Introduction

Vegetation is part of a biome that covers the land surface and is an important component of the global ecosystem1. Vegetation phenology is defined as a natural phenomenon that occurs during annual cycles of plant growth2. Most plants undergo a series of growth stages, such as germination, leaf development, flowering, growth, fruiting, yellowing, and defoliation, during the annual cycle3. Because changes in vegetation phenology affect energy flow, the water balance and the carbon cycle in ecosystems4 vegetation phenology is considered an important parameter for monitoring ecosystem changes at different regional scales5.

Moreover, vegetation phenology is also regarded as an important indicator of the response to climate change because changes in vegetation phenology can visually reflect the impacts of climate change on vegetation and the biosphere6. Therefore, accurately analysing long-term temporal and spatial changes in vegetation phenology is highly important for the in-depth study of terrestrial ecosystems and climate change at regional and even global scales. The continuous improvement in remote sensing satellite observation technology has provided strong data and technical support for long-term dynamic monitoring of vegetation phenology, which is especially useful in geographically complex areas.

The QTP, as one of the most complex geographic environments in the world, is not only the world’s largest low-latitude plateau but also the world’s largest high-altitude permafrost-covered area7,8. Owing to the interactions among climate, altitude, soil conditions and topography, a unique and complex ecosystem with many vegetation types has formed7,9. It is known as the world’s “third pole”, “Asian water tower” and “roof of the world” and is regarded as the “amplifier” of global climate change4. Therefore, in the context of global warming, it is necessary to accurately grasp the long-term temporal and spatial patterns of vegetation phenology on the QTP. This study provides a reference and guidance for future ecosystem management, vegetation protection and analysis of the response of vegetation to climate change.

Spatial and temporal variations in vegetation phenology are usually analysed by SOS, EOS and LOS10. To analyse the long-term spatial and temporal variation patterns of vegetation phenology on the QTP accurately, it is necessary to precisely determine the SOS, EOS and LOS of the vegetation in the region each year. In studies of methods for extracting the SOS, EOS and LOS11,12,13 the following three steps were used: (a) satellite remote sensing data with different observation periods and resolutions were acquired to extract long-term NDVI data; (b) different smoothing algorithms were selected to reconstruct the long-term NDVI data to reduce the influence of clouds, atmosphere, soil, background noise, etc., on the original NDVI data; and (c) the reconstructed NDVI data combined with different mathematical model algorithms were used to extract the SOS, EOS, and LOS. Researchers have focused mainly on improving the second and third steps in research related to vegetation phenology, which has led to the construction of many smoothing reconstruction methods and a variety of extraction algorithms based on mathematical models.

Smoothing reconstruction methods mainly include asymmetric Gaussian14, Savitzky‒Golay filtering15, double-logistic14, changing-weight16, harmonic analysis of NDVI time series17, Whittaker smoother18, Fourier transform19, best exponential slope extraction20 and time window within linear interpolation methods21 among others. Each of these methods has its individual applications; among them, the Savitzky‒Golay filtering method, also known as numerical smoothing polynomials or least squares, can be used to reduce random noise in time series data. This is because the method does not involve edge points away from most points in the fitting process, and it is more effective in removing noise that deviates from the normal growth trend line22. Therefore, the Savitzky‒Golay filtering method has been widely used in the smooth reconstruction of remotely sensed vegetation indices such as the Normalized Difference Vegetation Index (NDVI)23 and this method has also been widely used in some previous studies related to vegetation on the QTP11,24.

The main extraction algorithms based on mathematical models include the maximum change slope method25, delayed moving average method26, curve fitting method27, double-logistic method28 and dynamic threshold method14,29. Among them, the maximum change slope method, which is based on the established standard NDVI time series data, calculates the EOS in the declining phase of the NDVI curve and, conversely, calculates the SOS in the rising phase of the NDVI curve. The delayed moving average method is mainly used to analyse the vegetation growth cycle. It helps to identify trends and changes in vegetation growth by calculating the mean value of the vegetation index and forming a curve that fluctuates over time. The curve fitting method is based on fitting a logistic model during the vegetation growth process, and the SOS and EOS are determined on the basis of the local maxima and minima in the rate of change in the curvature of the logistic model. The double-logistic method, which is based on the relationship between the NDVI and time, involves fitting two logistic curves to extract the SOS and EOS. The dynamic threshold method determines the SOS and EOS of each image on the basis of preset relevant reference thresholds. Although there are no extraction algorithms based on mathematical models that can be applied to all regions of the globe and all types of vegetation, the dynamic threshold method, because of its simplicity and effectiveness, has been successfully applied even in regions with extremely complex geography, such as the QTP29.

Although many reconstruction methods and mathematical algorithms, as described above, have been generated for performing the second and third steps in current research on the SOS, EOS and LOS, these methods and algorithms are almost all centred around NDVI data. Although the NDVI is very simple in terms of algorithms and processing, it is highly susceptible to interference from factors such as the soil background and atmosphere30 and is prone to misestimating vegetation cover when analysing the SOS and EOS in sparsely vegetated areas31. Currently, there are more than 150 types of vegetation indices in the field of remote sensing, which have different applications32 and some indices may be more suitable for extracting the SOS and EOS than the NDVI in specific situations. Therefore, this work focused on the first step of extracting the SOS, EOS and LOS and selected vegetation indices suitable for extracting the SOS and EOS to replace the NDVI with the goal of obtaining more accurate vegetation phenology data. The QTP in this study, for example, has a relatively high proportion of bare soil, sparse vegetation, grasslands and low vegetation overall33. The MSAVI is adjusted for soil effects, is sensitive to grasslands and early vegetation, and works well even in areas with low vegetation cover, especially in the initial period of the season when seedlings begin to grow34. Compared with the NDVI, the MSAVI is more suitable for extracting SOS information on the QTP region. In addition, because the EVI can better adapt to atmospheric and soil noise, reduce saturation and decrease the bias of results when the EOS is extracted35, the EVI was used instead of the NDVI to extract EOS information on the QTP region in this study.

In summary, this study is based on the Savitzky‒Golay filtering method and the dynamic threshold method. First, the SOS and EOS on the QTP were extracted from the experimental group via the MSAVI and EVI, respectively, instead of the commonly used NDVI index. To compare the results with those of the experimental group, the SOS, EOS, and LOS were also extracted via each of the vegetation indices (MSAVI, EVI, and NDVI, respectively). Next, we analysed the major land cover type changes on the QTP. The results were subsequently validated on the basis of these land cover changes. Finally, the annual trend of vegetation phenology on the QTP was mapped.

Methods

Study area

The QTP is bounded by the Pamir Plateau in the west and the Himalayas in the south, with a geographic range from longitude 73°18′52″–104°46′59″ E to latitude 26°00′12″–39°46′50″ N. The terrain is low in the southeast and high in the northwest7, with an average elevation of approximately 4000 m amsl8. Figure 1 shows the study area of this paper, which is based on MODIS global land cover MCD12Q1 data (Doi: https://doi.org/10.5067/MODIS/MCD12Q1.006 (accessed on 24 October 2024)) combined with the geographic latitude and longitude ranges of the QTP. We reclassified the QTP in relation to the actual land cover on the QTP, and the reclassified QTP contained 10 main land cover types: forests, scrublands, meadows, grasslands, permanent wetlands, croplands, urban and built-up areas, snow and ice, bare ground and deserts, and water bodies.

Fig. 1
figure 1

Geographic extent of the QTP in the study area and reclassified land cover types. The map was created using ArcGIS 10.8.1 (https://www.esri.com/en-us/arcgis/).

Full size image

Data sources

A total of three types of data were used in this study. The first is the MCD12Q1 data (Doi: https://doi.org/10.5067/MODIS/MCD12Q1.006 (accessed on 24 October 2024)), which is an SIN grid product available yearly with a spatial resolution of 500 m. This product is a land cover attribute based on observations from Aqua and Terra data on an annual basis, which contains a variety of land cover type scenarios. Its land cover type 1, the International Geosphere-Biosphere Programme (IGBP), is used in this study. The uses of these data in this study include the mapping of land cover in the study area (QTP), the analysis of the trends of major vegetation types from 2001 to 2020, and the validation of the spatial and temporal results of vegetation phenology. The second is MOD09A1 data (Doi: https://doi.org/10.5067/MODIS/MOD09A1.006 (accessed on 24 October 2024)), which is a MODIS surface reflectance product available every eight days at a spatial resolution of 500 m. Because these data are extracted on the basis of a low viewing angle, high observational coverage, and no clouds, each of its pixels contains the best possible L2G observations over an 8-day period. In this study, these data were used to extract three vegetation indices, the NDVI, MSAVI and EVI, for the period 2001–2020. The third is the day-by-day temperature data (Doi: https://doi.org/10.11888/Atmos.tpdc.272349 (accessed on October 24, 2024)). These data were provided by the National Tibetan Plateau Data Center and contain day-by-day maximum and minimum temperatures from more than 400 weather stations around the QTP. These data cover the period 1957–2020 with a spatial resolution of 0.25 degrees. In this study, the calculated mean air temperatures for the period 2001–2020 were used to identify image elements on the QTP that were affected by alpine snow.

Calculation of three vegetation indices, MSAVI, NDVI and EVI

The MSAVI was calculated as follows34:

$$MSAVI=\frac{1}{2}\left( {2NB+1 - \sqrt {{{\left( {2NB+1} \right)}^2} - 8\left( {NB - RB} \right)} } \right)$$
(1)

The NDVI was calculated as follows35:

$$NDVI=\frac{{NB - RB}}{{NB+RB~}}$$
(2)

where RB is the surface reflectance in the red band and NB is the surface reflectance in the near-infrared band.

The EVI was calculated as follows35:

$$EVI=GB\frac{{NB - RB}}{{NB+~C{B_1} \times RB - C{B_2} \times BB+LB}}$$
(3)

where BB is the surface reflectance in the blue band. LB has an initial value of 1 for adjusting the soil and canopy background, GB has a gain factor value of 2.5, and CB1 and CB2 are used as the two coefficients for correcting atmospheric aerosol scattering, with values of 6 and 7.5, respectively35.

Timing reconstruction and averaging for MSAVI, NDVI and EVI

The main steps of timing reconstruction and mean value calculation for MSAVI, NDVI and EVI are shown below: (a) The required MOD09A1 data tiles were preprocessed in ENVI 5.6 by stitching, reprojection, mosaicking, and batch clipping according to the vector boundaries of the QTP. Equations (1–3) were combined to calculate every 8-day time series of MSAVI, NDVI and EVI for the period 2001–2020. (b) The QTP is a high-altitude area with more snow in the mountains. This affects the accuracy of the three vegetation indices, MSAVI, NDVI and EVI, which in turn interferes with the extraction of vegetation phenology information. In this study, the daily air temperature was used to find the image elements affected by snow. Before the temperature data were used, the spatial resolutions of the temperature data and vegetation index data were first harmonised via MATLAB2022B. After that, the pixels with temperatures below freezing for more than five consecutive days were identified. Finally, the closest vegetation index values that were not affected by snow were used for replacement36,37. (c) Reconstruction of the MSAVI, NDVI and EVI time series data in the TIMESAT platform was achieved with the Savitzky‒Golay filtering method to remove noise that deviated from the normal growth trajectory. (d) The monthly average VA(M, N,E) of MSAVI, NDVI and EVI for 2001–2020 was calculated. (e) The six months of the year with the highest monthly mean values of MSAVI, NDVI, and EVI were set as the vegetation growing season on the QTP, and the annual mean values of MSAVI, NDVI, and EVI were calculated via the VA(M, N,E) for these six months of the year. The calculated annual mean values are the values of MSAVI, NDVI and EVI for each growing season (GV(M, N,E)) on the QTP.

Extraction of vegetation phenology information (SOS, EOS and LOS)

In this study, the dynamic threshold method14 was used to extract vegetation phenology information (e.g., the SOS, EOS and LOS) on the QTP from 2001 to 2020. The method extracts thresholds for the SOS and EOS by calculating thresholds for rising and falling periods of vegetation growth each year. The formulas for calculating the thresholds for the SOS and EOS are as follows:

$${T_{SOS\left( {M,N,E} \right)}}={V_{min\left( {M,N,E} \right)}}+\left( {{V_{max\left( {M,N,E} \right)}} - {V_{min\left( {M,N,E} \right)}}} \right) \times {q_1}$$
(4)
$${T_{EOS\left( {M,N,E} \right)}}={V_{max\left( {M,N,E} \right)}} - \left( {{V_{max\left( {M,N,E} \right)}} - {V_{min\left( {M,N,E} \right)}}} \right) \times {q_2}$$
(5)

where Vmin(M, N,E) is the minimum value before the growing season of MSAVI, NDVI and EVI each year. Vmax(M, N,E) is the maximum value of the MSAVI, NDVI and EVI each year during the growing season. q1 and q2 are the threshold coefficients for the rising and falling periods of vegetation growth, respectively. Symmetry of q1 and q2 is assumed in this study, and q1 = q2 = 0.20 is set14. TSOS(M, N,E) is the SOS threshold indicating the value of the vegetation index as it increases from Vmin(M, N,E) to the q1 proportional amplitude. TEOS(M, N,E) is the EOS threshold indicating the value of the vegetation index as it falls from Vmax(M, N,E) to the q2 proportional amplitude. In the rising period of each year, the first point in time that occurs such that Vt(M, N,E) ≥ TSOS(M, N,E) is the SOS of that year. Similarly, in the falling period of each year, the first point in time that occurs such that Vt(M, N,E) ≤ TSOS(M, N,E) is the EOS of that year. The difference between the EOS and SOS is the LOS for that year. For the extraction of vegetation phenology information, three comparison groups were added in addition to the experimental group. For the experimental group, MSAVI was used to extract the SOS, and EVI was used to extract the EOS. The SOS and EOS of the comparison groups were extracted by each vegetation index (MSAVI, EVI and NDVI) separately. The LOS in each group was determined by the difference between the EOS and SOS.

Combination of Theil–Sen trend analysis and the mann‒kendall test

In this study, three vegetation indices (MSAVI, NDVI, and EVI), vegetation phenology information (SOS, EOS, and LOS), and land cover changes in major vegetation were analysed for 2001–2020. These analyses were performed on the basis of Theil-Sen trend analysis. The Theil–Sen trend analysis and the Mann‒Kendall test are often combined and widely used in the field of vegetation9. The Theil–Sen trend utilises the median value of the computed series, which effectively reduces the influence of outliers, missing data and noise on the analysis results. The Mann‒Kendall test is often used to test the significance of a change in trend and is a nonparametric statistical test. The method not only does not require the data to follow a normal distribution but is also highly resistant to errors in the data. It is very suitable for analysing and evaluating long-term trend changes in type and series variables38. When the standardised test statistic ZT < 0 or ZT >0, the data series is trending downwards or upwards, respectively. Suppose that given a significance level β, a significant trend exists if | ZT | ≥ ZT (1 - β/2). When β = 0.05, the ZT statistic is 1.96; i.e., the 95% significance test is passed when the confidence level is 0.05 and the statistic reaches 1.96.

Correlation analysis via pearson’s method

The Pearson method is used to analyse the strength of correlation between two variables and is also known as Pearson correlation analysis39. Owing to the unavailability of long-term continuous vegetation phenology data on the QTP, the Pearson method was used in this study to analyse the correlation of the following points on the QTP during 2001–2020: (a) Two-by-two correlations between the three different vegetation indices, MSAVI, NDVI and EVI; (b) correlations between interannual rates of change in the land cover of major vegetation types and interannual rates of change in the three vegetation indices (MSAVI, NDVI and EVI); (c) correlations between the interannual rates of land cover change of major vegetation types and the interannual rates of change in vegetation phenology information (SOS, EOS and LOS).

Results

Time series changes in the MSAVI, NDVI and EVI from 2001 to 2020

Figure 2 shows the time series of the maximum, minimum and mean values of the MSAVI, NDVI and EVI for January 2001 to December 2020. All trend lines passed the significance test except for the minimum value of the NDVI (slope = 4.00 × 10<–6>, p > 0.05). The maximum and mean values of the MSAVI, NDVI, and EVI tended to increase. The minimum values of the MSAVI and EVI tended to decrease, whereas the minimum values of the NDVI tended to increase. The trends of the MSAVI and EVI were similar, whereas the trends of the NDVI were different. In terms of oscillation amplitude, the NDVI fluctuated between − 0.50 and 1, whereas the MSAVI and EVI fluctuated between − 0.30 and 0.90. Overall, the monthly mean values of MSAVI, NDVI and EVI in each year started to increase from April, peaked during August, and then started to decline, falling to lower levels during November, with the highest mean values occurring between May and October. Therefore, the growing season of the QTP was May–October.

Fig. 2
figure 2

Time series of MSAVI, NDVI, and EVI maximum, minimum, and mean values for the period 2001–2020. The horizontal axis table is a month‒year abbreviation, e.g., J-01 for January 2001 and D-20 for December 2020.

Full size image

Annual changes of MSAVI, NDVI and EVI in the growing seasons during 2001–2020

Figure 3 shows a bar chart of the annual growing season means of MSAVI, NDVI and EVI for 2001–2020. The trends of the GMSAVI, GNDVI, and GEVI passed the significance test. Over a period of 20 years, the GMSAVI, GNDVI and GEVI experienced several incremental and decremental changes but generally showed a small increasing trend. Among them, GMSAVI (slope = 3.00 × 10<–4>, p < 0.05) and GEVI (slope = 3.00 × 10<–4>, p < 0.05) were closer to each other in terms of changes in yearly means, whereas the GNDVI (slope = 1.00 × 10<–4>, p < 0.05) was somewhat different from both. In terms of oscillation amplitude, the GNDVI fluctuated between 0.25 and 0.29, whereas the GMSAVI and GEVI fluctuated between 0.14 and 0.18. For the minimum values, the GMSAVI, GNDVI and GEVI all occurred in 2017. For the maximum values, the GMSAVI and GEVI appeared in 2019, whereas the GNDVI appeared in 2009.

Fig. 3
figure 3

Annual mean changes in MSAVI, NDVI, and EVI during the growing season from 2001–2020.

Full size image

Annual changes and spatial patterns of the SOS, EOS and LOS from 2001 to 2020

Figure 4 shows the spatial distributions of the annual averages of the SOS, EOS and LOS 2001 to 2020. Among them, the SOS did not obviously change in the central region; the situation in the southeast, south and north regions was slightly advanced, whereas that in the southwest and northeast regions was slightly delayed, and the trend of the whole region fluctuated and then advanced slightly. The change in the EOS in the southeast region was relatively late, whereas that in the northwest region gradually advanced, the whole region showed a delayed tendency. The LOS in the southeast region was relatively long, whereas in the northwest region, it gradually became shorter, showing an overall increasing trend.

Fig. 4
figure 4

Spatial distributions of the annual mean values of the SOS, EOS and LOS from 2001–2020. The legends for the SOS and EOS indicate the specific month in which the SOS and EOS occurred, and the LOS is measured in days. Maps were created using ArcGIS 10.8.1 (https://www.esri.com/en-us/arcgis/).

Full size image

Figure 5 shows the annual growing season averages for the SOS, EOS, and LOS for the period 2001–2020. Figure 5 (a) shows the experimental group where the SOS was extracted via MSAVI and the EOS was extracted via EVI. Figures (b), (c) and (d) show the comparison groups where the SOS and EOS were extracted via MSAVI, EVI and NDVI, respectively. The mean values of the SOS, EOS and LOS in Figure (a) were 131.10, 266.64 and 135.54 days, respectively. The SOS (slope = − 0.11, p > 0.05) showed a slight advancing trend at 0.11 d/a, the EOS (slope = 0.22, p < 0.05) showed a significant delaying trend at 0.22 d/a, and the LOS (slope = 0.33, p < 0.05) showed a significant increasing trend, with an increase of 0.33 d/a. In particular, the maximum value of the LOS occurred in 2019 at 144.68 days, and the minimum value of the LOS occurred in 2017 at 126.47 days. Notably, the rate of change in the EOS latency was approximately twice as high as the rate of change in the SOS advance, implying that the EOS latency played an important role in the significant increase in the LOS. The mean values of the SOS, EOS and LOS in Figure (b) were 131.10, 267.11 and 136.01 days, respectively. The SOS (slope = − 0.11, p > 0.05) showed a slight advancing trend at 0.11 d/a; the EOS (slope = 0.23, p < 0.05) showed a significant delaying trend at 0.23 d/a; and the LOS (slope = 0.34, p < 0.05) showed a significant increasing trend at 0.34 d/a. The mean values of the SOS, EOS and LOS in Figure (c) were 133.54, 266.64 and 133.10 days, respectively. The SOS (slope = − 0.13, p > 0.05) showed a slight advancing trend at 0.13 d/a; the EOS (slope = 0.22, p < 0.05) showed a significant delaying trend at 0.22 d/a; and the LOS (slope = 0.35, p < 0.05) showed a significant increasing trend at 0.35 d/a. The results of Figures (b) and (c) are consistent with those of Figure (a), and the delay in the EOS plays an important role in the significant increase in the LOS. The mean values of the SOS, EOS and LOS in Figure (d) were 129.65, 270.37 and 140.72 days, respectively. The SOS (slope = − 0.31, p < 0.05) significantly advanced by 0.31 d/a; the EOS (slope = 0.10, p > 0.05) was slightly delayed by 0.10 d/a; and the LOS (slope = 0.40, p < 0.05) significantly increased by 0.40 d/a. Figure (d) differs from the results of the other three figures in that the advancement of the SOS plays an important role in the significant increase in the LOS.

Fig. 5
figure 5

Annual changes in the SOS, EOS, and LOS over the period 2001–2020. (a) The experimental group where the SOS was extracted via MSAVI and the EOS was extracted via EVI. (b), (c) and (d) The comparison groups where the SOS and EOS were extracted via the MSAVI, EVI and NDVI, respectively.

Full size image

Annual trends in major land cover types between 2001 and 2020

Figure 6 shows the annual proportional changes in the five main land cover types (forests, scrublands, meadows, grasslands, bare ground and deserts) on the QTP. Bare ground and deserts showed a decreasing trend, whereas the remaining four vegetation types showed an increasing trend. Among them, bare ground and deserts experienced an annual decrease of 6.85‱. The annual rates of increase of the four vegetation types were 2.57‱ (grasslands), 2.43‱ (meadows), 0.45‱ (scrublands) and 0.15‱ (forests). Among the four vegetation types, grasslands and meadows presented greater increases, whereas forests and scrublands were relatively stable. The total annual growth rate of the four vegetation types was 5.60‱. The four vegetation types accounted for an average of 59.40% of the total area covered by the QTP during the period 2001–2020, with bare land and deserts accounting for an average of 37.79%. The total number of pixels on the QTP was 2,609,761, and the number of pixels corresponding to bare land and deserts was approximately 986,228, with a decrease of 67,557 pixels per year, whereas the number of pixels corresponding to the four types of vegetation was approximately 1,550,266, with an increase of 86,815 pixels per year. The difference between the number of pixels increased by vegetation and the number of pixels decreased by bare ground and desert was 19,258. Only the interconversion between vegetation and bare ground and desert areas was considered. This difference represents 12.48% of the total number of converted pixels and 22.18% of the total number of pixels increased by vegetation. The percentages of these two sets of differences per year are referred to as D1 and D2, respectively.

Fig. 6
figure 6

Ten-thousandths ratios (‱) of annual changes in the five major land cover types on the QTP.

Full size image

Validation results based on correlation analysis

The two-by-two correlations between the trends of the three indices MSAVI, NDVI and EVI for 2001–2020, including the correlations between the mean, maximum and minimum values, were calculated. All three vegetation indices were highly correlated at the mean value; only MSAVI and EVI (R = 0.95) were highly correlated at the maximum value; and no correlation was found at the minimum value except for a moderate correlation between MSAVI and EVI (R = 0.79).

Table 1 shows the correlation between the interannual rates of change in the major land cover types and the interannual rates of change in MSAVI, NDVI and EVI for 2001–2020. In addition, Table 1 contains groups D1 and D2 from subsection 3.4. The correlation between the MSAVI or EVI and interannual rates of change in different land cover types is much greater than that between the NDVI and interannual rates of change in different land cover types. The correlations between MSAVI or EVI and grasslands, meadows and forests were 0.78, 0.82 and 0.75, respectively. In addition, the correlations between the MSAVI or EVI and D1 and D2 were moderate (0.60) and high (0.90), respectively, and the correlations between the NDVI and D1 and D2 were moderate (0.80) and low (0.45), respectively.

Table 1 Correlation between the interannual rates of change in the major land cover types and the interannual rates of change in MSAVI, NDVI and EVI from 2001–2020.
Full size table

Table 2 shows the correlations between the interannual rates of change in major land cover types and the interannual rates of change in the SOS, EOS and LOS for 2001–2020. The SOS and EOS were extracted via the MSAVI, EVI and NDVI, respectively. LOS (a) corresponds to the LOS extracted by the experimental group in Fig. 5, and LOS (b), (c), and (d) correspond to the LOS extracted by the three comparison groups in Fig. 5. In the following, descriptions and comparisons of the correlation results are presented in absolute terms.

Table 2 Correlation between the interannual rates of change in major land cover types and the interannual rates of change in the SOS, EOS and LOS from 2001–2020.
Full size table

The correlation of the SOS (MSAVI) with different land cover types revealed that the SOS (MSAVI) had the highest correlation, 0.73, with forests, and the others had a correlation of less than 0.50. The correlations of the SOS (MSAVI) with D1 and D2 were high (0.88) and low (0.50), respectively. Unlike the SOS (MSAVI), most of the correlations between the EOS (MSAVI) and different land cover types passed the significance test and were relatively high. Among them, the correlations of the EOS (MSAVI) with grasslands and meadows were 0.85 and 0.90, respectively. The correlations of the EOS (MSAVI) with D1 and D2 were moderate (0.54) and high (0.96), respectively. These findings are similar to the results of the EVI group. The correlation of the SOS (EVI) with different land cover types revealed that the SOS (EVI) had the highest correlation, 0.72, with forests, and the others had a correlation less than 0.50. The correlations between the SOS (EVI) and D1 and D2 were high (0.84) and low (0.46), respectively. The correlations between the EOS (EVI) and different land cover types mostly passed the significance test, and the correlations were relatively high. The correlations of the EOS (EVI) with grasslands and meadows were 0.86 and 0.91, respectively. The correlations of the EOS (EVI) with D1 and D2 were moderate (0.57) and high (0.99), respectively. That is, during the process of vegetation cover growth, the SOS (MSAVI/EVI) did not advance in most areas of the QTP, whereas the EOS (MSAVI/EVI) was delayed.

Unlike the results of the MSAVI and EVI groups, the correlation of the EOS (NDVI) with different land cover types revealed that the EOS (NDVI) had the highest correlation, 0.64, with forests, and the other correlations were less than 0.50. The correlations of the EOS (NDVI) with D1 and D2 were moderate (0.77) and weak (0.44), respectively. The correlations of the SOS (NDVI) with different land cover types mostly passed the significance test and had relatively high correlations. The correlations of the SOS (NDVI) with grasslands and meadows were 0.84 and 0.79, respectively. The correlations of the SOS (NDVI) with D1 and D2 were low (0.41) and moderate (0.72), respectively. In other words, the SOS (NDVI) of most areas of the QTP advanced, whereas the EOS (NDVI) was not delayed during the growth of vegetation cover.

Aspects of the correlation between the LOS and different land cover types in the four groups were considered. Overall, the correlation of LOS (a) with different vegetation types was greater in the experimental group than in the three comparison groups. The correlations between the LOS and D2 of all four groups passed the significance test and were 0.67, 065, 0.63, and 0.55, respectively. That is, in terms of the accuracy of extracting the LOS of the QTP, MSAVI in combination with EVI was the best, followed by MSAVI, EVI, and finally NDVI.

Discussion

The results in Fig. 2 show that the mean and maximum values of the MSAVI, NDVI and EVI fluctuated and increased slightly between 2001 and 2020. This may be due to some changes in temperature and precipitation patterns on the QTP, which positively affected the growth of vegetation, resulting in an increase in vegetation cover40. Although the mean and maximum values of the NDVI also increased, the mean and maximum values of the NDVI were significantly greater than those of the MSAVI and EVI. This is because the atmospheric and soil backgrounds have not been adequately corrected in the calculation of the NDVI30. Even though the NDVI is reconstructed with smoothing, it is still affected by significant changes in the atmospheric or soil background, resulting in amplification of the vegetation signal. The MSAVI and EVI take these disturbing factors into account and improve the related algorithms, so they can recognise vegetation information more accurately34,35. In addition, the NDVI fluctuates greatly when the vegetation changes slightly, and the minimum value contains more noise interference. The MSAVI and EVI have relatively smooth responses to these small changes.

From the results in Fig. 4, the southeast, southwest, and south of the QTP show a slight advance in the SOS and a significant delay in the EOS. The southeastern and southern parts of the plateau are covered mainly by forests and meadows, which have tended to be warmer and wetter in recent years41 and the vegetation can meet the temperature and moisture conditions required for growth earlier in the year. The southwestern part of the plateau is covered mainly by grasses, which are relatively flexible in their response to temperature42; the SOS appears earlier, and the EOS arrives later in warming climatic scenarios. In the northeast and north, both the SOS and EOS were slightly delayed, with little overall change in the LOS. The northeast and north regions have drier climates with limited precipitation43. Although temperatures are increasing in these regions, the total cumulative temperature in spring and fall is relatively stable. Overall, the spatial heterogeneity of vegetation phenology on the QTP and the complexity of hydrothermal scenarios require more in-depth investigation.

From the results of the experimental group in Fig. 5a, the mean values of the SOS, EOS and LOS on the QTP during 2001–2020 in this study were 131.10 days, 266.64 days and 135.54 days, respectively. During this 20-year period, the SOS slightly advanced by approximately 0.11 d/a, the EOS was significantly delayed by approximately 0.22 d/a, and the LOS was significantly increased by approximately 0.33 d/a. These results are different from the results of a previous study in which vegetation climatic data were extracted from the QTP via the NDVI method. For example, Yuan et al. reported mean values of the SOS, EOS, and LOS of 128, 276, and 148 days, respectively, from 2000 to 2019, and the trends of the SOS, EOS, and LOS significantly advanced, slightly delayed, and significantly increased, respectively44. Huang obtained mean values of the SOS, EOS, and LOS of 129, 266, and 137 days, respectively, from 2001 to 2018, and the trends of the SOS, EOS, and LOS significantly advanced, did not significantly change, and significantly increased, respectively45. Feng et al. reported mean values of the SOS, EOS and LOS of 122, 282 and 160 days, respectively, from 2001 to 2020, and the trends of the SOS, EOS and LOS significantly advanced, with no significant change and a significant increase, respectively11. Chang et al. reported mean values of the SOS, EOS, and LOS of 123, 289, and 166 days, respectively, from 2003 to 2013, and the trends of the SOS, EOS, and LOS advanced, were complicated, and increased, respectively46. Among these sets of results, Huang produced results closest to those of the experimental group in this study. However, the years 2019 and 2020 were not analysed in Huang’s study. As shown in Fig. 2, the NDVI values for these two years were in the high range between 2001 and 2020. This means that the LOS values obtained by Huang could have been larger if he had also studied the years 2001–2020. This is verified in the NDVI-based comparison group in Fig. 5d.

To analyse the results of the experimental group in greater detail, several additional comparisons were performed. These comparisons were performed on the LOS results of different study regions near the QTP. The LOS growth rate in this study was 0.33 d/a, which was lower than that of 1.03 d/a in the Three-River Headwaters area47 higher than that of 0.19 d/a on the western Sichuan plateau48 and close to that of 0.36 d/a in the Qilian Mountains49. Although the Three-River Headwaters area and the Qilian Mountains are both part of the QTP, the former has a higher LOS growth rate. This is because the Qilian Mountains are an ecological region dominated by mountainous landscapes49, whereas the Three-River Headwaters area is a typical wetland ecological region where many rivers originate47. Compared with mountainous ecoregions, wetland ecoregions have more favourable hydrothermal conditions for vegetation growth. Therefore, the increasing rate of the LOS in the Three-River Headwaters area was relatively high. The QTP as a whole is characterised by mountainous terrain, and the area of mountainous terrain is much greater than the area of wetlands. This study was conducted on the QTP as a whole, so the growth rate of the LOS was closer to that of the Qilian Mountains.

A comparison with the results of previous studies revealed essentially the same trend of a significant increase in the LOS. However, the difference is mainly in two aspects. First, the SOS values obtained from previous studies44,45,46 are relatively small, whereas the EOS values are relatively large, which means that the magnitude and range of the increase in LOS in previous studies are relatively large. Second, the trend of the SOS reported in previous studies44,45,46 was a significant advanced, whereas the trend of the EOS appeared to be more complicated, which means that in previous studies44,45,46 the increase in the LOS resulted from the fact that the SOS played a dominant role. However, in the experimental group of this study, the increase in the LOS was smaller in magnitude and scope than it was in previous studies, and the increase in the LOS in this study was mainly dominated by the EOS, whereas the SOS was only slightly advanced, and the change was relatively smooth. The main reason for this result is that the SOS and EOS were analysed via MSAVI and EVI in the experimental group, whereas previous studies44,45,46 analysed the SOS and EOS via the NDVI. The MSAVI and EVI are relatively stable and do not tend to amplify vegetation signals or background noise34,35 so the increase in and range of the LOS obtained in this study are relatively small. This was also verified in the comparison groups based on MSAVI and EVI in Fig. 5b and c.

From the point of view of indirect argumentation and in conjunction with Fig. 2, the results obtained in Fig. 5a may be more closely related to vegetation phenology changes on the QTP. First, in the selection of vegetation indices, the MSAVI and EVI are clearly superior to the traditional NDVI in terms of principle and applicability34,35. Second, the correlation between the MSAVI and EVI in terms of the mean, maximum, and minimum values is indeed much better than that of the NDVI. In addition, we found that there was no correlation between the performance of the NDVI in terms of minimum values and the MSAVI and EVI, which explains why it is difficult for the NDVI to perform its identification function in areas of low vegetation cover, and many areas on the QTP are covered by sparse vegetation and low vegetation33. There is a large bias in the SOS results obtained on the basis of the NDVI in sparsely vegetated and snow-covered areas50. In addition, the spectral mixing effect in sparsely vegetated and snow-covered areas leads to delayed EOS results obtained on the basis of the NDVI51. This may be the main reason for the early identification of the SOS and late identification of the EOS in previous NDVI-based studies11,44,45,46 as shown in Fig. 5d.

In addition to indirect argumentation, the reliability of the experimental results should be further verified from the perspective of direct argumentation. The best way for direct validation is to use the measured data of vegetation phenology on the QTP from 2001 to 2020 for comparison. Owing to the geographical constraints of the QTP, field stations are scarce, and it is difficult to obtain enough measured data for validation. However, most of the vegetation phenology data in the literature are based on NDVI extraction44,45,46,47,48,49. Considering the above factors, we directly validated the results of this experiment through changes in vegetation cover. Table 1 shows the correlations between the three vegetation indices and the interannual rates of change of different land cover types, and the results of MSAVI or EVI are much better than those of NDVI. The most covered grasslands on the QTP are taken as an example. The correlation coefficient between the interannual rates of change of MSAVI (or EVI) and those of grasslands was 0.78, whereas the correlation coefficient between the interannual rates of change of NDVI and grasslands was only 0.39. These findings indicate that the MSAVI and EVI can more effectively reflect the changes in vegetation on the QTP. Furthermore, the extracted vegetation phenology information in the comparison groups (b) and (c) in Fig. 5 is close enough to that in the experimental group (a) in Fig. 5. Both of these findings support the conclusion that the EOS plays a dominant role in vegetation phenology changes on the QTP. In addition, from the results in Table 2, the EOS plays a dominant role in the change in vegetation phenology when vegetation phenology information is extracted on the basis of MSAVI and EVI. When vegetation phenology information is extracted on the basis of the NDVI, the SOS plays a dominant role in the change in vegetation phenology. However, the EOS (MSAVI/EVI) was greater than the SOS (NDVI) in terms of correlations with different vegetation types and D2. Moreover, the correlation between the LOS (a) and D2 was greater than that between the LOS (d) and D2. This finding also indicates that the method of extracting vegetation phenology via MSAVI combined with the EVI is more applicable to the QTP than the NDVI is.

In summary, the reliability of the experimental results of this study was demonstrated through indirect and direct validation. However, due to various limitations in time and space, some limitations remain in this study. First, due to the long observation period, the vegetation index changed relatively slowly over short periods of time. In this study, MOD09A1 data provided every 8 days were used. However, the vegetation index during the growing season is not precise enough when only measured every 8 days. Second, the QTP is a vast area with obvious spatial heterogeneity, and this study did not conduct a detailed zoning study. Finally, this study did not analyse in depth the specific effects of changes in the hydrothermal environment and human activities on vegetation phenology. For these shortcomings, we will gradually improve upon them in subsequent studies.