COVID-19 pandemic changes the recreational use of Moscow parks in space and time: Outcomes from crowd-sourcing and machine learning

The limited access to urban green spaces during the COVID-19 pandemic had a negative impact on the human-nature interaction in cities and human well-being. Number of visitors to green areas, initially declined due to imposed restrictions, was restored after they were lifted as established by several studies across the globe However, little is still known about changes in behavior and preferences of park visitors in the post-COVID time. In this study, we investigated spatial-temporal patterns of recreational activities in the three urban parks in Moscow (Russia) prior, during and after the COVID-19 lockdown (in 2019 and 2020). The selected parks represent two different types: a centrally located park with much infrastructure and open landscapes (Gorky Park) and parks located at the outskirts of the city center with a more forested landscape and little infrastructure (Timiryazevski and Sokolniki parks). Recreational activities were identified based on the analysis of social media photos using machine-learning algorithms. As expected, park closures during lockdown resulted in overall decrease in the number of taken photos. After the parks were re-opened, however, the number of photos did not grow immediately. The number of photos only restored after almost three months, and the visiting peak shifted to autumn. Differences between parks were related to the type of the park and its landscape structure. The lowest decrease in the number of photos was observed for the Timiryazevsky park – a semi-natural green area, while the centrally located Gorky Park was the most affected, likely due to the strictest control measures. In comparison to 2019, photos in 2020 were more evenly distributed across the area in all the three parks. Besides, ‘natural’ areas became the main attractors for the visitors - photos under ‘nature observation’ category became the most popular. Spatial distribution of the recreational activities in post-lockdown period was characterized by larger distances between photos, likely corresponding to the social distancing. COVID-19 pandemic highlighted the value of green areas for citizens, but also changed their recreational preferences and overall behavior in parks. The observed shift from high density of visitors around entertainments and attractions in 2019 to a more homogeneous and less dense distribution along the natural zones in 2020 reveals a new pattern in visitors behavior and preference, which shall be considered in spatial planning of the parks. Increasing availability of natural green areas and their integration in urban green infrastructures can become the most relevant policy to consider the crucial role of urban nature as a source of resilience in turbulent times.

The limited access to urban green spaces during the COVID-19 pandemic had a negative impact on the humannature interaction in cities and human well-being. Number of visitors to green areas, initially declined due to imposed restrictions, was restored after they were lifted as established by several studies across the globe However, little is still known about changes in behavior and preferences of park visitors in the post-COVID time.
In this study, we investigated spatial-temporal patterns of recreational activities in the three urban parks in Moscow (Russia) prior, during and after the COVID-19 lockdown (in 2019 and 2020). The selected parks represent two different types: a centrally located park with much infrastructure and open landscapes (Gorky Park) and parks located at the outskirts of the city center with a more forested landscape and little infrastructure (Timiryazevski and Sokolniki parks). Recreational activities were identified based on the analysis of social media photos using machine-learning algorithms. As expected, park closures during lockdown resulted in overall decrease in the number of taken photos. After the parks were re-opened, however, the number of photos did not grow immediately. The number of photos only restored after almost three months, and the visiting peak shifted to autumn. Differences between parks were related to the type of the park and its landscape structure. The lowest decrease in the number of photos was observed for the Timiryazevsky parka semi-natural green area, while the centrally located Gorky Park was the most affected, likely due to the strictest control measures. In comparison to 2019, photos in 2020 were more evenly distributed across the area in all the three parks. Besides, 'natural' areas became the main attractors for the visitors -photos under 'nature observation' category became the most popular. Spatial distribution of the recreational activities in post-lockdown period was characterized by larger distances between photos, likely corresponding to the social distancing. COVID-19 pandemic highlighted the value of green areas for citizens, but also changed their recreational preferences and overall behavior in parks. The observed shift from high density of visitors around entertainments and attractions in 2019 to a more homogeneous and less dense distribution along the natural zones in 2020 reveals a new pattern in visitors behavior and preference, which shall be considered in spatial planning of the parks. Increasing availability of natural green areas and their integration in urban green infrastructures can become the most relevant policy to consider the crucial role of urban nature as a source of resilience in turbulent times.

Introduction
In December 2019, in China, a new coronavirus (SARS-CoV-2) and associated disease  emerged and then rapidly spread around the world. By March 2020, it has been declared as a global pandemic by the World Health Organization (WHO, 2020). By 2022, the total number of COVID-2019 victims includes over 5.5 M people dead and over 330 M people infected globally. Quarantines, lockdowns, and other restrictive measures taken by governments to combat the COVID-19 spread have been affecting literally everyone. While these measures allowed decreasing the burden on healthcare system and ultimately deaths, self-isolation, physical distancing and limited access to services and recreation had a negative effect on human well-being, physical and mental health (Dushkova et al., 2021;Mell and Whitten, 2021;Zacher and Rudolph, 2021).
The benefits of outdoor recreation for human well-being include stress reduction and relaxation, conditions and motivations for physical activity, inspiration for culture and arts, educational (e.g. birdwatching) and aesthetic values (Andersson et al., 2014;van den Bosch and Ode Sang, 2017). In times of the COVID-19 pandemic, the access to urban greenery became even more crucial for mental and physical health. For example, studies done in Oslo, Norway (Venter et al., 2020), in Brisbane, Australia (Berdejo-Espinola et al., 2021) and Mexico City, Mexico (Mayen Huerta and Utomo, 2021) have demonstrated the crucial role of urban green space in mitigating some of the pandemic-related stresses. Visits to the nature areas as well as views of greenery were associated with more positive emotions among residents of Spain, United Kingdom, Germany, France, United States, Portugal, Italy, New Zealand and Mexico (Pouso et al., 2021). Moreover, urban greenery could compensate the reduction in opportunities for leisure-time physical activities during the pandemic, when most of sports facilities were closed (Yang et al., 2021). Conversely, a limited access to urban green spaces was reported among the most negative consequences of the pandemic restrictions (Scott, 2021;Ugolini et al., 2020). For example, the mental health risks imposed by a limited access to urban green areas have already been reported for cities in Russia and Australia (Dushkova et al., 2021). Italian and Spanish residents who could not visit urban green areas during lockdowns reported deficiency of these visits, which was alleviated in minor extent for those with some view of greenery from the window (Ugolini et al., 2021). In Singapore, where stay-at-home regime lasted for seven weeks in 2020, a significant increase in depressive brain symptoms was observed on brain scans (Olszewska-Guizzo et al., 2020). These and other social surveys highlight the increased public demand for outdoor recreation after the lockdowns and other measures lifted but lack spatial explicitness and are not related to the landscape structure of urban green areas. At the same time, spatially explicit information reflecting pandemic-driven changes in recreational use of green areas and behavior, is still scarce but very necessary to rethink their spatial structure, management and services in order to facilitate adaptation to the post-COVID realities (Geneletti et al., 2022;Mell and Whitten, 2021).
Public participation GIS tools can provide a relevant solution to capture the spatial distribution of recreational activities attached to urban green spaces (Ives et al., 2017;Tyrväinen et al., 2007). Lately, crowdsourcing has often been used to collect spatial data generated by large groups of people on the web, for example publicly available data on social media platforms such as Flickr and Instagram (Dunkel, 2015;Huang, 2016;Huang et al., 2018) or check-in data obtained from social media platforms such as Weibo or WeChat (Fangzheng et al., , 2020c. In comparison to the social surveys and expert interviews, social media data is less structured, but on the other hand includes various data categories (e.g., solicited audio and written diaries, photos, pedestrian and riding tracks), which can be interpreted to analyze public perception of urban green areas and provided benefits by identifying the points of attraction. Such approach allows reaching and investigating a representative sample at a low cost and without involving direct interaction with respondents, which is particularly convenient during pandemic (Langemeyer et al., 2018;Tenerelli et al., 2016;Yoshimura and Hiura, 2017). Moreover, when applied together with advanced technologies, such as image detection by machine-learning, analysis of large datasets becomes less labor-intensive, and allows assessing spatial variability and temporal changes in recreational use and activities of urban green spaces (Bubalo et al., 2019;Levin et al., 2017;Richards and Friess, 2015). For example, Nițȃ et al. (2021) collected and analyzed Instagram posts for various parks in Bucharest (Romania) demonstrating patterns of number of photos while focusing on text analysis done by manual coding. Huang et al. (2022) investigated how public values associated with New York City parks has shifted between pre-COVID (i. e., from March 2019 to February 2020) and post-COVID (i.e., from March 2020 to February 2021) using a short modelling technique on Twitter posts. Cheng et al. (2021) explored the effects of urban parks on residents' happiness using a regression model on data from geo-tagged posts on Sina Weibo (Chinese Twitter).
In this study, we investigated spatial-temporal patterns of recreational activities in the three urban parks in Moscow (Russia) prior, during and after the lockdown related to the COVID-19 outbreak. Parks were chosen to represent different types: from a centrally located park with much infrastructure and open landscapes (Gorky Park) to other two located at the outskirts of the city center with a more forested landscape and little infrastructure (Timiryazevski and Sokolniki parks). To do so, we conducted machine-learning analysis of social media photos collected for two years (2019 and 2020) representing periods prior, during and after the lockdown. We hypothesized the following: 1) an increase of the overall number of visits in post-lockdown period; 2) changes in the spatio-temporal patterns of the taken photos and the recreational activities depicted on them; 3) interrelation between these patterns, landscape structure and availability of infrastructure and facilities in the parks.

Study areas
Moscow city (56N; 37E) is the capital of Russia and one of the largest urban areas in Europe. Sustainable development of urban green infrastructure is claimed among the main priorities by the Moscow government (Kulbachevksy, 2021). Green areas including urban forests, city and district parks, courtyards and green lawns cover more than half of the city area  and provide important ecosystem services, among which climate mitigation, air quality control and recreation are considered the most valuable (Dushkova et al., 2022;Klimanova et al., 2018;Matasov et al., 2020). The research focused on the three green spaces in Moscow: Sokolniki, Gorky and Timiryazevsky parks. They are comparable in area and distance from the city center, but different in functionality and landscape structure. Timiryazevsky park (~ 400 ha) is in fact an urban forest. For many years, the territory had a status of a natural protected area and was mainly used for the educational and research purposes. After 2010, the area is open for public, but still is not intended for mass recreation due to underdeveloped infrastructure. The average number of visitors is less than 5 thousands per day (Dovletyarova et al., 2017;Reitz et al., 2021). The Gorky Park (~ 375 ha) is the central and historically the most famous park in Moscow (Kalyukin et al., 2015;Shaw, 2011) attracting tourists and citizens from all over the city. An average daily number of visitors is 40 thousand during the week and up to 250 thousand on weekends. The park area is very diverse including the central part, a historical Neskuchny garden, Vorobyovy Gory natural area and an entire embankment. Sokolniki park (~ 640 ha) is a more homogeneous green space, which used to be a suburban forest and has been significantly transformed for recreational purposes during the last decades. In 2019, the park hosted over 10 million visitors or on average 30 thousand visitors per day (Fig. 1).
The measures that were adopted in Moscow during the COVID-19 pandemic have been among the strictest in the country, and included almost 1.5 months of a complete lockdown. This regime took place from 30th March to 9th June 2020 and included among other restrictive measures the stay-at-home order, movement control (e.g. permits for using transport) as well as a ban of visits to recreational areas, including green zones (Dushkova et al., 2021). The restrictions to visit urban parks were lifted on the 1st of June 2020. Starting from this date, the walks in the parks were allowed but with the obligatory wearing of a mask and respecting social distancing. The restrictions were canceled on June 23, 2020.

Data collection and analysis
The research aim was addressed by the following consequent steps: i) Data acquisition and preparation: geo-tagged photos for the three parks were collected via API query from the social media for the period 2019-2020 and the raw data was cleaned to exclude commercial photos, indoor locations and other irrelevant photos; ii) Recreational activities detection -the recreational activities for the photos were identified  based on the pre-trained machine-learning algorithm; iii) Spatial analysis: land cover maps of the parks were derived based on the satellite images and the infrastructures of the park were mapped based on the OpenStreetMap (OSM) features classes and the spatial relationships between the recreational activities and land covers/infrastructures were analyzed; iv) COVID-19 effects analysis -the dynamics in the recreational activities and their spatial distribution within the parks was analyzed to investigate the effects of COVID-19 lockdowns and other factors (such as public holidays and weather conditions) (Fig. 2).

Data acquisition and preparation
VK (short for its original name VKontakte, vk.com) is a Russian online social media and social networking service which supports sharing georeferenced image and video content. According to Sim-ilarWeb, VK is the 8th most visited social media network in the world (SimilarWeb) and the most popular in Russia (SimilarWeb). Unlike Instagram, it still has an open API for querying images by geolocation, moreover it uses a real shot position (GPS longitude/latitude of the phone) from the image meta information rather than a geo-referenced tag as it is implemented in Instagram. These two factors along with VK's popularity determined its selection as the data source for the study.
We used an API-based dynamic programming scrapping algorithm to collect all the available data by the following steps: i) choose 100 m query radius to hit the optimal cells; ii) divide all of the key plot's areas into regular grid of 100 × 100 m; iii) query each grid cell center starting with full time limits (2019-2020); iv) iterate pages until actual limit, caching responses and repeat until pages estimate matches the real count of obtained items. Totally, we achieved 49,387 photos: Timiryazevsky park -3018, Gorky park -31,621, Sokolniki park -14,748.
The obtained data included a significant number of irrelevant photos and advertisements. To reduce their share in the datasets, the three consequent filtering procedures were implemented. First, the number of photos per owner was analyzed and the owners with over 600 images were identified as commercial accounts and deleted. Second, the photos which locations overlaid with the OSM buildings polygons were excluded since they were either irrelevant or commercial. Finally, the mis-tagged photos with broken coordinates or photos from cars/trains/ other indoor locations outside buildings polygons were also filtered out. In result, the clean dataset included 15,067, 7639 and 1985 photos for the Gorky park, Sokolniki park and Timiryazevsky park respectively.
After preparation of all the images, the users' / groups' API was used to collect the information about the image owners. In the further analysis, additional information including user ID, sex and age was considered.

Recreational activities detection
An object detection neural network was used to determine depicted activities on the images. This approach allows identifying different activity patterns on one image rather than classifying it as a single unit. The Google OpenImages V6 dataset (https://storage.googleapis.com/ openimages/web/index.html) was taken as a starting point to develop and train the model. It has approximately 9 million images labeled in more than 600 classes. These classes were grouped into the following 10 recreational activities: recreation with children, dog walking, horseriding, nature observation, active sports, water recreation, winter recreation, music and drawing, passive recreation, working in a park, eating food (Fig. 2, Suppl. Material A). An extra class 'person' was added to label all photos with persons. All the photos not allocated to any of the above classes were referred as the 'others' class.
For each of these grouped classes, we collected ~10k images from Google OpenImages API and split them for training and validation datasets. A modern YOLOv5x network (https://github.com/ultralytics/ yolov5) pretrained on imagenet was used for object detection. In comparison to other neural network architectures (e.g., Faster R-CNN or RetinaNet) YOLOv5x has an advantage in terms of both accuracy and speed. Specifically, YOLOv5x achieved an average precision of 83.5 % on the COCO dataset, while Faster R-CNN and RetinaNet achieved an average precision of 78.0 % and 79.5 %, respectively. Furthermore, YOLOv5x was able to run at a faster speed, with a processing time of just 33 ms per image, compared to Faster R-CNN's processing time of 150 ms per image and RetinaNet's processing time of 116 ms per image. Additionally, YOLOv5x has several features that make it well-suited for object detection tasks, including its ability to handle multiple scales and aspect ratios, its robustness to occlusion, and its high level of accuracy even at low resolutions. For our research, YOLOv5x network was used for the object detection in 2 steps: i) 30 epochs (8000 train/1000 validation), learning rate = 0.01, ii) 3 × 10 epochs (120,000 train/2600 validation), learning rate = 0.0032 (Suppl. Material B). The final quality of the trained model was measured by the mAP (mean Average Precision) metric: -mAP@0.5 = 0.55 (average across all classes for fixed IoU=0.5 threshold); -mAP@0.5:0.95 = 0.4 (average for all classes and for all IoU thresholds from 0.5 to 0.95 with step = 0.05) (common metric used for object detection, also called box-AP).
The resulting model showed industry average mAP quality comparable with the recent leading scientific results for the similar detection tasks (https://paperswithcode.com/sota/object-detection-on-coco) (Everingham et al., 2010(Everingham et al., , 2015Russakovsky et al., 2015). Further, we have also merged activities detected on photos taken from the same point (similar geolocation), by the same user (similar user ID) and during the same day. This has reduced the total number of photos to 20, 299 including 12,721, 6045 and 1533 for the Gorky park, Sokolniki park and Timiryazevsky park respectively. Hereinafter we refer to "photos" meaning the cleaned and merged dataset of recreational activities in 2019 and 2020.

Park's landscape and socio-economic mapping
The OSM open spatial data and Sentinel-2 satellite imagery data with a spatial resolution of 10 m were used to relate the photos' locations to the infrastructure/land cover of the parks. The OSM are freely distributed spatial datasets (point, line and polygon) classified according to their type: e.g. road network, land use types, points of interest etc. The OSM data were accessed through osmnx library (Boeing, 2017) within python 3 environment. Only the following spatial feature classes relevant for identifying the spatial patterns in the distribution of photos within parks were used: i) Points of Interest: cultural and arts facilities, amusement rides, bars, fast food outlets, beauty salons, shops, bicycle rental outlets, cafes, flower shops; ii) Transport: car parks, public transport stops; iii) Footpaths; iv) Water objects; v) Sports facilities: stadiums, sports grounds etc. For all the uploaded photos, the minimum distance to all the mentioned above spatial objects was calculated using python 3 geopandas library to check statistically significant links between the types of activity on the photos and their proximity to the certain infrastructure in the park. The following main landscape classes were retrieved for all the parks: water, sealed soils, tree and shrub vegetation, lawns, and open ground as a result of the rule-based classification of Sentinel-2 mosaic (cloud free images acquired June 6, 2019). The classification consisted from the retrieval of sealed surface based on soil-adjusted vegetation index threshold and further sub-pixel classification of remaining classes based on spectral signatures of each class (see Dvornikov et al., 2021 for details). The total validation accuracy was 80 %. Each photo was related to one of the landscape classes based on spatial analysis. The spatial relationship between the recreational activities on photos and land cover classes/proximities to POIs and natural objects was analyzed to explore connections between the dominating recreational activities and internal structure of the park. We separately investigated the chances of observing activities on photos for each park and year (2019/2020) by building logistic regression models where the activity labeled (0/1) was a response variable and distances to different objects/land cover class were used as independent variables. Statistical analysis was performed in R environment (stats package) (R Core Team, 2017).

Spatio-temporal dynamics of recreational use
To analyze the effect of the COVID-19 lockdown on the temporal (before, during and after lockdown) and spatial (distribution within a park) patterns in the recreational use, a chronology of the pandemic mitigation measures was compiled. The timeline of the COVID-19 mitigation measures (i.e., self-isolation and lockdown, closure of the shops, cafes, parks, etc) was created based on the consequent analysis of the official documents, mainly decrees of the Moscow Mayor. Weather conditions (noon temperature, humidity, and wind speed), public holidays, weekends and social event (e.g., festivals) were also considered as factors contributing to spatial-temporal dynamics of the recreational activities. The data on these factors was collected for 2019 and 2020 from the open sources. Weather conditions were obtained from the RP5 service. The number of photos was analyzed in a relation to median noon air temperature, humidity, and wind speed on a daily and weekly basis (multiple linear regression, some variables were log-or sqrt-transformed to fulfill the normality of distribution requirement). Public holidays, working and non-working days were distinguished based on the official calendar. The official websites and pages in social networks were used to determine the days of entertainment, sports, cultural or educational events in the parks. See in Fig. C1 in Suppl. Material C for an example of the timeline overlaying air temperature data and the lockdown period. Differences between the number of photos (statistics by months/years), ages of social network users, and the distances between geo-tagged photos were analyzed by applying a one-way ANOVA test.

Results
We present our findings in tree sub-sections. In Section 3.1 we focus on the total number of photos and unique users who took them, specifically if there were any changes between 2019 and 2020 (pre-, during and post-lockdown). Then we investigate differences between parks, considering socio-demographic characteristics of visitors and weather conditions. In Section 3.2 we focus on the spatial distribution of taken photos and explore their relationships with landscape structure and available infrastructure. We also analyze the effect of COVID-19 restrictions on the number, spatial distribution and the average distance between the photos in parks. In Section 3.3, temporal dynamic in the dominant type of recreational activities and their spatial distribution is explored in connection to landscape structure, available infrastructure, weather conditions, public holidays, or other events.

Changes in the number of photos and unique users' profiles
The COVID-19 pandemic and, specifically, lockdown in April-June 2020 affected the recreational use of the three Moscow parks considerably. As expected, the restricted access to the parks resulted in a rapid decrease in the number of unique users in comparison to the same months in the pre-pandemic 2019 year (Fig. 3A). The dynamics of unique active VK users posting photos in the parks has shown a "shifted peak" in 2020 with a maximal number observed in September (Fig. 3B). We also found a significant increase in the age of all park visitors for the period of April and May: 32 years median in 2019 versus 38 years in 2020 (one-way ANOVA, F-value: 85.2, p < 0.05). However, in July, the median age of the visitors in 2020 were 6 years younger compared to 2019: ~30 and 36 years correspondingly (F-value: 38.1, p < 0.05). Starting from August, a significant difference in age of the visitors before and after lockdown was not observed. The average visitors' gender ratio did not change between the years. Females posted two times more photos than male visitors. Moreover, the shifted peak of recreational activities in autumn 2020 was mainly provided by female visitors.
We have also observed different changes in the number of photos between the three parks (Fig. 4, Suppl. Material D & E). The average number of photos per day during lockdown in 2020 in comparison to the same period in 2019 decreased in 2.7 times in Sokolniki park (Fig. D1 in Suppl. Material D); in 1.1 times in Timiryazevsky park (Fig. D2 in Suppl. Material D) and in 2.4 times in Gorky park (Fig. D3 in Suppl. Material D). The highest decrease in monthly average number of photos was shown for the Gorky Park -8 times less in April and 10 times less in May, based on the month-to-month comparison between 2019 and 2020 (Table E3 in Suppl. Material E). A similar pattern was obtained for Sokolniki, where the number of photos decreased 4 times in April and 7 times in May (Table E2 in Suppl. Material E). The difference between the months is likely explained by the cancellation of the traditional May holidays in 2020.
Lifting the lockdown restrictions did not result in an immediate return to the pre-pandemic number of photosin June 2020 it was still slightly lower or similar compared to the 2019. A rapid growth in the number of photos was shown for the Gorky park already in July, when it was 4-5 times higher than in 2019 based on the month-to-month comparison. It continued in August (almost 3 times), September (almost 4 times) and October (almost 5 times). Even in November, when outdoor activities are usually hampered by cold weather, the number of photos in 2020 reached 850, which was comparable to the summer period in 2019 (Table E3 in Suppl. Material E). In Sokolniki, an increase in the photos' number was gradual in July and August and rapid in September-November, when the results were up to 3 times higher than in the prepandemic year (Table E2 in Suppl. Material E). Holidays and events could have triggered the rapid increase in September. However, afterwards it continued without any external impact and regardless the weather conditions. Apparently, the effect of the postponed demand in green infrastructure resulted in this time lag. Timiryazevsky park was less affected by lockdown. A rapid decrease (5 times compared to 2019) was shown only for April, whereas already in May there was a slight increase in the number of photos (Table E1 in Suppl. Material E). In August-October, the visiting activity was continuously growing achieving the total number of photos 2-3 times higher than in the prepandemic time. Likely, the lockdown measures were more strictly regulated in the central parks (i.e., Gorky and Sokolniki), that reoriented the citizens' demand in green infrastructure towards the local district green areas (i.e., Timiryazevsky park).
Weather condition can clearly influence the number of citizens. To consider it, we compared the regression model built up for 2019 and 2020, where the number of photos were related to the meteorological parameters (i.e., median daily noon temperature, humidity and wind speed). In 2019, the weather conditions significantly contributed to the social activity in parks on a weekly basis (linear model, R 2 adj = 0.71, p < 0.01). In 2020 though, the weather conditions explained 20 % of the variance in the total number of photos in all parks on a daily basis (linear model, R 2 adj = 0.2, p < 0.01, all parameters --were significant predictors), and 32 % of the variance -on a weekly basis (linear model, R 2 adj = 0.32, p < 0.01, median daily noon temperature was the only moderately significant predictor). The main outliers in the model were week numbers associated with the time period starting from the second half of the August till October. It confirms that the shifted peak in the number of visitors was not caused by the weather, but, in contrast, occurred contrary to unfavorable weather conditions.

Changes in spatial distribution of photos
In comparison to 2019, in 2020 photos in all the three parks were more evenly distributed across the area and were re-directed towards green infrastructure and natural sites (Fig. 5). Comparison between the parks showed different patterns in internal spatial distribution of photos, which was explained by differences in the spatial structure of the parks. In Gorky Park, the three hotspots with the highest number of photos were clearly identified in 2019. The first one in the north-eastern part corresponded to a large area including the main entrance and the central 'parade' zone where all the facilities and exhibitions occurred. The second attraction area coincided with the embankment of the Moscow River. The third one was the viewpoint on the top hill of the Vorobyovy Gory, with the highest density of photos (i.e., number of photos per area). In 2020, the number of photos was partly redistributed from the three hotspots to the in-between territory including semi-natural areas of the Vorobyovy Gory slopes and a historical Neskuchny garden. Although this area was dominated by the woody vegetation with less infrastructure for recreation or sports, the total number and density of photos here considerably increased in 2020 (Figs. 5C, 5F). Spatial distribution of the photos taken in Sokolniki park in 2019 clearly showed two main areas of attraction: the main entrance zone and the eastern part of the park where sport stadiums and playgrounds are located (Fig. 5B). In 2020, the number and the density of the photos in this sport zone decreased and at the same time more photos were distributed across the natural areas and landscape park zone (Fig. 5E). The total number of photos in Timiryazevsky park was considerably less compared to the others, which is likely explained by scarcity of the recreational facilities and infrastructure. Lack of infrastructure and particular attraction points also resulted in the most homogeneous spatial distribution of photos, which mostly concentrated along the walking paths with a slight increase near the entrances (Figs. 5A, 5D). At the same time, in 2020, there was an increase in the number and density of photos taken near the Bolshoi Garden Pond and along the Zaranka River in the northern part of the park (Fig. 5D).
During the lockdown months (March-May), the distances between the photos was significantly (one-way ANOVA, p < 0.05) higher than for the other moths, reflecting the limited number of visitors and social distancing. After lockdown was lifted and visitors returned to the parks, the average distance between the photos taken in 2020 remained higher compared to 2019 for the Gorky Park and Sokolniki park for most of the season. This can be considered a spatial proxy of post-lockdown outdoor recreation, explained by the social distancing requirement and reorientation of the visitors' preferences towards nature observation activities. In Timiryazevsky park, the same pattern was obtained for June-July and November-December, whereas the opposite was observed in August-October (Fig. 6). In 2020, walking activity has been shifted further from the walking paths to viewpoints (p < 0.05) in all parks. Art objects, fountains, benches and pitches became the local attractors of the visitors, whereas the number of photos taken at the sealed areas (close to building or roads) reduced considerably.

Changes in the recreational activities depicted on photos
The COVID-19 pandemic's restrictions affected not only the number of photos taken in the parks, but also the breakdown of recreational activities depicted in them using machine learning. In all three parks, in 2020, compared to 2019, the popularity of nature observation activities increased (Fig. 7). Likely, a growing interest to nature was facilitated by a long stay indoors. However, a limited access to some attractions remaining after lockdown could also reorient the visitors to more natural areas. In Sokolniki, nature observation, water and winter activities partly substituted sports activities. From June to November the increased number of photos with children was observed (Fig. F1, F2 in Suppl. Material F). In Gorky Park, nature observation became more popular whereas the number of photos with passive recreation reduced. In Timiryazevsky park, nature observation became a predominant activity regardless the season (Fig. F2 in Suppl. Material F). The popularity of sport activities and water observations also increased especially in June-September reducing the share of photos with food and dogs (Fig. F1, F2 in Suppl. Material F). Interestingly, nature and water observation were more popular during the working days or weekends, than on holidays and summertime.
Remarkable changes in the spatial distribution of recreational activities in the parks were observed after lockdown (Suppl. Material G). Nature observation (Fig. 8) and passive recreation (Fig. G1 in Suppl. Material G) were spread over considerably larger areas in 2020 compared to 2019. The differences in landscape structure as well as the availability of infrastructures in the three parks were also reflected in the spatial distribution of recreational activities. The odds of finding passive recreation on photos were higher if the photo was located within the non-green surface area (sealed/bare soil) whereas nature observation activity coincided with natural and urban green zones (p < 0.05). The frequency of recreation with children was also spatially correlated with the green infrastructures ( Fig. G2 in Suppl. Material G). More photos with children were observed within urban tree patches and lawns closer to bench and picnic points but at the distance to paths and viewpoints. Sport activities were mainly related to the stadiums and playgrounds, however the distance to the paths in 2020 was slightly smaller compared to 2019 (Fig. G3 in Suppl. Material G). In general, the de-centralization pattern was more evident in 2020 in Sokolniki and Gorky Park, whereas in the Timiryazevsky park the same pattern was already observed in 2019. Weather conditions (noon temperature and wind speed) had a significant influence on the nature observing recreation in all parks (the increase in temperature and less stronger winds increase the chances of finding the nature observation activities on photos, p < 0.05). The observed relationships between the prioritized activities and landcover/land-use categories can be considered for the landscape planning and maintenance of the green areas in the post-COVID times. For example, the increasing demand in nature observation shall be supported by expanding the natural (non-disturbed) areas and facilitating the access to existing ones.

The effect of COVID-19 lockdown on recreational use and activities
As expected, the restrictions brought by COVID-19 lockdown in April 2020 in Moscow (i.e. park closures) resulted in a drastic depletion in recreational use of the parks. The total number of park visitors who have taken photos in April-May 2020 decreased 3-10 times based on the month-to-month comparison with the previous year. Similar outcomes were reported for many cities where lockdown measures included a limited access to public green spaces (da Schio et al., 2021;Geng et al., 2021, Noszczyk et al., 2022. In some places where restrictions didn't  Fig. 7. Breakdown of number of photos taken in Timiryazevsky, Sokolniki and Gorky Parks in 2019 and 2020 by recreational activities depicted on them ("passive" = passive recreation; "nature" = nature observation; "water"-= water recreation; "winter" = winter recreation; "children" = recreation with children; "sport" = active sports; "dog" = dog walking; "horse" = horse-riding; "food" = eating food; "music" = playing music and drawing).
involve parks closure (e.g., in Hong Kong, Singapore, Tokyo, Seoul and Oslo) an opposite pattern with an increased recreational use of urban green spaces was observed (Venter et al., 2020). In Moscow, the central Gorky Park was the most affected by the lockdown restrictions, whereas the number of visitors in the 'semi-natural Timiryazevsky park decreased less. Partly it could be explained by lower enforcement or at least lower control of the lockdown restrictions. More likely, the landscape structure dominated by woody natural areas without particular local attractors (e.g., recreational facilities or viewpoints) resulted in a more homogeneous and less dense distribution of the visitors. Generally speaking, the observed recreational activities (even though in a very limited amount) in the parks which were officially closed for visitors is an interesting outcome, reflecting the ambiguous reaction of citizens to the lockdown measures and possible mistrust to the governmental pandemic measures (Vardavas et al., 2021). According to the survey done by High School of Economics in September 2020 (HSE Office for Expert Analysis, 2020) about one third of Moscow population were identified as so-called "COVID dissidents" i.e., people who did not believe in the pandemic, rejected to follow restrictions, and therefore could ignore the ban for accessing green areas during lockdown. Moreover, the lockdown period and limited access to urban parks coincided with the days traditionally spent on picnics and barbecues. As these events were also prohibited, it could have generated more denial and illegal use (Erdönmez and Atmiş, 2021).
The  (Venter et al., 2021). However, lifting lockdown in June didn't not result in an immediate restoration of the visitors' number, and it took several months for citizens to return to the parks. This effect could likely be explained by an increased concern over health risks from spending time in public places during the first COVID-19 wave (Xiao et al., 2021;Bristowe and Hechert, 2023;Huang et al., 2022). We observed a shifted peak in the number of visitors in September, October and even November, which are usually not popular for outdoor recreational activities due to the cold weather. In September, days off and events could have triggered the rapid increase in photos, however, this trend continued even afterwards without any external impact and regardless weather conditions. Apparently, the effect of the postponed demand in green infrastructure resulted in this time lag. This postponed demand for recreational activities not only restored the pre-pandemic visitation load, but even increased it 2-4.9 times. Female visitors were mainly responsible for this shifted peak in the number of photos taken in parks in autumn 2020. Other studies (Burnett et al., 2021;da Schio et al., 2021;Lin et al., 2014) have also demonstrated that women tended to value nature and give more priority to recreation in urban green spaces than man.
The COVID-19 lockdown in April-June 2020 had a remarkable effect on the spatial distribution of the recreational activities in parks. The photos taken in 2020 were more evenly spread throughout the parks, with the largest areas covered by nature observation and walking. A reduction of spatial density could be partly explained by the social distancing requirement and escaping over-crowded areas as it was reported in other studies (Lopez et al., 2021;Shoari et al., 2020). This outcome is in coherence with the similar studies conducted in Croatia, Israel, Italy, Lithuania, Slovenia and Spain that showed changes in reported visitation of urban green spaces in terms of frequency, motivation and distance travelled to reach those places as a result of the measures imposed by authorities to contain the outbreak during the COVID-19 pandemic (Ugolini et al., 2020). Australian respondents demonstrated changes in visitation of green spaces, but also in their motivations from improvement of personal well-being rather than occurring community capital (Berdejo-Espinola et al., 2021). Similarly to the results obtained for Moscow parks, a considerable shift of recreational activities to nature observation was reported for Warsaw (Poland) (Grzyb et al., 2021). In Beijing (China), the highest visitors' priority was given to 'observing the scenery', especially plants (Zhu and Xu, 2021). A reported increase in outdoor sports highlights the significant role in shaping physical activity (Yang et al., 2021).

Methodological limitations and future research directions
Combing social media data with machine learning is a novel and promising approach to explore changes in recreational activities, however at least two methodological limitations shall be considered: 1) uncertainties in image analysis and 2) limitations of social media as a data source.
Uncertainties in image analysis are caused by a high number of irrelevant photos and advertisements/posts from commercial accounts in the initial dataset, machine-learning limitations for detecting objects of recreation activities, and the accuracy of GPS location (Guerrero et al., 2016;van Zanten et al., 2016). The identification of the 13 recreation-related classes from 600 classes labeled in the Google OpenImages V6 dataset cannot be rigorously characterized as the only correct variant therefore an additional verification is needed. The model accuracy could be increased if additional labeling were used on the subset of obtained data to eliminate possible training/testing distribution skew (for example, additional labeling of some classes of photos by hand). Calibrating and testing the model for a bigger set of green area will allow reducing the limitations of the machine-learning algorithm as it was demonstrated for New York City (Hamstead et al., 2018) and Xuzhou City (Dai et al., 2019).
A concern about social media as a representative data source is a more fundamental limitation of the approach . Although the number of social media users increases every year, the average user profile is biased in terms of age, education and other demographic criteria as well as ability or motivation to use social media and post photos there (Tenerelli et al., 2016;Yoshimura and Hiura, 2017). In our study, the median age of park visitors was 35 in both 2019 and 2020, but the distribution was skewed to the younger age. Another limitation refers to the fact that people are likely enthusiastic to create posts on social media with bright and positive photos rather than sharing their negative emotions or experiences. However, significant correlation between the number of photos and number of visitors in recreational areas was demonstrated by previous studies (Levin et al., 2017). Besides, photos are usually taken to depict sites with high aesthetic value (Figueroa-Alfaro and Tang, 2017), thus they shall be considered relevant proxy indicators of the visitors' preferences. The limitations can be reduced by complementing social media data with social surveys including Volunteered Geographical Information (VGI) and participatory mapping or non-participatory observations in parks as it is proposed by recent studies (e.g., Herman and Drozda, 2021;Johnson et al., 2019;Sim and Miller, 2019;Wang et al., 2018). Another way to reduce uncertainty and verify the results is to analyze text comments to photos in social networks (e.g., Dai et al., 2019) or to use data from image-based and text-based platforms that present different types of content (tweets, photos, comments, statistics and etc) (e.g., Lyu and Zhang, 2019;Nițȃ et al., 2021;Pickering et al., 2020). Therefore, despite the discussed limitations, social media remains an important source of data for researchers to reveal people's behavior and preferences for the recreational activities in urban green spaces (Guerrero et al., 2016;Kothencz and Blaschke, 2017;Roberts et al., 2017).

Implications for urban planning
Our findings have highlighted the crucial role of urban nature as a source of resilience, which is especially relevant in times of high turbulence (Bristowe and Heckert 2023;Dushkova et al., 2021;Rice et al., 2020). In the COVID-pandemic time, availability of urban green areas becomes a vital resource of recreation for citizens, who are generally bound by the tightest restrictions (Rice et al., 2020;Bristowe and Heckert, 2023). In Moscow, the availability of this resource is far not equal as it was demonstrated by Dushkova et al. (2021Dushkova et al. ( , 2022. The quality, functionality, and location of open green spaces varied considerably between the neighborhoods, and in many cases several communities from a neighborhoods suffered from a limited access to blue and green infrastructure. Provisioning of the equal (or at least comparable) availability of urban green areas throughout the entire city area is one of the clear messages for urban planning and policy implication.
The second important message is changes in behavior and preferences of the visitors in the parks after lockdown, which shall be considered in park planning and maintenance. Urban green areas are the sources of resilience, places for well-being, physical exercise and alternative ways of socializing. It is important to re-organize the space in order to facilitate social distancing whilst maintaining physical and mental health effects (Slater et al., 2020). Changes in preferences in favor of nature observations and walking across green areas revealed by the study stress that the balance between integrated urban nature and the preservation of more 'natural' and biodiverse landscapes is needed (Venter et al., 2021). The most 'natural' Timiryazevskiy park showed less reduction of the visitors during lockdown. This underlines the importance of less rigidly designed, multifunctional spaces that give their users a more flexible approach, allowing spontaneous tactical use (Herman and Drozda, 2021).
Management of parks and activities within parks had to be rethought to adapt to these challenges by expanding forested areas and outdoor recreation spaces. There is less need in complete reorganization of green spaces to areas with an excess of amenities and formal functions. Sometimes it is enough to update the concept or philosophy of the place and just add some elements of infrastructure like parklets in the public spaces, streets or park embarquements; fitness equipment along park paths; children playgrounds or even working places in green areas. Learning from experiences in other places, for example, how Italian urban agricultural farms have adapted their activities during COVID-19 restrictions could also be beneficial (Rusciano and Gatto, 2022). Taking such actions would be most beneficial when done in consultation with local stakeholders and residents identifying and addressing the diversity of human-nature relations and values (Pascual et al. 2017).
The observed increased demand in natural areas with less density of visitors and facilities for individual or small group recreation is likely a new reality for spatial planning and maintenance of urban green areas. Increasing availability of natural green areas and their integration in urban green infrastructures is a strategic solution to be considered by the urban planning policy in Moscow megalopolis.

Conclusions
Based on the machine learning analysis of social media data, we showed that lockdown measures taken in Moscow in April-June 2020 restricted citizens' use of the parks and altered their recreational behavior and preferences. The central and most popular among the citizens Gorky Park was the most affected by restrictions -the number of photos here decreased more than 10 times during lockdown. Lifting lockdown did not result in immediate restoration of the visitation rates, however the postponed demand for outdoor recreation shifted the visitation peak to September-November even with less favorable weather conditions. Overall, the total monthly number of photos after lockdown increased 2-4.9 times compared to the pre-pandemic period, however the spatial distribution of recreational activities changed. Nature observation became the most preferable activity in all parks after lockdown, whereas the number of photos depicting other categories (e. g., recreation with children, dog walking or music and drawing) decreased. The spatial distribution of visitors in the parks became decentralized and less dense. We clearly showed larger average distances between the photos and more photos related to natural woody areas of the parks compared to the sealed areas and central parts which used to be attractors for the recreational activities before lockdown. Apparently after lockdown, visitors prefer spending time in more quiet and isolated spaces with fewer interactions with strangers which probably corresponds to less risks of getting infected. These new patterns in recreational use and activities shall be considered by landscape planners and managers to adapt the spatial structure of urban green spaces to the new visitors' demands and preferences in the post-pandemic times. Moreover, a trained model developed for photo detection also has a practical importance and can be used for similar studies to identify ecosystem services and recreational activities in various green areas.

Funding
Social media data collection and spatial analysis were supported by the Russian Science Foundation project # 19-77-30012. Data analysis and publication preparation were supported by the RUDN University Scientific Projects Grant System.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.