates. Since the photographs showing the landscape were shared on social media from 2004 onwards, they were grouped in such a way as to cover the years between 2004 and 2020. A total of 6,091 geotagged photographs were found to have been shared on Flickr and Google Earth. The geotagged photographs collected from Flickr were saved using the photosearcher R library (Fox et al. 2020), while the photographs collected from Google Earth (previously Panoramio) were saved manually. The Flickr API enables users to retrieve georeferenced photographs falling within a bounding box, and the photosearcher R library accesses the Flickr API. This procedure provides the researcher with the photographs themselves and a great number of details such as their location (latitude and longitude), the dates when they were taken and uploaded, hashtags, user identities and view counts. The time of uploading and the view count (the number of times other people have viewed the photograph) were extracted and recorded for each of the geotagged photographs. In the case of the photographs downloaded from Google Earth, the geolocations were saved with the help of the Geoalgorithms tool of the QGIS 2.8.8 software. Copyright was observed when collecting both sets of data and the downloading of photographs was avoided. Images that do not depict the landscape (personal photographs such as selfies) were excluded.

In the second step, an accessibility analysis was carried out using road data clipped to match the geographical extent of the study area. The goal of this analysis was to correlate the distribution of the geotagged photographs with the accessibility of the locations. The geolocations of the photographs were analyzed on the basis of their proximity to roads, using intervals of 0-200 m, 200-400 m, 400-600 m, 600-800 m and 800–1,000 m. The Euclidean Distance tool in ArcMap 10.7 was used for the accessibility analysis.

In the third step, an executable script was written in R v. 4.0.3 (R Core Team 2020) and fluency metrics were computed using the image fluency R library as described in the following paragraphs (See Mayer and Landwehr (2018) for details. See also Figure 3).

Figure 3. Illustration of the image fluency metrics.

For the simplicity variable, a value of between one (very simple) and zero (highly complex) is determined based on the compressibility of the image (Simple images can be compressed into much smaller files). Meanwhile, a contrast value is obtained by calculating the standard deviation of the intensity of the pixels, normalized if required. If the deviation is high, this means that contrasts are great and results in a high contrast value.

As for the self-similarity variable, this is determined from the slope of the power spectrum log-log plot in the OLS. A steeper slope points to more low-frequency components and hence to lower self-similarity (See Simoncelli and Olshausen 2001; Redies, Hasenstein, and Denzler 2007), which is indicated by a value approaching zero. With color images, separate calculations are made for each color channel and the weighted average is obtained (cf. Mayer and Landwehr 2018).

Finally, the symmetry variable requires the calculation of symmetry variables for a range of horizontal and vertical mirror axes. The highest of the variables obtained in this way is taken to represent the value for the perceived mirror axis, which may not be in the center of the image.

Here, aesthetic preference is represented by the number of views. The view count is a response variable used to develop an OLS regression model in the R statistical software. View count data is only available from Flickr. Consequently, only the photos originated from Flickr (284 geotagged photographs) were used in this model. A natural log transformation is used because of the skewed nature of the view count.

Mayer and Landwehr (2018) offer a model that includes the four properties of the images as predictors. In this model, rank order is not used for purposes of prediction. The model equation is as follows:

(1) $\begin{array}{c}\log {\left(\mathrm{VIEWS}\right)}_i=b_1\mathit{ }\mathrm{x TIM}{\mathrm{E}}_{\mathrm{i}}\mathit{ }+\mathit{ }{b}_2\mathit{ }\mathrm{x SIMPLICIT}{\mathrm{Y}}_i\mathit{ }\mathrm{\backslash break}+\mathit{ }{b}_3\mathit{ }\mathrm{x SYMMETR}{\mathrm{Y}}_i\mathit{ }+\\ b_4\mathit{ }\mathrm{x CONTRAS}{\mathrm{T}}_i\mathrm{\backslash break}+b_5\mathit{ }\mathrm{x SELF}-\mathrm{SIMILARITY}\mathit{ }+\mathit{ }{\epsilon }_i\end{array}$(1)

In the formula above, VIEWCOUNT represents the number of views of the geotagged photographs, TIME the year of publication (direct value of the year of uploading), SIMPLICITY the visual simplicity value, SYMMETRY the visual symmetry value, CONTRAST the visual contrast value, SELFSIMILARITY the visual self-similarity value and ε the error term. Fluency values [log(VIEWS)] were assumed to be determinative for visual aesthetic liking. Visual aesthetic liking was then categorized into three groups (high, medium or low) using quantiles of the values in the OLS model results.

In the fourth step of the work flow, GIS-based viewshed analyses were conducted using the geotagged photographs. The purpose of the viewshed analyses was to determine which type of landscape is visible in the photographs taken in places where people are present. Raster surface data were used in the viewshed analyses and the geolocations of the photographs used in the study were used as observer data.

In the fifth step, LCA (Swanwick 2002) was conducted using three environmental layers: land use, gradient, and elevation (Atik et al. 2015). These variables were selected as they are determinative when visually describing landscapes. For the purpose of the LCA, the current land use map, which also shows the vegetation, was obtained from Muğla Municipality, while elevation and gradient maps were produced from 27 m resolution DEM using the ASTER GGDEM – Global Digital Elevation Model. The elevation map used the classifications low-altitude (0-250 m), medium-altitude (250-360 m), elevated (360-600 m), and high-altitude (600–2,000 m) (Haslam and Anne Wolseley 1981). The gradient map used the categories low (0–15%), moderate (15–30%), medium (30–45%) and high (45–65%). The three maps were overlaid in ArcMap 10.7 with overlay analysis and the types of landscape were coded accordingly (Atik et al. 2017). Figure 4 gives an example of the coding.

Figure 4. Sample landscape code.

In the sixth step, the viewshed analyses and the map of the types of landscape were overlaid. Considering that the visual aesthetic liking of the photographs may vary with the distance between the type of landscape shown and the geolocation of the viewer, the analysis was conducted separately for “near-zone” (within 50 m), “medium-zone” (200 m) and “far-zone” (1,000 m) landscapes. These parameters were determined in the light of the variety of landscape types, the scale of the study area, the average distances between the type of landscape shown and the viewer, and the distribution of the geotagged photographs used in previous studies (Tenerelli, Püffel, and Luque 2017). For each zone, the codes representing the types of landscape visible were listed by level of liking. Unique code(s) were assumed to be determinative when comparing levels of visual aesthetic liking. The text data consisting of codes were translated into multiple vectors in the R environment and common codes were found using the “reduce” command. The “setdiff” command, which performs asymmetric difference analysis on vector data, was then run to find unique landscape codes.

In the final step of the work flow, content analysis was performed to understand which features of the landscape attracted people visually. The landscape features in the geotagged photographs with high levels of visual aesthetic liking were identified using the Google Cloud Vision API. The Vision API allows developers to detect features through image labeling, face and landmark detection, optical character recognition, and tagging of explicit content (Benliay and Altuntaş 2019). Based on a machine-learning algorithm, image labeling gives broad information about the entities in an image (Ramos et al. 2020). Each photograph depicted one or more landscape feature, as clearly demonstrated by the labels acquired from the Google Cloud Vision API. To classify these labels, the approach proposed by Richards and Friess (2015) for the conduct of content analysis on a fine spatial scale was adopted. Eight categories were identified (vegetation, waterscape, geological feature/landform, historical landmark/architecture, animal, recreational facility/transportation, built structure/human settlement, and trail/path), and each photograph was classified into one or more than one category according to its labels (Table 2). Boolean logic was utilized to classify the labels and a binary heat map was created to visualize the landscape features.

Table 2. Classification of landscape features.

Landscape feature class	Description	Sample labels
Vegetation	Photographs showing trees, flowers, shrubs etc.	Tree, plant, scrublands, flower
Waterscape	Photographs illustrating lakes, rivers etc.	Lake, water, water resources
Geological feature/landform	Photographs depicting rocks, mountains etc.	Mountain, hill, highland, rock, bedrock, mountainous landforms
Historical landmark/architecture	Photographs depicting historical infrastructure (e.g. historical buildings, ruins)	Ruins, historic site, ancient history
Animal	Photographs focusing on animals such as birds, cows, donkeys etc.	Bird, cattle, donkey
Recreational facility/transportation	Photographs showing boats on Lake Bafa	Boat, vehicle
Built structure/human settlement	Photographs depicting buildings	House, chapel
Trail/path	Photographs illustrating roads, paths, trails	Road, trail

3. Results

A total of 651 photographs depicting landscape, taken by 106 different users, were used in the study. As can be seen in Figure 5, the geotagged photographs were concentrated to the east of Lake Bafa.

Figure 5. The spatial distribution of the geotagged photographs 2004–2020, types of road, and Euclidean distances.

We observed a significant concentration of photographs in four historical places: Kapıkırı Island, the ancient cities of Herakleia and Latmos, and the Yediler Monastery. A few photographs were also taken of the historical areas of the islets in Lake Bafa. The reason why fewer photographs were taken in other historical areas may well be related to their lower accessibility, as an examination of the accessibility map suggests. Interestingly, people appear to make more use of the secondary roads than the main roads in the study area. When all types of road are considered, 347 of the 651 photographs are seen to be located within 0-200 m of a road, 123 within 200-400 m, 57 within 400-600 m, 41 within 600-800 m, and 18 within 800–1,000 m. The remaining 65 photographs were taken in areas farther than 1,000 m from the nearest road.

When the four fluency metrics (visual simplicity, visual symmetry, visual self-similarity, and visual contrast) were computed for each geotagged photograph, and an ordinary least square (OLS) regression model was developed, the parameter estimates and associated p-values for all variables as per the results of the OLS regression model were as shown in Table 3. The visually preferred photographs and their locations are given in Appendix A.

Table 3. Parameter estimates of the model (Std. Error: Standard error. Pr: probability. VIF: variance inflation factor. Time is the year the photograph was uploaded on Flickr.com. n = 284, + p < .10^.* p < .05^.** p < .01*** p < .001.).

Term	Estimates	Std. Error	t value	Pr(>|t|)	VIF
(Intercept)	3.13351	0.20880	15.007	<2e-16 ***	.
Contrast	−0.21744	0.10818	−2.010	0.0454 *	1.253171
Self-similarity	−0.12665	0.10542	−1.201	0.2306	1.190055
Simplicity	−0.58889	0.12692	−4.640	5.37e-06 ***	1.724806
Symmetry	0.15860	0.12920	1.228	0.2207	1.787404
Time	0.29424	0.03273	8.991	<2e-16 ***	1.024363

The study indicates that time (t = 8.99, p < 0001), simplicity (t = −4.64, p ≤ 0.0001) and contrast (t = −2.01, p = 0.04) are each statistically significant predictors of the logged geotagged view counts. The slope estimates for contrast and simplicity are −0.217 and −0.589 respectively, so that for each unit increase in contrast, the response falls by exp(−0.217) = 0.80 units (i.e. by −0.217 units on the natural logarithmic scale), while for each unit increase in simplicity, it falls by 0.55 units. There is no indication of multicollinearity between the predictors, since all the variance inflation factor (VIF) scores – which measure the degree of multicollinearity in a set of multiple regression variables – are smaller than 2. A plot of residuals versus predicted values is shown in Appendix B. According to the plot, the residuals are randomly scattered around zero with constant variance and no apparent outliers. Thus, the data fit the model adequately. The values varied between 589.41 and 594.30 and were categorized using quantiles as low, medium, or high. On this basis, there were 165 geotagged photographs with a low level of fluency (values of 589.41–591.75), 323 with a medium level of fluency (values of 591.76–593.83) and 163 with a high level of fluency (values of 593.84–594.30). Some example photographs with high, medium, and low evaluation results based on our model and computation are given in Appendix C.

The geographical distributions of the geotagged photographs with low, medium, and high levels of fluency across the study area differed from one another. Unlike geotagged photographs with low and high levels of fluency, geotagged photographs with medium levels of fluency were to be found not just in the east of Lake Bafa, but also in the north and south of the lake and east of the Beşparmak Mountains. As shown in the viewshed analyses in Figure 6, the locations visible in the photographs differ to some extent for each level of fluency. While all parts of Lake Bafa can be seen from the locations where the geotagged photographs with low and medium levels of fluency were taken, the north and north-western parts of the lake are not visible from the locations where photographs with high levels of fluency were taken. In all, 51.1% of the study area is visible for the geotagged photographs with low levels of fluency, 39.5% for the photographs with medium levels of fluency, and 40% for the photographs with high levels of fluency.

Figure 6. Viewshed analyses based on geotagged photographs with low, medium, and high levels of fluency.

Figure 7 shows the spatial distribution of the types of landscape in the study area. Among the 103 types of landscape identified, the three which cover the largest proportions of the study area are “low-altitude, low-gradient black pine forests (l_lo_bp)” (10.19%), “Lake Bafa” (14.48%) and “low-altitude, low-gradient agricultural areas (l_lo_ag)” (15.79%). These are followed by “low-altitude, moderate-gradient black pine forests (l_mo_bp)” (4.88%) and “high-altitude, moderate-gradient black pine forests (h_mo_bp)” (4.26%). “Low-altitude, low-gradient macchia (l_lo_ma)” “elevated, low-gradient black pine forests (e_lo_bp)”, “low-altitude, moderate-gradient agricultural areas (l_mo_ag)”, “elevated, moderate-gradient agricultural areas (e_mo_ag)”, “elevated, moderate-gradient black pine forests (e_mo_bp)” and “medium-altitude, low-gradient black pine forests (m_lo_bp)” each account for between 3.33% and 3.93% of the total study area. Eighty-three of the remaining types of landscape cover less than 1% of the area. Of the 651 photographs taken in the study area, 250 were taken of “Low-altitude, low-gradient macchia (l_lo_ma)”, 120 of “low-altitude, low-gradient agricultural areas (l_lo_ag)”, 63 of “low-altitude, low-gradient artificial areas (l_lo_ar)”, 53 of “low-altitude, moderate-gradient macchia (l_mo_ma)” and 50 of “Lake Bafa”.

Figure 7. Spatial distribution of the types of landscape (l: low-altitude, m: medium-altitude, e: elevated, h: high-altitude, lo: low-gradient, me: medium-gradient, mo: moderate-gradient, hi: high-gradient, ag: agricultural areas, ar: artificial areas, ba: bare areas, bp: black pine forests, du: dunes, ft: treeless forest areas, ma: macchia, sp: stone pine forests, sw: swamps).

Table 4 gives the disaggregation of the types of landscape visible in photographs with different levels of visual aesthetic liking by zone (near: 50 m, medium: 200 m, and far: 1,000 m). In the case of the near zone, only two types of landscape (by unique landscape code) fall within the visibility border of the photographs with high visual aesthetic liking. These are “medium-altitude, low-gradient artificial areas (m_lo_ar)” and “high-altitude, moderate-gradient stone pine forests (h_mo_sp)” – an artificial landscape and a forested landscape. A wide variety of types of landscape (by unique visibility code) are depicted in photographs with medium visual aesthetic liking – namely, “high-altitude, high-gradient bare areas (h_hi_ba)”, “elevated, low-gradient artificial areas (e_lo_ar)”, “elevated, low-gradient treeless forest areas (e_lo_tf)”, “medium-altitude, low-gradient macchia (m_lo_ma)”, “elevated, medium-gradient black pine forests (e_me_bp)”, “low-altitude, medium-gradient agricultural areas (lo_me_ag)”, “high-altitude, medium-gradient bare areas (h_me_ba)”, “high-altitude, moderate-gradient black pine forests (h_mo_bp)”, “elevated, moderate-gradient treeless forest areas (e_mo_tf)” and “low-altitude, moderate-gradient swamps (l_mo_sw)”. The types of landscape (by unique visibility code) which fall within the visibility border of photographs with low visual liking are “elevated, low-gradient bare areas (e_lo_ba)”, “low-altitude, moderate-gradient treeless forest areas (l_mo_tf)” and “high-altitude, moderate-gradient treeless forest areas (h_mo_tf)”. A common feature of these areas is a lack of vegetation.

Table 4. Types of landscape by zone and level of visual aesthetic liking based on visibility analyses and unique landscape codes.

	Landscape types
	Near Zone (50 m)	Medium Zone (200 m)	Far Zone (1000 m)
High visual aesthetic liking	Medium-altitude, low-gradient artificial areas (m_lo_ar) High-altitude, moderate-gradient stone pine forests (h_mo_sp)	Medium-altitude, low-gradient artificial areas (m_lo_ar) High-altitude, moderate-gradient agricultural areas (h_mo_ag) Medium-altitude, moderate-gradient artificial areas (m_mo_ar)	Medium-altitude, low-gradient artificial areas (m_lo_ar) Medium-altitude, low-gradient stone pine forests (m_lo_sp) High-altitude, medium-gradient artificial areas (h_me_ar) High-altitude, moderate-gradient artificial areas (h_mo_ar) Medium-altitude, moderate-gradient artificial areas (m_mo_ar)
Medium visual aesthetic liking	High-altitude, high-gradient bare areas (h_hi_ba) Elevated, low-gradient artificial areas (e_lo_ar) Elevated, low-gradient treeless forest treeless (e_lo_tf) Medium-altitude, low-gradient macchia (m_lo_ma) Elevated, medium-gradient black pine forests (e_me_bp) Low-altitude, medium-gradient agricultural areas (l_me_ag) High-altitude, medium-gradient bare areas (h_me_ba) High-altitude, moderate-gradient black pine forests (h_mo_bp) Elevated, moderate-gradient treeless forest areas (e_mo_tf) Low-altitude, moderate-gradient swamps (l_mo_sw)	High-altitude, high-gradient bare areas (h_hi_ba) High-altitude, high-gradient black pine forests (h_hi_bp) Elevated, low-gradient artificial areas (e_lo_ar) High-altitude, low-gradient artificial areas (h_lo_ar) Elevated, medium-gradient bare areas (e_me_ba) High-altitude, medium-gradient stone pine forests (h_me_sp) Elevated, moderate-gradient artificial areas (e_mo_ar) Elevated, moderate-gradient bare areas (e_mo_ba)	High-altitude, high-gradient agricultural areas (h_hi_ag) High-altitude, high-gradient stone pine forests (h_hi_sp) High-altitude, low-gradient artificial areas (h_lo_ar) Elevated, medium-gradient bare areas (e_me_ba) Elevated, moderate-gradient bare areas (e_mo_ba)
Low visual aesthetic liking	Elevated, low-gradient bare areas (e_lo_ba) Low-altitude, moderate-gradient treeless forest areas (l_mo_tf) High-altitude, moderate-gradient treeless forest areas (h_mo_tf)	Elevated, high-gradient macchia (e_hi_ma)	Low-altitude, medium-gradient treeless forest areas (l_me_tf)

Turning to the medium zone, the types of landscape (by unique landscape code) within the visibility range of the photographs with high visual aesthetic liking are “medium-altitude, low-gradient artificial areas (m_lo_ar)”, “high-altitude, moderate-gradient agricultural areas (h_mo_ag)” and “medium-altitude, moderate-gradient artificial areas (m_mo_ar)”. It is worth emphasizing that two of these are artificial areas. Once again, many types of landscape (by unique landscape code) appear within the visibility range of the photographs with medium visual aesthetic liking. These are “high-altitude, high-gradient bare areas (h_hi_ba)”, “high-altitude, high-gradient black pine forests (h_hi_bp)”, “elevated, low-gradient artificial areas (e_lo_ar)”, “high-altitude, low-gradient artificial areas (h_lo_ar)”, “elevated, medium-gradient bare areas (e_me_ba)”, “high-altitude, medium-gradient stone pine forests (h_me_ba)”, “elevated, moderate-gradient artificial areas (e_mo_ar)” and “elevated, moderate-gradient bare areas (e_mo_ba)”. In the medium zone, there is only one type of landscape (by unique landscape code) within the visibility range of the photographs with low visual aesthetic liking – namely, “elevated, high-gradient macchia (e_hi_ma)”.

Finally, the far zone analysis shows that several types of landscape (by unique landscape code) appear within the visibility range of the photographs with high visual aesthetic liking. These are “medium-altitude, low-gradient artificial areas (m_lo_ar)”, “medium-altitude, low-gradient stone pine forests (m_lo_sp)”, “high-altitude, medium-gradient artificial areas (h_me_ar)”, “high-altitude, moderate-gradient artificial areas (h_mo_ar)” and “medium-altitude, moderate-gradient artificial areas (m_mo_ar)”. Again, there is a preponderance of artificial areas. The types of landscape (by unique landscape code) within the visibility range of the photographs with medium visual aesthetic liking in this zone are “high-altitude, high-gradient agricultural areas (h_hi_ag)”, “high-altitude, high-gradient stone pine forests (h_hi_sp)”, “high-altitude, low-gradient artificial areas (h_lo_ar)”, “elevated, medium-gradient bare areas (e_me_ba)” and “elevated, moderate-gradient bare areas (e_mo_ba)”. The only type of landscape (by unique landscape code) within the visibility border of the photographs with low visual aesthetic liking is “low-altitude, medium-gradient treeless forest areas (l_me_tf)”.

When the landscape features in the 163 photographs found to have a high level of fluency are binary-categorized through the Google Cloud Vision API, vegetation and geological forms emerge as the dominant landscape features, followed by waterscapes and built structures/human settlements (Figure 8). The incidences of historical landmarks/architecture and trails/paths are similar to one another, and both of these occur more frequently than recreational facilities/transportation and animals.

Figure 8. Binary heat map showing landscape features of geotagged photographs with high levels of fluency.

4. Discussion

4.1. Limitations of the study

Contemporary technology makes it possible to use combinations of methods to correlate types of landscape with visual aesthetic value. The visual aesthetic liking of landscape can be assessed objectively based on spontaneous spatial-temporal information obtained from social media platforms. However, social media platforms, particularly the frequently-used Facebook and Instagram, have recently stopped sharing data because of privacy policies. Since 2019, it has not been possible to gather data from these platforms through any software package or interface. Moreover, even if Facebook and Instagram were to change their privacy policies and enable data sharing, it would not be possible to access the view counts for geotagged photographs. One of the limitations of this study, therefore, is the inaccessibility of data from widely-used social media platforms. For this reason, the study was limited to the data sets pertaining to geotagged photographs available on Flickr and Google Earth. Google Earth provides geotagged photographs, the time when the geotagged photographs were taken, and user information. However, it does not share the numbers of views the photographs have received. Consequently, for the OLS model, the study was only able to use the photographs obtained from Flickr.

Another limitation of this study is the lack of any assessment of the accuracy of the measure of visual aesthetic liking. More research is required to investigate the accuracy of fluency metrics and prevent bias. In order to test the reliability of the metrics, experts would need to compare their results with findings from traditional landscape aesthetics methodologies such as surveys and questionnaires. Besides, there is nothing to prevent the objective methods used here from being combined with subjective methods; previous studies have used the combination extensively.

The Google Vision Cloud API is helpful for determining objectively which feature each photo shows. In general, the accuracy of the machine learning algorithms of the Google Vision Cloud API is reliable. This study found that 80% of the labels were assigned correctly. However, when each label was checked carefully, some labels were missing. For instance, one geotagged photograph showed a view which included Lake Bafa but none of the attached labels referred to waterscape elements such as “lake” or “water”. For this reason, the missing labels were added manually to make the analysis 100% accurate.

4.2. Comparison with previous research

The fluency theory of Mayer and Landwehr (2018) has made it possible to study visual preferences objectively. In addition, the metric calculation package (image fluency) which they have made available free of charge has greatly benefitted researchers working on similar issues and made it possible to re-test their theory. There are similarities between the conclusions these researchers have arrived at using landscape images and the outputs of the present study. In both cases, a significant negative relationship was found between the time when the geotagged photographs were shared on Flickr (date of uploading) and visual simplicity. Previous studies have stated on several occasions that visual simplicity is an important criterion for determining visual aesthetics

Visual simplicity and its antonym, visual complexity, have been the topic of various debates in visual perception studies. Our findings could be interpreted as indicating that people like landscapes of higher visual complexity (i.e. lower visual simplicity) – such as landscapes (See Appendix A) in which geological features are surrounded by vegetation, or the lake is embraced by cultural elements such as historical. While some studies support our findings that humans’ visual aesthetic preferences favor complex landscapes, some others claim that landscapes with low complexity (i.e. higher visual simplicity), such as water bodies or the sky, are better liked (Ode, Hagerhall, and Sang 2010). Day (1967) asserts that people prefer medium-level complexity to either low- or high-level complexity. Noting that there are studies which suggest both that visual liking increases with the complexity of the landscape and that it decreases, Nadal et al. (2010) argue that these contradictory findings could reflect unresolved connections or manipulations. For this reason, it is not possible to measure the importance of various visual aesthetic evaluation criteria by a standard method. The statistical significance of various criteria may be debated, but the conclusions will not be entirely accurate. In the present study, for example, the time when the photograph was uploaded to Flickr, visual simplicity, and visual contrast were found to be significant in determining visual aesthetic liking. However, in another study conducted with a different data set, different fluency metrics, such as visual self-similarity or visual symmetry, might turn out to be influential. Thus the relative importance of different metrics is related to the information included in the data set used.

The fact that the geotagged photographs were not homogeneously distributed across the study area can be explained by the different levels of accessibility. The accessibility analysis, which showed that the number of photographs decreases with distance from the nearest road, supports this conclusion. According to the information obtained as a result of the LCA, the topography and the water body are influential here. Transport services are not available to the historical sites within Lake Bafa, and personal vehicles such as canoes or boats are the only means of reaching them. This explains why fewer photographs of these areas were shared. In the doctoral study she conducted in the same area, Hetemoğlu (2019) emphasizes that visitors to the area’s cultural heritage sites experience serious access problems.

As suggested by the accessibility analysis, people may simply have preferred the most physically accessible parts of the study area. At the same time, humans use their visual evaluation mechanisms a great deal when perceiving a landscape and assigning value to it, and they tend to prefer and visit landscapes that they find aesthetically pleasing. Looked at from this angle, there are two essential reasons why people might not favor a landscape. The first is that the area might not be accessible due to its topographical structure or gradient, and might therefore appeal only to people who are interested in particular recreational activities such as trekking or mountaineering. The second reason is that the landscape potential of the area might not be sufficiently known. People are unlikely to favor landscapes the potential of which is either unknown or known only to small groups of people such as researchers or land managers.

Our findings are compatible with the results obtained by Wang, Zhao, and Liu (2016) who proved that types of landscape have a significant influence on aesthetic preference. Among the types of landscape that generate high visual aesthetic liking in various zones (i.e. at various distances), it is striking that “medium-altitude, low-gradient artificial areas (m_lo_ar)” are represented in the near, medium, and far zones alike. When the photographs showing this type of landscape are examined, the artificial areas in the photographs turn out to consist mostly of historical buildings or remains (Figure 9). This suggests that historical sites are not just important for their cultural value, but also for their visual aesthetic value. Conversely, the category of low visual aesthetic liking (Table 4) was consistently associated with poorly vegetated areas.

Figure 9. Photographs depicting ‘medium-altitude, low-gradient artificial areas’ (the ruins in the Kapıkırı village).

Most studies on the visual aesthetics of landscape have found that water bodies receive the most attention. It is interesting that although the geotagged photographs in this study were mostly taken in and around Lake Bafa, people were primarily attracted to geological landform and vegetation. The results of the study suggest that geological landform and vegetation produced a general positive response. This is reminiscent of Wang, Zhao, and Liu (2016), who showed that an environment that is well maintained and offers ample vegetation may arouse increased aesthetic preferences. Other researchers too have found a correlation between vegetation and environments that are aesthetically attractive (Ulrich 1979; Smardon 1988; Lindemann-Matthies, Junge, and Matthies 2010). Wang, Zhao, and Liu (2016) suggest a possible explanation based on evolutionary theory: flowers may indicate potential future resources, trees may offer shelter and protection, and grassland may facilitate the surveillance of threats (Kaplan and Kaplan 1989).

4.3. Future application/recommendation

The findings of this study have the potential to inform, much-needed restoration and conservation efforts in the study area. The integration of visual aesthetic preferences into land planning decisions can contribute to rigorous and efficient planning and it has been recommended that the spatial information which humans share spontaneously should play a part in landscape planning. Dramstad et al. (2006) asserted that integrating visual aesthetic liking into land monitoring and management enhances the decision-making process. They suggested that the values attached to landscapes could be ascertained through simple methods and used to shape alternative development scenarios. As one aspect of landscape values, landscape aesthetics provide us with a better understanding of how to design desirable future landscapes. Quantification methods such as those presented in this study could be relevant to the success of future landscape design and planning. The methodological approach presented here may serve as a useful tool to quantify the visual aesthetic liking both of different types of landscapes and of specific landscape features.

In combination with other site attributes (e.g. infrastructure, fine-scale features), the complexity of the types of landscape in the study area (Figure 7) far exceeds the resolution of the dependent data which we derived from the photographs. A finer-grained analysis would require a more comprehensive data set, which would likely be resource-intensive and possibly less cost-effective. It is therefore necessary to take a broad view of the results of this study and their implications. The following paragraphs present a broad view of the implications of the study for fluency theory, aesthetic preferences and potential uses.

The finding that accessibility has a major effect on aesthetic liking outcomes implies that landscape planners should work toward providing better access to historical and scenic sites. Indeed, a holistic assessment of the study area based on the metrics used here would suggest that improving accessibility could be the most effective single means of increasing aesthetic appreciation. We note that accessibility is not a metric of fluency as proposed by Mayer and Landwehr (2018); in our OLS analysis, accessibility functions as a control variable.

The study also suggests that (re)vegetating some of the barren areas could make a positive difference. Making the site both more accessible and more attractive to visitors could have significant benefits for tourism and hence the local economy.

It is striking that low-altitude areas inspire only medium and low levels of visual aesthetic liking, and are not represented among the landscapes that generate high visual aesthetic liking, regardless of the distance involved. Given that the study site is a mountainous area, this result may not be surprising. Nevertheless, one might have expected that the lake and valleys, which are situated at low elevations, would be of considerable aesthetic interest. These findings require further investigation.

Several additional inferences could potentially be drawn from the data and analysis presented here. Ultimately, however, aesthetic appreciation (liking) is subjective, and varies with each person who views a landscape. Theory and objective analysis can, at best, approximate the factors and processes involved in aesthetic judgment. Nevertheless, objective analysis provides useful insights and offers a path to consensus in landscape evaluation and planning.

The results of this study illustrate the value of employing machine learning methods to understand how people judge the visual merits of physical settings. As underlined by Jahani et al. (2021), one of the benefits of machine learning algorithms could be their use in image processing and predicting the relationship between spatial features and aesthetic preferences in landscapes.

According to Gosal et al. (2019), social media data can be harnessed to develop a better understanding of the locations to which visitors attach value. Analysis of geotagged photographs in conjunction with content analysis provides us with a better understanding not only of which places users of the environment visit, but also of what it is that attracts them. By using a combination of image content analysis and LCA in this study, we have been able to show how landscape features can be grouped according to the visual aesthetic liking of visitors.

5. Conclusions

This study has used geotagged photographs to examine the relationship between the visual aesthetic value of landscape and landscape types. Accessibility, although not a standard metric of fluency, has been identified as a significant positive factor when evaluating this relationship, whereas lack of vegetation and low elevation have been identified as negative factors. Historical sites have been found to have a positive impact on visual aesthetic liking through the use of viewshed analyses as a way of evaluating the different types of landscape determined through Landscape Character Assessment, taking different zones (distances) into consideration. More generally, the findings of this study indicate that visual aesthetic liking varies by the type of landscape. People are intrigued by natural forms such as geological features or landforms and vegetation.

Advanced algorithms and public data can potentially be used to explore landscape aesthetics which in turn can help to combat the loss or degradation of cultural and natural heritage. Aesthetic preferences can help us to understand visual aesthetic liking and to assess it effectively by objective means, partially offsetting subjective factors. In a formal evaluation and planning context, objective analysis should be seen as an essential tool for comparison and decision-making.

The threshold values and weightings used to assess visual preferences in this experiment have been assigned by the authors or previous researchers, rather than objectively or dualistically. Most landscape research or environmental psychological research is inter-subjective or a combination of objective and subjective. The factors measured by the fluency framework are fairly novel. The fluency concept can be supported with experimental data, and new algorithms can be developed to predict people’s visual aesthetic preferences in various types of landscape. Further research of this kind could bolster the robustness of the method.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

There is no data related to this work.

Additional information

Funding

This research received no external funding.

Notes on contributors

Derya Gülçin

Derya Gülçin received the PhD degree in landscape architecture from Çukurova University, Turkey, in 2018. Her research interests include the spatial analysis of land use/land cover and landscape analysis.

Nermin Merve Yalçınkaya

Nermin Merve Yalçınkaya received her PhD degree in landscape architecture from Çukurova University, Turkey, in 2019. Her research interests include urban landscape planning and impact assessment systems.

Appendix A.

Visually preferred photographs and their locations

Appendix B.

Residual by predicted plot

Appendix C.

Some example photographs with high, medium, and low evaluation results based on OLS model