Introduction

Natural heritage sites, core areas for global ecological conservation, are important in ecological, cultural, and scientific research1,2,3. Green landscapes such as forests and grasslands are crucial components of ecosystems in natural heritage sites. Their dynamic changes reflect the health status of regional ecosystems and directly affect the functioning of ecological services, maintenance of biodiversity, and aesthetic value of landscapes4,5,6,7. In recent years, the intensification of global climate change and human activity has had a profound impact on vegetation phenology. The phenological characteristics of green landscapes in natural heritage sites have also inevitably been affected, posing challenges to the ecological protection, management, and sustainable utilization of these heritage sites8,9,10,11.

Phenology is an important field at the intersection of ecology and geography and is often used to explore the interactions between plants and environmental factors, as well as the response mechanisms of ecosystems to climate change12,13,14. Studies have shown that global warming significantly alters the timing of vegetation phenology, including earlier sprouting, extended growing seasons, and changes in the peak periods of the green period15,16,17. However, most of these studies have focused on large-scale regions or the global scale, and there has been little in-depth exploration of the change characteristics of the phenology of green landscapes at a fine scale in natural heritage sites with complex terrains and ecological sensitivity. In summer (e.g., June–July), when the productivity of green landscapes reaches its peak, the phenological characteristics during this critical period are directly related to the carbon absorption capacity of ecosystems, soil moisture regulation function, and the interaction between vegetation and animals10,18,19.

Phenological changes have great significance for ecosystems themselves and have a profound impact on the tourism industry of natural heritage sites. Tourism activities at natural heritage sites rely heavily on the aesthetic value and seasonal characteristics of ecosystems, and phenological changes directly affect the visual attractiveness of landscapes and tourist experiences. For example, phenological characteristics such as vegetation greenness and color changes, are key factors that attract tourists20,21,22. Tourists tend to visit destinations when the vegetation is lush, and the landscape is at its most beautiful, which generally coincides with the peak phenological period23,24. Therefore, phenological changes determine the seasonal distribution of tourism resources and directly affect the spatial and temporal distribution of tourists and the sustainability of tourism activities. Tourism management at heritage sites must clarify whether tourism activities depend on landscape phenology. However, research on how to effectively construct a relationship between the growth cycle of green landscapes and the tourist season is lacking. Thus, a universal theoretical framework and methodology are needed25. It is also unclear whether there is a lagged relationship between seasonal landscape changes and tourist behavior, as well as the role of socioeconomic factors in modulating this relationship.

In recent years, with the rapid development of the tourism industry at natural heritage sites, the impact of human activities on ecosystems has become increasingly evident26,27. For example, the expansion of land use and an increase in artificial heat sources (such as nighttime lights) may have a profound impact on phenological changes28,29. Although many studies have investigated the effects of climate change, little attention has been paid to the complex dynamic relationships between climate change, tourism activities, and phenology at natural heritage sites, leaving a research gap in the ecological management of heritage sites and the optimization of recreational resources. Quantifying the dominant factors influencing green landscape changes at heritage sites helps enhance our understanding of the underlying mechanisms. This will enable the development of targeted measures to counteract the negative trends arising from global climate change and regional socioeconomic development.

Numerous studies have used remote sensing satellite data to depict grassland growth and extract phenological indicators30. Additionally, internet platform data have been employed to represent tourist destination traffic and forecast peak tourist seasons31,32,33. The daily long-time-series nature of these two datasets offers possibilities for analyzing the relationship between grassland phenological periods and recreation periods at the temporal precision level. Therefore, this study quantified long-term grassland growth trends using satellite remote sensing data, modeled tourist flow trends on actual and potential scales using multi-platform data, and innovatively linked grassland growth to tourist visitation data at grassland tourist sites. Using multi-source remote sensing data, meteorological, and human activity data, we explored the spatiotemporal change characteristics (from 2000 to 2022) of the phenology of green landscapes (grasslands) and analyzed the comprehensive action mechanism of climate factors, topographic factors, and human activities on the phenology of grasslands. The main objectives of this study included: (1) identifying the dependence relationship between the phenological periods of grasslands in natural heritage sites and the recreation periods; (2) revealing the spatiotemporal distribution patterns and long-term dynamic trends of the phenological changes of grassland landscapes in natural heritage sites; and (3) quantifying the relative contributions of different driving factors to phenological changes and exploring the comprehensive action mechanism of climate change and human activities. Through the novel Pheno-Tourism, landscape ecology and tourism geography are linked to provide suggestions for the ecological protection and tourism management optimization of natural heritage sites.

Methods

Study design

The phenological stages of the grassland growing season phenological stages were determined using daily grassland Normalized Difference Vegetation Index (NDVI) data. Regional grassland sightseeing periods were captured based on multi-year and multi-platform online tourism-related big data. The correlation between the grassland growth seasons and tourist sightseeing behavior was analyzed using Spearman correlation analysis and dynamic time warping (DTW) methods. Spatiotemporal changes in grassland phenology were explored using the Theil–Sen median slope estimation method and the Mann–Kendall (M–K) test. A multilevel dataset of influencing factors was built, and the Geodetector method was applied to analyze the key factors affecting grassland phenological indicators. Additionally, the potential socioeconomic factors underlying tourist responses to phenological changes in grassland were investigated (Fig. 1).

Fig. 1: Research ideas and methodological design.
figure 1

a Research ideas (shows the grassland NDVI and tourist sightseeing trends over the years; where t is time, SOS is the Start of the Growing Season, EOS is the End of the Growing Season, LGS is the Length of the Growing Season, and Lag is the difference between the date of NDVI peak and the date of sightseeing peak). LGS_1, LGS_2, LGS_n” now denote the Length of the Growing Season in different years, illustrating interannual comparisons of grassland phenology, while “Lag_1, Lag_2, …, Lag_n” represent the time lag (in days) between the NDVI peak and the tourism peak in corresponding years; b methodological design.

Full size image

Study area

Narat Shan is located in the eastern Tianshan Mountains of Xinjiang, China (Fig. 2a, b). It is part of the “Xinjiang Tianshan,” a UNESCO World Heritage Site and is a typical mountainous natural heritage34. These sites are characterized by dramatic terrain, rich biodiversity, and ecological sensitivity, making them crucial for global conservation efforts. Narat Shan is situated between 80°E and 86°E longitude and 42°N and 44°N latitude, with altitudes ranging from 1500 m to over 4000 m (Fig. 2c). The diverse terrain supports different ecological zones, and grasslands and coniferous forests form unique habitat mosaics.

Fig. 2: Study areas.
figure 2

a Location of Tianshan Mountains. b Location of Narat Shan in Tianshan Mountains. c Geographical conditions of Narat Shan. d Distribution of grassland in Narat Shan. e Illustration of grassland phenological indicators based on NDVI (Map Review Number: GS(2020)4619).

Full size image

The region has a continental mountain climate, with an average annual temperature ranging between −2 °C and 5 °C, and precipitation varying between 200 and 600 mm depending on the altitude. Summer is a crucial season for plant growth and the key to studying plant phenology. The green landscape in this area mainly comprises alpine grasslands that support rich biodiversity, including several endemic and endangered species (Fig. 2d).

The unique climatic conditions, topographic complexity, and human-environment interactions of Narat Shan make it an ideal research area for exploring the spatiotemporal dynamics of summer phenology in natural heritage sites. As part of mountainous natural heritage, understanding these dynamics can provide important insights for managing similar mountain ecosystems under the dual pressures of climate change and human activities.

Data sources and processing

Remote sensing image data: The NDVI data used in this study were obtained from the NOAA Climate Data Record (CDR) Program (https://doi.org/10.7289/V5ZG6QH9), covering 1981–2023 (truncated to the first 10 days of that year). These daily raster products were provided at a spatial resolution of 0.05° × 0.05° and were derived from an Advanced Very High Resolution Radiometer (AVHRR) sensor. Daily precipitation and temperature data ranging from 1951 to 2022 were sourced from the China Meteorological Science Information Center (CMSIC; https://data.cma.cn/). Based on regional meteorological stations, the Inverse Distance Weighting method was used for interpolation to obtain the regional daily average precipitation and temperature raster data. The land-use data were sourced from the Level-1 and Atmosphere Archive & Distribution System Distributed Active Archive Center (LAADS DAAC, https://ladsweb.modaps.eosdis.nasa.gov/search/order/1/MCD12Q1--6), and similar to the NDVI data, they were derived from MODIS data products.

Socio-economic data: Data on trends in the online search popularity of tourist destinations were sourced from the Baidu Index (https://index.baidu.com/v2/index.html). Daily index data were obtained by collecting keywords (such as Narat Grassland) from the Baidu Index API daily. Baidu is a mainstream online search platform used in China. The Baidu Index is a data analysis platform based on the massive behavioral data of Baidu users. It is mainly used for the statistical analysis of the scale, trends, user portraits, and related public opinion information of keywords in Baidu searches. In the tourism field, the Baidu Index is used to analyze the online attention to scenic areas, predict passenger flow, and study tourism public opinion31,32,33. The actual review data of tourist destinations were obtained from Ctrip (https://vacations.ctrip.com/) and Dianping (https://www.dianping.com/), and the data on actual consumption behaviors were integrated from these two platforms. As a leading domestic online travel service platform, Ctrip (https://vacations.ctrip.com/) has an intelligent itinerary planning system and provides full-chain services covering air tickets, hotels, and tickets. Dianping (https://www.dianping.com/), by virtue of its advantage in deeply cultivating local life, has included in-depth reviews of merchants in >500 cities, among which tourism-related consumption reviews account for 37%, covering subdivided dimensions such as ticket experience, tour guide services, and transportation connections. Cross-verification of data from the two platforms can effectively restore the specific arrival trends of tourists and actual sightseeing periods at tourist destinations.

Geospatial infrastructure: A digital elevation model (DEM) with a resolution of 30 m was obtained from NASA/USGS (https://lpdaac.usgs.gov/products/astgtmv003/), re-projected onto the World Geodetic System 1984 (WGS84) coordinate system, and mosaicked to fill the gaps between tiles. The boundary data of the heritage sites and tourist destinations were obtained from the World Database on Protected Areas (WDPA) (https://www.protectedplanet.net/) and Baidu Maps (https://map.baidu.com/), respectively. Data on roads, rivers, and land cover (with a resolution of 500 m) were sourced from OpenStreetMap (https://data.maptiler.com/) and the LAADS DAAC MCD12Q1 product (https://ladsweb.modaps.eosdis.nasa.gov/). All data were uniformly projected onto the WGS84 coordinate system, and the resolution was resampled to a uniform 0.05° × 0.05° to facilitate data analysis. ArcGIS Pro 3.0.0 (https://pro.arcgis.com/) was used for data processing. Detailed information is provided in Table 1.

Table 1 Datasets and sources
Full size table

Grassland phenological extraction

First, using land-use data from 2000 to 2022, areas where the grassland plots had never undergone conversion were extracted as sample points to ensure the uniqueness of the plot attributes. Based on the NDVI data from the NOAA CDR program, the NDVI data at the locations of the grassland units were screened and cleaned to remove invalid values and ensure the accuracy of subsequent calculations. Then, the moving average method was employed to smooth the NDVI time series to reduce the noise interference caused by factors such as clouds and atmospheric conditions, while retaining the periodic variation characteristics of NDVI. The Savitzky–Golay filtering method was used for smoothing. This method fits the data through local polynomial regression, effectively filling in missing values and improving data accuracy, thus ensuring the integrity and accuracy of the daily NDVI data.

After preprocessing the NDVI data, the key phenological node indicators were calculated using a code based on R 4.4.1 software (https://cran.r-project.org/). For each pixel, SOS, EOS, and LGS were determined by analyzing the NDVI time series.

Methods for extracting phenological parameters based on NDVI mainly include the threshold, median, and maximum slope methods. In general, the dynamic threshold method selects 20% and 50%35,36. Therefore, in this study, the dynamic threshold of 20% and 50% (Feasibility analysis can be found in the supplementary materials) and the maximum slope method were selected to extract the phenological indicators (SOS and EOS) of the grasslands. The LGS is the difference in the number of days between the SOS and EOS and is used to measure the length of the growing season. The dynamic threshold method is expressed as follows:

$${{NDVI}}_{{ratio}}=\frac{{NDVI}-{{NDVI}}_{\min }}{{{NDVI}}_{\max }-{{NDVI}}_{\min }}$$
(1)

where the \({{\boldsymbol{NDVI}}}_{{\boldsymbol{ratio}}}\) is the fitted \({\boldsymbol{NDVI}}\) on a given day, and \({{\boldsymbol{NDVI}}}_{{\boldsymbol{\max }}}\) and \({{\boldsymbol{NDVI}}}_{{\boldsymbol{\min }}}\) are the maximum and minimum \({\boldsymbol{NDVI}}\) each year, respectively.

Trend analysis of grassland phenological change

After extracting the key phenological node indicators, this study combined the calculation of the Theil–Sen median slope estimator method and the M–K test to conduct a trend analysis of the phenological indicators of each pixel37,38. For the phenological indicator time series (SOS, EOS, and LGS) of each pixel, the data points were arranged in chronological order, and the slope between each pair of data points was calculated. The Theil–Sen median slope estimator method was determined by taking the median of the slopes. This median effectively avoids the influence of outliers in the data, thereby yielding a robust linear trend estimate.

After calculating Sen’s value for each pixel, the M–K test was performed to determine whether the trend of the time series was significant. This test is a non-parametric method suitable for determining whether a monotonic trend exists in the data series. Trend analyses were conducted at the pixel level (0.05° grid) using the Sen median slope estimator method and the M–K test.

$$\beta ={Median}\left(\frac{{X}_{j}-{X}_{j}}{j-i}\right),\,\forall j > i$$
(2)

where \({X}_{j}\) and \({X}_{i}\) are variables in years \(j\) and \(i\), respectively. \(\beta\) is the Theil–Sen slope based on the median values. \(\beta\) > 0 indicates an increasing trend in variables (i.e., SOS, EOS, LGS); \(\beta\) < 0 indicates a decreasing trend. The significance of Theil–Sen’s slope was further examined using the M–K test39,40 by calculating \(Z\) values from Eqs. 3–6:

$$S=\mathop{\sum }\limits_{i=1}^{n-1}\mathop{\sum }\limits_{j=i+1}^{n}{\text{sgn}}({X}_{j}-{X}_{i})$$
(3)
$$sgn({X}_{j}-{X}_{i})=\left\{\begin{array}{l}1,\,\left({X}_{j}-{X}_{i}\right) > 0\\ 0,\,\left({X}_{j}-{X}_{i}\right)=0\\ -1,\,\left({X}_{j}-{X}_{i}\right) < 0\end{array}\right.$$
(4)
$${\rm{Var}}(S)=\frac{n(n-1)(2n+5)}{18}$$
(5)
$$Z=\left\{\begin{array}{c}\frac{S}{\sqrt{{\rm{V}}{\rm{ar}}(S)}},\,S > 0\\ 0,\,S=0\\ \frac{S+1}{\sqrt{{\rm{V}}{\rm{ar}}(S)}},\,S < 0\end{array}\right.$$
(6)

where n denotes the number of years. The absolute values of \(Z\) > 1.65, 1.96, and 2.58 indicate significance at the 90%, 95%, and 99% confidence levels, respectively.

Using the above steps, a trend analysis was conducted on the phenological indicator time series of each pixel to determine its trend direction (positive or negative) and significance level. The research results can be used to analyze the spatiotemporal patterns of phenology. Finally, the analysis results were spatially visualized to display the spatial patterns of changing trends in phenological indicators. In this study, the trend package in R software was used to calculate Sen’s slope values and conduct the M–K test. Combined with the previous results of extracting key phenological node indicators, through trend test analysis, the changing trends and significance of the SOS, EOS, and LGS for each pixel were obtained.

Simulate actual and potential tourist numbers

The number of tourists is an important indicator of the main sightseeing period at a tourist destination. However, because of the difficulty in obtaining tourist data from tourist attractions, it is challenging to acquire real tourist data directly from specific scenic areas to accurately identify the best sightseeing time for a certain place. Therefore, scholars have adopted methods that use indirect data to predict tourist numbers. For example, studies have used the Baidu Index to predict the number of tourists in scenic areas, constructing prediction models by analyzing the correlation between search popularity and tourist flow, whereas others have combined online review data with social media data to simulate the visitation of tourists in scenic areas, achieving good verification results31,32,33. Drawing on these studies, we propose a method to simulate the temporal change trend in the number of tourists from two aspects: (1) using the Baidu Index to represent and simulate the Simulated Potential Tourist Numbers (SPTNs) and (2) simulating the Simulated Actual Tourist Numbers (SATNs) based on actual consumption data from Ctrip.com and Dianping.com.

The Baidu Index is an analytical tool developed by Baidu Inc., based on users’ search behaviors and other relevant indicators on its platform, which reflect the online popularity and attention of a certain keyword. The Baidu Index data cover a long period (dating back to 2011) and show strong timeliness and trends. In this study, we selected the keyword “Narat Grassland” and obtained relevant data through the API interface of the Baidu Index to represent temporal trends in the potential number of tourists. The potential number of tourists reflects the size of the user group interested in a certain tourist destination and represents the willingness of tourists to travel. Through the API interface, we obtained the review data (with a time scale from 2016 to 2022) of the ticket sales entrances of “Nalati Scenic Area,” “Kuerdening Scenic Area,” and “Karakulun Scenic Area” on Ctrip.com and Dianping.com, and used these to simulate the actual consumption behaviors of tourists, further reflecting the changing trend in the actual number of tourists.

To ensure the accuracy and reliability of the data, all the acquired raw data underwent strict data cleaning and preprocessing. First, for the review data, we removed obvious outliers, such as irrelevant comments, misreviews or malicious comments. Second, we extracted valid fields (such as timestamps and review contents) and removed irrelevant information, and then used the normalization method to standardize the data from different sources to the same scale for comparison and analysis. In addition, to reduce noise interference in the data, we performed Savitzky–Golay filtering and smoothing on the original Baidu Index and reviewed the data to make the trend changes clearly visible. Finally, we obtained data for the SPTNs and SATNs.

The normalization formula is:

$${X}_{{Normalization}}=\frac{X-{X}_{\min }}{{X}_{\max }-{X}_{\min }}$$
(7)

where \({{\boldsymbol{X}}}_{{\boldsymbol{\min }}}\) and \({{\boldsymbol{X}}}_{{\boldsymbol{\max }}}\) represent the minimum and maximum values of each dataset during the study period (2000–2022), respectively.

We employed Pearson correlation analysis to conduct a cross-validation of the data from the Baidu Index (SPTNs) and Ctrip/Dianping (SATNs). The logic of cross-validation is that if data sequences from multiple independent sources with distinct bias characteristics maintain a high degree of consistency in trends, they can serve as a reasonable “proxy consistency validation” method in the absence of official visitor arrival data or tourism statistics41,42. The results showed that the two datasets were highly correlated between 2016 and 2022 (R = 0.805, p < 0.01). Moreover, the correlation coefficient for each year was >0.8, and all passed the significance test at the 0.01 level. To ensure temporal consistency between vegetation phenological indicators and tourism behavioral indicators, this study conducted a strict temporal alignment and processing of NDVI, SPTNs, and SATNs. First, we interpolated the original 16-day composite MODIS NDVI data into a daily time series using S-G filtering to obtain continuous daily NDVI data. Second, the SPTN (Baidu Index) and SATN (Ctrip/Dianping) data have daily temporal resolutions. Finally, we performed an alignment based on the Gregorian calendar date to ensure that the phenological indicators and tourism behavioral indicators have a strictly corresponding relationship at the same time points. This process provides a reliable basis for the subsequent time-lag analyses (Table 2). Among these years, the correlation was the highest in 2017 (R = 0.959), whereas it decreased to 0.807 in 2020 owing to the impact of the COVID-19 pandemic; nevertheless, a strong level of consistency was maintained. This indicates that despite differences in platform sources, the stability of these datasets in depicting trends in tourist behavior was relatively high, thereby supporting the use of SPTNs/SATNs as proxy indicators.

Table 2 Pearson correlation analysis results between SPTNs and SATNs
Full size table

To ensure temporal consistency between vegetation phenological indicators and tourism behavioral indicators, this study conducted a strict temporal alignment and processing of NDVI, SPTNs, and SATNs. First, we interpolated the original 16-day composite MODIS NDVI data into a daily time series using S-G filtering to obtain continuous daily NDVI data. Second, the SPTN (Baidu Index) and SATN (Ctrip/Dianping) data have daily temporal resolutions. Finally, we performed an alignment based on the Gregorian calendar date to ensure that the phenological indicators and tourism behavioral indicators have a strictly corresponding relationship at the same time points. This process provides a reliable basis for the subsequent time-lag analyses.

Correlation and dynamic time warping analysis

To quantify the relationship between the phenological periods of grasslands and tourism sightseeing periods, this study adopted Spearman correlation analysis and the DTW method. The Spearman correlation coefficient was used to measure the monotonic relationship between the NDVI and SPTNs and SATNs. Specifically, the NDVI time series was aligned with the SPTNs and SATNs time series to ensure that their time resolutions were consistent. Then, the “cor.test” package in the R software is used to calculate the Spearman correlation coefficient and evaluate the correlation between phenological changes and tourist behaviors.

The DTW method is used to assess the similarity between the NDVI, the SPTNs, and SATNs, and it can handle nonlinear time shifts between time series43,44. The specific steps include normalizing the NDVI and the SPTNs and SATNs to eliminate the differences in dimensions, and then using the “dtw” package in R to calculate the DTW distance between them. The smaller the DTW distance, the more similar the two time series.

Through these two methods, the temporal relationship between the phenological periods of grasslands and tourist sightseeing periods was quantified, providing a scientific basis for understanding the response mechanisms of tourist behaviors to phenological changes.

Exploration of factors influencing the spatiotemporal variation of grassland phenology

In this study, the Geodetector method was adopted to explore key factors influencing changes in grassland phenology. The Geodetector method revealed the degree of influence of different factors on SOS, EOS, and LGS, as well as their interaction effects, by quantifying the explanatory power of various environmental factors on phenological changes. The selected environmental factors included climate, topography, and human activity. Detailed descriptions and calculation methods for these factors are presented in Table 3.

Table 3 Factor selection and processing
Full size table

The Geodetector method is suitable for complex systems affected by multiple factors, and can effectively handle non-linear relationships and spatial heterogeneity. This method is based on the principle of spatially stratified heterogeneity and has unique advantages for dealing with categorical variables. The core hypothesis of this method is that if an independent variable has a significant influence on a dependent variable, the spatial distributions of the independent and dependent variables should be similar45. Geodetector can quantitatively explain spatial differentiation through factor and interaction detection. Its statistical framework is based on zonal variance analysis and does not rely on the assumption of residual independence required by traditional regression models, thus reducing the interference of spatial dependence on the results to a certain extent45,46. We used the “GD” package in the R software to execute Geodetector.

The factor detector in Geodetector uses the q value to measure the extent to which a given factor can explain the spatial heterogeneity by comparing the total variance of one factor46. The following formula was used:

$$q=1-\frac{{\sum }_{h=1}^{L}{N}_{h}{\sigma }_{h}^{2}}{N{\sigma }^{2}}$$
(8)

where \({\boldsymbol{q}}\) represents the explanatory power of a specific factor (q ∈ [0,1]), indicating that this factor explains (100 × q)% of the spatial heterogeneity of phenological indicators; \({\boldsymbol{h}}\) represents the layer of factor; \({\boldsymbol{N}}\) and \({{\boldsymbol{N}}}_{{\boldsymbol{h}}}\) represent the size of total sample units and the size of layer \({\boldsymbol{h}}\), respectively; and \({{\boldsymbol{\sigma }}}^{{\boldsymbol{2}}}\) and \({{\boldsymbol{\sigma }}}_{{\boldsymbol{h}}}^{{\boldsymbol{2}}}\) represent the total variance of the region and the variance in layer \({\boldsymbol{h}}\).

Results

Trend correlation between phenological and tourism periods of the grassland

As illustrated in Fig. 3a, the SATNs follow a unimodal seasonal trajectory that aligns closely with the NDVI peaks (June–September, NDVI > 0.5), indicating that tourists concentrate on visits during periods of maximum grassland greenness. The statistical analysis in Table 4 shows that before COVID-19 (from 2016 to 2019), the inter-annual correlation coefficient (Pearson’s R) between the SATNs and NDVI was consistently high, reaching 0.857 in 2016, 0.888 in 2017, 0.813 in 2018, and 0.917 in 2019 (all p < 0.01). This stability confirms that phenological signals strongly guided tourism flow during the pre-pandemic period. In 2020, R declined sharply to 0.669 (p < 0.01), reflecting the disruptions caused by COVID-19 restrictions and a shift toward short-distance travel. In contrast, 2021 showed an unusually high synchrony (R = 0.965, p < 0.01), likely reflecting a compensatory travel demand. In 2022, R (0.864, p < 0.01) returned to pre-pandemic levels, suggesting a re-established dependence on phenological cues. Lag phase analysis revealed shifting dynamics: during 2016–2019, tourists typically arrived 14–25 days after NDVI peaks, whereas in 2020–2021, lags turned negative (–15 to –32 days), indicating earlier visits driven by pandemic restrictions. The extreme –32 days in 2020 coincided with strict lockdowns. In 2022, the lag was shortened to 10 days, suggesting a partial recovery toward phenology-driven timing, but earlier arrivals were enabled by better information and infrastructure (Table 5).

Fig. 3: Trend relationship between the grassland phenology and tourism periods.
figure 3

a Trend relationship between grassland phenological period and SATNs; b relationship between grassland phenological period and SPTNs.

Full size image
Table 4 Trend correlation between the grassland phenological period and SATNs
Full size table
Table 5 Trend correlation between the grassland phenological period and SPTNs
Full size table

Figure 3b demonstrates a strong match between SPTNs and NDVI from 2011 to 2019 (R = 0.960–0.991, p < 0.001), confirming the intrinsic coupling between tourism demand and vegetation greenness, and the high sensitivity of web-based attention metrics. In 2020–2021, the correlations weakened (R = 0.657 and 0.773), reflecting shifts in online search behavior where safety concerns outweighed phenology-related interests.

Three distinct evolutionary phases emerged from lag analysis: the pre-pandemic era exhibited consistent 6–15 day lags in SPNT maxima relative to NDVI peaks, aligning with the established information processing theory (dissemination–decision–access). Pandemic conditions (2020–2021) reversed this relationship with negative Lag_POS values (−25 to −21 days) attributed to pandemic-induced shifts in search priorities. Post-2022 shows near-synchronization (2-day advance), signaling dual restoration mechanisms: (1) phenological driver recalibration in tourist decision-making and (2) technological accelerants in information diffusion pathways enhancing temporal matching precision.

Spatio-temporal trends of phenological indicators of grassland

Local Moran’s I was used to conduct spatial autocorrelation analysis of the multi-year average phenological indicators. The results revealed that all indicators exhibited significant spatial agglomeration characteristics (SOS Moran’s I = 0.772; EOS Moran’s I = 0.452; LGS Moran’s I = 0.709; all p < 0.05; Fig. 4). From the perspective of the LISA cluster map, the high-high and low-low clusters of SOS were mainly distributed on the southern and northern slopes of Narat Shan, respectively; those of LGS were primarily located on the northern and southern slopes of Narat Shan, respectively. Specifically, the northern slopes featured an earlier SOS and a longer growing season, whereas the southern slope exhibited the opposite pattern. These two phenological indicators display distinct spatial differentiation characteristics.

Fig. 4: Local Moran’s I and LISA cluster analysis of phenological indicators.
figure 4

a SOS; b EOS; c LGS.

Full size image

Figure 5 shows the pronounced regional heterogeneity of the three phenological parameters across the Narat Shan grassland system. Spatiotemporal analysis demonstrated latitudinal divergence in SOS dynamics (Fig. 5a): north-facing aspects exhibited progressive SOS advancement, with 7.01% showing extreme significance (p < 0.01) and 11.11% moderate significance (p < 0.05), forming continuous early SOS clusters along northern slopes and intermontane basins. Conversely, the southern slopes manifested SOS delays in 3.03% (p < 0.01) and 6.84% (p < 0.05) of the areas, establishing mirror-symmetric spatial patterns across the orographic divide.

Fig. 5: Spatial distribution of variation trend of phenological indicators in Narat Shan.
figure 5

a SOS; b EOS; c LGS (each dot represents a 0.05° × 0.05° pixel and color gradients indicating the direction and magnitude of phenological trends).

Full size image

EOS patterns (Fig. 5b) demonstrated axial complexity: northeast-central zones showed predominant delays covering 7.35% (p < 0.01) and 6.15% (p < 0.05), whereas flanking central regions exhibited localized advancement in 3.14% (p < 0.01) and 3.76% (p < 0.05) of the areas. This asymmetric distribution yields13.5% delayed versus 5.3% advanced coverage, indicating predominantly delayed EOS trends that critically influenced the duration of the growing season and ecosystem productivity.

LGS changes (Fig. 5c) revealed altitudinal differentiation; low-elevation sectors display significant LGS extension in 8.21% (p < 0.01) and 6.50% (p < 0.05) of the areas, in contrast to the high-altitude LGS reduction observed in 2.39% (p < 0.01) and 6.50% (p < 0.05). Net extension prevailed (14.71%, p < 0.05), with spatial heterogeneity underscoring topography-mediated phenological responses.

Influencing factors of grassland phenological changes

Geodetector analysis (Fig. 6a) revealed a clear regulatory hierarchy for SOS, with environmental factors exerting different explanatory strengths. Elevation (ELE) emerged as the dominant controller (q = 0.323, Table 6), which was stronger than the cumulative influence of all the other single factors. The PTR (q = 0.200) and PPT (q = 0.189) ranked second and third, respectively, and formed a climate-driven regulatory cluster. Critically, the ELE-PTR interaction achieves a much stronger explanatory power (q = 0.451), confirming strong topography–climate synergism, which is differentiated by altitude (Fig. 6a–c).

Fig. 6: Geodetector results of influencing factors of phenological indicators.
figure 6

a SOS; b EOS; c LGS. The radial bar chart shows the q value of the single-element Geodetector results, and the lower triangular heat map shows the q value of the interactive detection results (PTR pre-season temperature rate, PoTR post-season temperature rate, PPC pre-season precipitation change, PPT pre-season precipitation total, PoPC post-season precipitation change, PoPT post-season precipitation total, ELE elevation, SLO slope, ASP aspect, DD distance from road, DR distance from river).

Full size image
Table 6 Single-factor results of Geodetector
Full size table

Low-elevation belt (1000–2000 m): Persistent thermal advantages in northern Narat Shan sustain SOS advancement despite declining PTR trends (q = 0.39 for ELE-PTR interaction).

Mid-high elevation transition (2000–3000 m): Accelerated SOS advancement stems from PTR increases that counteract the high-altitude thermal constraints.

Southern highlands (>2000 m): Steep elevational gradients combined with PTR decrease drive the altitudinal amplification of SOS delays.

For EOS, altitude (ELE, q = 0.297) remained the strongest driver, higher than any other climatic factor, highlighting the topographic regulation of EOS (Table 6). The dominant climatic factors were PoTR (q = 0.287) and PoPT (q = 0.193). Interaction detection analysis showed that ELE–PoTR was particularly strong (q = 0.494), reflecting the indirect topographic control of EOS through temperature regulation (Fig. 6b). Specifically, the PoTR declined more steeply at higher altitudes (>2500 m), delaying the EOS. The lag effect was most evident at 1500–3000 m on the northern slope and 2500–3500 m on the southern slope (Fig. 7 d-f).

Fig. 7: Relationship between phenological indicators and key factors.
figure 7

a SOS_slope and ELE; b SOS_slope and PTR_slope; c PTR_slope and ELE; d EOS_slope and ELE; e EOS_slope and PoTR_slope; f PoTR_slope and ELE. The gray interval is the 95% confidence interval; N and S are the northern and southern parts of Narat Shan, respectively.

Full size image

The EOS is the dominant factor influencing the LSG, with q = 0.342, which is far stronger than that of the SOS (q = 0.124), confirming that the EOS delay is the key driver of the LGS extension (Table 6). In contrast, the single-factor q value of SOS is 0.12, indicating that the advancement of SOS contributes less to LSG. Among the climatic variables, PoPT (q = 0.170) and PoTR (q = 0.099) indirectly extended the LGS by delaying the EOS. Interactions amplify these effects; PoTR–EOS reaches q = 0.494, indicating that PoTR regulates LGS mainly through its impact on EOS (Fig. 6c).

Discussion

The high correlation between the phenological periods of grasslands in heritage sites and tourism visitation indicates that phenology is a key determinant of visitor behavior. The time-series analysis showed that the NDVI peak (maximum greenness) typically preceded the tourism peak, reflecting a lagged response to the optimal viewing period. Information flow, transportation, and market promotion influence this process. Media and online platforms amplify the grassland greenness, attract visitors, and produce a tourism peak lagging several weeks. This lag underscores the fact that seasonal ecosystem dynamics strongly shape tourist decision-making and are highly sensitive to ecological change.

The changes during the pandemic have further revealed tourism’s reliance on phenology and its vulnerability to external shocks47,48,49. In 2020 and 2021, Lag_POS shifted from lag to advance, likely due to restrictions (e.g., travel bans and crowd limits) and heightened safety concerns47,48,49. Studies have shown that such policies have reshaped planning, shortened travel windows, and prompted earlier trips49,50,51. However, limited policy data prevent causal inference; therefore, observed shifts are best viewed as associations. Despite weaker correlations during the COVID-19 pandemic, SATNs and SPTNs remained strongly linked to NDVI, reaffirming the central role of phenology. After the relaxation of pandemic restrictions, the recovery of Lag_POS indicated a partial return to phenology-driven tourism, although lasting behavioral shifts (e.g., avoiding crowds) likely advanced to 2022 peaks. This aligns with studies that highlight the long-term impact of health crises on travel preferences.

Across the pre-, during-, and post-COVID periods, grassland phenology consistently drove tourism behavior. Seasonal dynamics shape the overall visitation and the spatial–temporal distribution of peaks. This dependence illustrates the adaptation of tourism to natural cycles and highlights the central role of grassland ecology in regional tourism, providing a scientific basis for optimizing the management of grassland tourism resources (e.g., targeted monitoring during the main season, June–September).

Since the 1980s, Northwest China has experienced a warm-wet trend, profoundly affecting vegetation52,53,54. Pre-season temperature was a key driver of phenological changes, exhibiting distinct slope-related differences (Fig. 8). On the northern slope, a higher pre-season temperature primarily advanced SOS (dominant in the negative SOS trend areas), whereas on the southern slope, it delayed SOS, producing a mirrored effect. For EOS, increased preseason temperatures led to a delay in the northern slope (dominant in positive EOS trend areas); however, no significant impact was observed on the southern slope. Overall, regional warming altered mountain phenology, with effects significantly modulated by topography. Tourist sightseeing periods based on grassland phenology are also inevitably affected. The main tourist sightseeing areas in the Narat Shan region are Nalati, Kalajun, and Kuerdening. The phenological characteristics of the grasslands in these three areas were characterized by the advancement of the SOS and the delay of the EOS (Fig. 9). The extension rates of LGS in Kuerdening, Kalajun, and Nalati are 0.89, 0.44, and 0.86 day/year, with an average extension of 19.58, 9.68, and 21.52 days, respectively (Fig. 10). These changes indicate the sensitive response of the grassland ecosystem in this region to climate change, and are highly consistent with the broader climate trends in Central Asia and the Tianshan region in the context of global climate change. Such SOS advances and EOS delays are consistent with patterns across China and the Northern Hemisphere, underscoring the broader representativeness and environmental significance of our findings30,55.

Fig. 8: Relationship between variations in phenological indicators and key climatic indicators.
figure 8

a SOS and PTR on the northern slope; b SOS and PTR on the southern slope; c EOS and PoTR on the northern slope; d EOS and PoTR on the southern slope.

Full size image
Fig. 9: Changes in phenological indicators in the three tourism core areas.
figure 9

a The proportion of grassland SOS in three tourism core areas with significant change trends; b the proportion of different EOSs with significant change trends in the grassland in the three tourism core areas; c the proportion of grassland LGS with significant change trends in three tourism core areas.

Full size image
Fig. 10: Variation trend of regional LGS.
figure 10

a Kuerdening; b Kalajun; c Nalati.

Full size image

The positive correlation between the growing season and the best sightseeing period indicates that extending the growing season extends grassland sightseeing opportunities25, which may attract more tourists and promote the development of local tourism. However, it is crucial to consider the local communities and indigenous knowledge when developing tourism in this ecologically sensitive region. Indigenous people have accumulated rich experience in grassland management and utilization, offering valuable guidance for sustainable tourism. By integrating traditional knowledge with modern tourism practices (e.g., herder–manager co-produced “Best Grass-viewing Bulletin” that merges field green-up reports with NDVI updates to release a daily greenness index and recommended visiting route), tourism activities can be optimized to minimize ecological impacts and promote the preservation of local culture and biodiversity. Furthermore, involving the local community in tourism decision-making and revenue-sharing mechanisms strengthens stewardship and supports community-oriented conservation. However, increased tourist activity may also affect the regional grassland ecosystems. Therefore, establishing preventive mechanisms that combine community participation with indigenous knowledge and ecological monitoring is crucial. For instance, local communities can participate in eco-tourism initiatives by applying their indigenous knowledge of grassland management to optimize tourism activities and minimize ecological impacts.

Our study supplements and extends the existing literature in three main respects. First, it lays essential empirical and methodological groundwork for the future construction of a process-based “climate–phenology–recreation” framework. Although we do not claim to establish the full framework here, our long-term, empirical coupling of phenology and tourism provides necessary data, analytical procedures, and conceptual validation that are directly relevant to climate-adaptive management in World Natural Heritage sites. Second, this work advances ecological tourism scholarship and supports the development of “tourism phenology” by shifting the focus from visually dominant phenological events (e.g., flowering, autumn foliage) to grassland growing-season dynamics, and by introducing time-series tools and indicators suited to ecological-tourism coupling. Third, we present the first systematic quantification of synchrony and lag (Lag_POS) between grassland phenology (SOS, EOS, LGS) and tourism visitation in a high-mountain grassland system, using dynamic time warping to characterize nonlinear temporal alignments; this fills a notable gap in the literature and provides new analytical avenues for future research.

Although this study demonstrates the strong dependence of tourism behavior on grassland phenology, there are still some limitations. First, the review data from Ctrip and Dianping predominantly represent younger, high-income demographics and potentially underreport the behaviors of older or rural tourists. Future studies should integrate official tourism statistics, scenic area visitor logs, and regional datasets to triangulate the online indicators and improve their representativeness. Second, the temporal alignment between NDVI-derived phenology and behavioral indices requires verification using in-situ data (e.g., visitor entry records and ground phenological monitoring). Third, although large-scale land conversion was excluded, localized grazing and micro-scale land management could still confound the NDVI trends. However, given the strict conservation measures implemented after the World Natural Heritage designation (e.g., grazing prohibition), climate change remains the primary disturbance. Additionally, the absence of ground-truthing for phenological metrics introduces uncertainties in heterogeneous terrains because the 0.05° resolution of AVHRR may obscure species-specific responses. The 0.05° grid used here matches the NDVI and climate data resolution; however, aggregation at this scale may weaken the extreme values and fine-scale heterogeneity. Future research should employ multi-scale sensitivity analysis (e.g., 1, 5, 10 km) to test robustness and mitigate the modifiable areal unit problem. Similarly, although the dynamic threshold method has been validated against the GIMMS dataset, integrating multiple approaches (e.g., dynamic threshold and double logistic slope methods) and computing weighted or averaged estimates would further reduce uncertainties30.

Predictive modeling can enhance the adaptability of tourism management. Based on key drivers (altitude and temperature dynamics), machine learning methods such as Long Short-Term Memory networks can be used to build NDVI early-warning systems. For example, when greening rates exceed thresholds during the SOS, managers can launch tourism campaigns 10–15 days earlier, bridging the phenology–tourism response gap. Furthermore, a “phenological digital twin” integrating multi-scale monitoring data (e.g., Landsat-8 30 m, MODIS 250 m) could provide 10-day precision carrying-capacity alerts for high-frequency tourist zones. Strategically, climate-adaptive tourism planning should incorporate phenological projections for various SSP scenarios. For instance, simulating optimal sightseeing windows for 2040–2060 under SSP2-4.5 and SSP5-8.5 would enable heritage managers to pre-adjust infrastructure and visitor flows decades in advance. Such predictive tools can shift management from reactive to proactive, thereby building climate resilience at UNESCO sites under accelerating changes.