Comparison of Phenological Parameters Extracted from SIF, NDVI and NIRv Data on the Mongolian Plateau

Abstract

The phenological parameters estimated from different data may vary, especially in response to climatic factors. Therefore, we estimated the start of the growing season (SOS) and the end of the growing season (EOS) based on sunlight-induced chlorophyll fluorescence (SIF), the normalized difference vegetation index (NDVI) and the near-infrared reflectance of vegetation (NIRv). The SIF, NDVI and NIRv breakpoints were detected, and the trends and change-points of phenological parameters based on these data were analyzed. The correlations between the phenological parameters and snow-related factors, precipitation, temperature, soil moisture and population density were also analyzed. The results showed that SIF and NIRv could identify breakpoints early. SIF could estimate the latest SOS and the earliest EOS. NDVI could estimate the earliest SOS and the latest EOS. The change-points of SOS_SIF were mostly concentrated from 2001 to 2003, and those of SOS_NDVI and SOS_NIRv occurred later. The change-points of EOS_SIF and EOS_NIRv were mostly concentrated from 2001 to 2007, and those of EOS_SIF occurred later. Differently from the weak correlation with SOS_SIF, SOS_NDVI and SOS_NIRv were significantly correlated with snow-related factors. The correlation between the meteorological factors in the summer and autumn and EOS_SIF was the most significant. The population density showed the highest degree of interpretation for SOS_NIRv and EOS_NDVI. The results reveal the differences and potentials of different remote-sensing parameters in estimating phenological indicators, which is helpful for better understanding the dynamic changes in phenology and the response to changes in various influencing factors.

1. Introduction

Phenology is a sensitive and key feature of the life history of plants. It can reflect how changes in climate affect vegetation growth, improve our understanding of vegetation ecological processes and allow us to quantify the impacts of meteorological changes on terrestrial ecosystems [1].

The Mongolian Plateau, located in arid and semiarid areas, is highly sensitive to climate change [2]. Its temperate continental climate forms large natural pastures [3]. Additionally, the grassland ecosystem of the Mongolian Plateau plays key roles in the carbon cycle in East Asia and globally [4]. The wide availability of low-cost remote sensing data facilitates the study of vegetation ecology on the Mongolian Plateau [5,6]. Previous studies have shown that the normalized difference vegetation index (NDVI) can be used to effectively monitor vegetation phenology [7]. Therefore, NDVI has frequently been used to estimate vegetation phenological parameters in studies of Mongolian Plateau phenology [1,2,8,9]. However, limitations may exist to using only NDVI when retrieving vegetation phenology information for evergreen forests [10,11]. Thus, evaluating the advantages and disadvantages of different types of remote sensing data in retrieving vegetation phenology and using a combination of approaches to obtain detailed observations are very important for high-precision phenology research [12].

In contrast to traditional remote sensing indicators based on changes in vegetation “greenness”, sunlight-induced chlorophyll fluorescence (SIF) can directly reflect the photochemical reactions of plants [13,14]. Additionally, SIF is very sensitive to abiotic stress factors and can be effectively used to monitor vegetation stress [15,16]. SIF has unique technical advantages in the field of monitoring vegetation physiological conditions and changes in the surrounding environment [17,18,19]. Furthermore, it represents a new potential method for monitoring the physiological dynamics of vegetation from space [20].

Specifically, SIF is the energy flux released by plant chlorophyll molecules after absorbing sunlight in the wavelength range of 600–800 nm, which represents less than 5% of the light energy absorbed by plants [21]. Many studies have investigated the SIF products of the Greenhouse Gas Observation Satellite (GOSAT) and Global Ozone Monitoring Experiment-2 (GOME-2), but GOSAT and GOME-2 (GOME-2: 40 × 80 km; GOSAT: 10 km) cannot meet the need for high-accuracy research [22,23]. The SIF products estimated by the Orbital Carbon Observatory-2 (OCO-2) provide a better spatial resolution (1.3 × 2.2 km) than the SIF products of GOSAT and GOME-2 [24]. However, the OCO-2 SIF products with discontinuous OCO-2 still have limitations. Fortunately, Li and Xiao developed a new SIF product based on OCO-2 SIF and meteorological data, called “GOSIF”, with a temporal resolution of 8 days and a spatial resolution of 0.05° [25]. This product has a better temporal and spatial resolution than existing products, and can greatly contribute to research using SIF to monitor vegetation physiology. Previous studies have shown the feasibility of GOSIF for retrieving vegetation phenology [26,27]. However, in arid and semiarid areas of the Mongolian Plateau, the characteristics of vegetation phenology estimated by SIF are not clear. At the same time, whether or not the high sensitivity of SIF to environmental stress has a higher advantage in terms of vegetation phenology also needs further exploration.

A recently developed vegetation index, the near-infrared reflectance of vegetation (NIRv), has the advantages of reducing the number of mixed pixels and minimizing the noise caused by soil pollution [28,29,30]. Previous studies have shown that NIRv can be used to monitor early changes in canopy development in grassland ecosystems. Thus, NIRv, as an index that can effectively reflect the growth status of vegetation, may also have great potential in retrieving vegetation phenology. Simultaneously, the advantages and disadvantages of sensitivity to climate factors are not clear.

Studies on the climatic factors affecting vegetation phenology have considered the effects of temperature, precipitation and soil moisture [31,32,33,34]. However, the effects of snow cover in winter and the snow melting period in spring at the start of the growing season (SOS) cannot be ignored. In addition, the phenological parameters of vegetation based on SIF (GOSIF), NDVI and NIRv and their responses to different climatic factors are also different. Understanding the response of phenological parameters estimated from different data to different climatic factors can enable us to more deeply understand the mechanism by which vegetation is affected by the environment. Concurrently, comparing the sensitivity of phenological parameters to influencing factors can help us to understand the advantages and disadvantages of the phenological retrieval abilities of different types of remote sensing data. The advantage of using different types of phenological parameters is of great value for retrieving phenological parameters with higher accuracy and improving our understanding of the relationship between vegetation phenology and climate factors in the future. The main objectives of this study were as follows:

• To compare the characteristics of vegetation dynamics estimated by SIF, NDVI and NIRv;
• To extract and compare phenological parameters and their dynamics based on SIF, NDVI and NIRv;
• To determine the influence mechanism of snow-related factors on phenological parameters.

2. Materials and Methods

2.1. Study Area

The Mongolian Plateau is located in Central Asia (87°43′–126°04′E, 37°22′–53°23′N) [35], and it is surrounded by high-altitude mountains in the north and is adjacent to the Gobi Desert in the south [36,37]. Affected by the climate, the annual precipitation on the Mongolian Plateau is low, and drought is frequent. The average annual precipitation in most areas is 200 mm, and precipitation mainly occurs from June to August. Windy conditions are prevalent in winter and spring. The average temperature in January is −26 °C, and in July it is 17 °C [38]. The plateau has complex ecological zones, which mainly include meadow steppe, typical steppe, alpine steppe, desert steppe, coniferous forest, broadleaf forest and the Gobi desert (Figure 1). The vegetation-type data used to produce the figure are from the National Atlas of Mongolia. The vegetation map with a scale of 1:1,000,000 was rasterized to 0.083° [39]. Finally, we chose GOSIF, NDVI and NIRv to estimate the vegetation phenological parameters for the Mongolian Plateau.

2.2. Data Sources

2.2.1. Sunlight-Induced Chlorophyll Exposure

Li and Xiao developed a new global SIF data set with high spatiotemporal resolution (0.05° × 0.05°) based on OCO-2, using discrete OCO-2 SIF detection, remote sensing data from the moderate-resolution imaging spectroradiometer (MODIS) and meteorological reanalysis data. Compared with coarse-resolution SIF data aggregated directly from OCO-2 sounding, sunlight-induced chlorophyll exposure (GOSIF) has a higher spatial resolution and longer records, and it provides continuous global coverage [25].

The website used to access GOSIF data is as follows: http://globalecology.unh.edu/data/GOSIF.html (accessed on 3 April 2022).

In this study, we used GOSIF data sets from 2001 to 2020 with a spatial resolution of 0.05° × 0.05° and a temporal resolution of 8 d. To match the temporal resolution of NDVI and NIRv, we synthesized the GOSIF data of 8 days into a data set with a temporal resolution of 16 days, using the maximum-value composite (MVC) method.

2.2.2. MODIS Data Set

The vegetation indices we used included NDVI and near-infrared reflectance (NIR), both of which came from the MOD13A1 product of Terra MODIS version 6.1. The spatial resolution was 500 m, and the temporal resolution was 16 d. The website used to access these data is as follows: https://lpdaacsvc.cr.usgs.gov/appeears/ (accessed on 3 April 2022). NDVI and NIRv are calculated as follows:

$$\mathrm{NDVI}=\frac{{\rho }_2-{\rho }_1}{{\rho }_2+{\rho }_1}$$

(1)

$$\mathrm{NIRv}=\left(\mathrm{NDVI}-C\right)\times {\rho }_2$$

(2)

where ρ₁ is the red reflectance (620–670 nm), ρ₂ is the NIR (841–876 nm), and the parameter C was set to 0.08, according to Badgley [29]. In this study, we used NDVI and NIRv data sets from 2001 to 2020. We resampled the spatial resolution of the NDVI and NIRv data (the source data are 500 m) to 0.05° × 0.05°, to match the spatial resolution of GOSIF. (Downsampling of all high-resolution data in this study is realized through the aggregate tool of ArcGIS Pro3.0 software).

We used the daily snow-cover product of MOD10A1 from 2000 to 2020, with a spatial resolution of 500 m. The winter (December of the previous year and January and February of the current year) snow-cover data set (SFC) from 2001 to 2020 was calculated. In addition, we defined the snowmelt date as the period when the snow cover in spring had no value for five consecutive days, and we used Interactive Data Language (IDL, an application program and programming language for data analysis and visualization) to calculate the item-by-item snow-melting-period data set (SMD) from 2001 to 2020. To match the spatial resolution of the phenological data set, SFC and SMD were resampled to 0.05° × 0.05°. The data were acquired from https://www.environment.snu.ac.kr/bess-rad (accessed on 3 April 2022).

2.2.3. Extraction of Vegetation Phenological Parameters

Before extracting the phenological parameters, we discarded the areas with an annual average NDVI < 0.1, because such areas have been identified as unvegetated [40]. Additionally, we masked GOSIF and NIRv with NDVI of the discarded unvegetated area as the template to obtain the GOSIF and NIRv data sets of the discarded unvegetated area. Then, we used double logistic filtering to fit the smooth curve from the time series of the GOSIF, NDVI and NIRv data [41]. The threshold method is usually used to extract phenological vegetation parameters, and it mainly includes the fixed-threshold method and dynamic-threshold method [42,43]. However, the fixed-threshold method has limitations in areas with different land-cover types and soil backgrounds. Therefore, we used a more flexible dynamic-threshold method to extract the phenological parameters for the start of the growing season (SOS), the end of the growing season (EOS) and the length of the growing season (LOS, i.e., LOS = EOS − SOS). The dynamic threshold we selected was 20% [44,45,46]. For example, SOS was defined as the first day of the year that the GOSIF ratio exceeded 0.2 [47]. The GOSIF ratio is calculated as follows:

$${\mathrm{GOSIF}}_{ratio}=({\mathrm{GOSIF}}_t-{\mathrm{GOSIF}}_{min})/({\mathrm{GOSIF}}_{max}-{\mathrm{GOSIF}}_{min})$$

(3)

where GOSIF_t is the value of GOSIF at time t and GOSIF_min and GOSIF_max are the minimum and maximum GOSIF values, respectively, in the annual GOSIF time series. The extraction of the above parameters was completed using TIMESAT 3.3 software. The software download website is http://web.nateko.lu.se/timesat/ (accessed on 4 April 2022).

In this study, we used SIF (GOSIF), NDVI and NIRv to extract phenological vegetation information, and obtained the following:

• Phenological parameters based on SIF: SOS_SIF, LOS_SIF and EOS_SIF;
• Phenological parameters based on NDVI: SOS_NDVI, LOS_NDVI and EOS_NDVI;
• Phenological parameters based on NIRv: SOS_NIRv, LOS_NIRv and EOS_NIRv.

2.2.4. ERA5-Land Climate Data Set

ERA5-Land is a reanalysis data set that has a higher resolution than ERA5 and provides conventional longitude and latitude grid data with a spatial resolution of 0.1° × 0.1° (approximately 10 km). We obtained the data set of the average monthly values from 2001 to 2020, and the data were divided into the three types of climate data, as explained below.

We selected the total precipitation in m; the skin temperature, k; and the volume of the soil water layer 1 (0–7 cm) m³/m⁻³). The meteorological data sets of spring (March to May, PRE_Spring, TEM_Spring, and SW_Spring), summer (June to August, PRE_Summer, TEM_Summer, and SW_summer) and autumn (September to November, PRE_Autumn, TEM_Autumn, and SW_Autumn) were calculated. To match the spatial resolution of the phenological data, the data were resampled to 0.05° × 0.05°. The website providing detailed descriptions and access to the above ERA-Land climate data is as follows: https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land-monthly-means?tab=overview (accessed on 3 April 2022).

2.2.5. Population Density Data

The population density data were obtained from open-space population data and previous research. We obtained the annual population-density data set from 2001 to 2020, with a spatial resolution of 1 km and a final resampling to 0.05° × 0.05°. The data were obtained from https://hub.worldpop.org/geodata/listing?id=77 (accessed on 3 April 2022).

2.2.6. Terrain Data Set

The terrain data were obtained from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM) V1 (https://www.gscloud.cn). Through an image mosaic, we obtained the DEM of the study area with a spatial resolution of 30 m (accessed on 3 April 2022).

2.3. Methods

2.3.1. The BFAST Package

The Breaks for Additive Season and Trend algorithm (BFAST) method was first proposed by Verbesselt in 2010 [48], and was initially used to identify disturbances in time-series models (with trend/season/regressor terms), using remote sensing data. The BFAST algorithm, which is widely applicable for meteorological, hydrological, economic and other fields, employs the parameters of fitted piecewise-linear trends and a seasonal trend model [49]. The formula of BFAST for long time-series is defined as follows:

Y_t = T_t + S_t + e_t, t = 1, 2, 3……n

(4)

T_t = α_i + β_it

(5)

M = (α_i−1 − α_i) + (β_i−1 − β_i)t

(6)

$$S_t={\displaystyle \sum }_{k-1}^j{\alpha }_{j,k} sin\left(\frac{2\pi kt}{f}\right)+{\delta }_{j,k}$$

(7)

where t is the time step of a long time-series; Y_t is the 16-d SIF from 2001–2020 on the Mongolian Plateau, T_t is the trend component (trend), S_t is the seasonal component (seasonal), and e_t is the remainder (remainder). The long-term trend component, T_t, is segmented, has a specific slope (β_i) and intercepts (α_i) on m + 1 different segments. Thus, there are m breakpoints, τ_i−1, …, τ_m, where i = 1, 2, …… m, and we define τ₀ = 0 and τ_m + 1 = n.

The order of magnitude of the breakpoints, M, can be calculated using the slopes α_i and intercepts β_i of the trend component T_t between t_i−1 and t_i. To fit the frequency, we define t₀ = 0 for t_j < t ≤ t_j + 1, where j is the breakpoint position, j = 1, … m, with m breakpoints in total; k is the number of harmonic terms; and α_j, k and δ_j,k are the segment-specific amplitude and phase, respectively. In this study, the SIF, NDVI and NIRv values during the growing season (from April to October) on the Mongolian Plateau from 2001 to 2020 were used for breakpoint analysis, and the differences in the vegetation-change trends retrieved from the three types of remote sensing data were evaluated. We also analyzed meadow-steppe, typical-steppe, broadleaf-forest and coniferous-forest data. The main purpose of this part of the study is to evaluate the sensitivity of three types of remote sensing data for retrieving the dynamic changes in vegetation and the ability to capture the dynamic changes in different vegetation types.

2.3.2. Theil–Sen Median Trend Analysis

Theil–Sen median-trend analysis and the Mann–Kendall analysis are effective for assessing the trends of long time-series data. The combination of these methods has the advantages of not requiring the data to obey a certain distribution, and having strong resistance to data errors [50]. Additionally, the Mann–Kendall method provides a significance test of linear trends. The Theil–Sen estimator calculates the median of the SOS data-combination slope as follows:

$$S_{SOS}=Median\left(\frac{SO{S}_i-SO{S}_j}{j-i}\right), 2001\le i<j\le 2020$$

(8)

S_SOS > 0 indicates an increasing SOS trend, while a negative value indicates a decreasing SOS trend.

The Mann–Kendall test can determine whether a trend is significant. Similar to the Theil–Sen median trend, this test does not require the sample to obey a certain distribution, and is robust to the presence of a few outliers [51,52].

$$Z=\{\begin{array}{c}\frac{S-1}{\sqrt{s\left(S\right)}}, S>0\\ 0, S>0\\ \frac{S+1}{\sqrt{s\left(S\right)}}, S>0\end{array}, S={{\displaystyle \sum }}_{j=1}^{n-1}{{\displaystyle \sum }}_{i=j+1}^{n}sgn\left(SO{S}_j-SO{S}_i\right)$$

(9)

$$sgn\left(SO{S}_j-SO{S}_i\right)=\{\begin{array}{c}1,SO{S}_j-SO{S}_i>0\\ 0,SO{S}_j-SO{S}_i=0\\ -1,SO{S}_j-SO{S}_i<0\end{array},s\left(S\right)=\frac{n\left(n-1\right)\left(2n+5\right)}{18}$$

(10)

where SOS_i and SOS_j represent the SOS in years i and j, respectively; sgn is the signum function; n represents the length of the time series, and the value range of the statistic Z is (−∞, +∞). At a given significance level α, when |Z| > u_1−α/2, the changes in the trend are considered significant. We evaluated the significance of a change in the SOS time-series trend at the 0.05 confidence level; thus, we applied α = 0.05 [53]. In this study, we analyzed the trends of SOS_SIF, EOS_SIF, SOS_NDVI, EOS_NDVI, SOS_NIRv, and EOS_NIRv and the phenological parameters of different vegetation-types from 2001 to 2020. At the same time, we compared and evaluated the differences in the trends of the three phenological parameters.

2.3.3. Detection and Analysis of Change-Points of Phenological Parameters

We used the change-point detection tool in ArcGIS Pro3.0 to detect the change points of the three phenological parameters. The change-point detection tool can recognize the time step when some statistical attributes of a time-series change. The tool can detect changes in the average value, standard deviation or slope (linear trend) of continuous variables, and changes in the average value of the counting variables. In this study, we used the slope change of the phenological parameters over 20 years to detect the change points. To prevent all time steps from being detected as change points, one of two constraint types must be applied, using method parameters. Algorithms for the optimal identification of segment neighborhoods (SegNeigh) enable the user to determine the number of change points to be detected, using the number of change-points parameters This option uses the “segment neighborhood” [54] algorithm to find the lowest cost segment among all possible segments with the specified number of changepoints. The optimal detection of changepoints with a linear-computational-cost (PELT) option uses the pruned exact linear Ttme [55,56] algorithm to estimate the number and location of change points.

The idea behind PELT is that for a time step to be detected as a change point, the value of its segmentation cost reduction must be more than the added penalty value. If the cost reduction is less than the increased penalty value, the penalty cost will increase, and the time step will not be detected as a change point. The penalty value used in PELT is determined by the detection-sensitivity-parameter value. The sensitivity is provided in digital form between 0 and 1, where a higher sensitivity will detect more change points by using lower penalty values. For the change in slope (linear trend), a more conservative penalty formula is used:

$$\mathrm{Penalty} \mathrm{Value}=6^{2-2·sensitivity}$$

(11)

In this study, we used PELT to detect change points. Because change-point detection based on slope (step size 2) requires conservative control-sensitivity, we finally defined the sensitivity as 0.35. This sensitivity detects one change-point at most.

2.3.4. Pearson Correlation Analysis and Partial Correlation Analysis

The Pearson’s correlation coefficient method can be used to analyze the degree of relatedness between two variables. By calculating the correlation coefficient, R, between the climate factors and SOS and EOS, we determined the relationships between them. The range of potential values of the correlation coefficient is [−1, 1], and the calculation formula is as follows:

$$R_{SOS,SFC}=\frac{{{\displaystyle \sum }}_{i=1}^n\left(SO{S}_i-\overline{SOS}\right)\left(SF{C}_i-\overline{SFC}\right)}{\sqrt{{{\displaystyle \sum }}_{i=1}^n{\left(SO{S}_i-\overline{SOS}\right)}^2}\sqrt{{{\displaystyle \sum }}_{i=1}^n{\left(SF{C}_i-\overline{SFC}\right)}^2}}$$

(12)

where n is the number of years, n = 20 (2001–2020); i and j represent years i and j, respectively; $\overline{\mathrm{SOS}}$ represents the average SOS value over the 20 years from 2001 to 2020, and R_SOS,SFC are the correlation coefficients of SOS and SFC. In addition, we used a t-test to test whether the correlation coefficient was significant. A p value < 0.05 indicated a significant correlation.

Partial correlation analysis: In a multifactorial system, the relationship between two factors is inevitably disturbed by other factors. Therefore, excluding the influence of other factors is necessary to obtain a relatively pure relationship between the two variables. Suppose there are k (k > 2) variables x₁, x₂…x_k; then, after excluding the influence of other variables, the correlation coefficient of variables i and j is called the hth-order (h ≤ k − 2) partial correlation coefficient, which is calculated as follows:

$$R_{ij,l_1·l_2\dots l_h}=\frac{R_{ij,l_1·l_2\dots l_{h-1}}-R_{i{l}_h,l_1·l_2\dots l_{h-1}}·R_{j{l}_h,l_1·l_2\dots l_{h-1}}}{\sqrt{(1-R_{i{l}_h,l_1·l_2\dots l_{h-1}}^2) (1-R_{jh,l_1·l_2\dots l_{h-1}}^2)}}$$

(13)

where $R_{ij,l_1{l}_2\dots l_h}$ is the partial correlation coefficient of variables i and j, after excluding the influence of other factors. Suppose that there are five variables (k = 4): SOS(S), TEM_Spring(T), PRE_Spring (P) and SW_Spring (W). Consequently, the order of the partial correlation coefficient of the variable SOS and TEM is h = 2, and the formula is as follows:

$$R_{ST,PW}=\frac{R_{ST,P}-R_{SW,P}·R_{TW,P}}{\sqrt{(1-R_{SW,P}^2) (1-R_{TW,P}^2)}}$$

(14)

In this study, we conducted a Pearson correlation analysis of three SOS values by SCF and SMD, and then we conducted a partial correlation analysis of spring precipitation, temperature and soil water and three SOS values, and a partial correlation analysis of summer and autumn precipitation, temperature and soil water, on three EOS values.

2.3.5. Path Analysis

Path analysis is an extension of multiple regression analysis. The causal relationship in path analysis is multileveled, and intermediate variables are added between causal variables, which makes the path-analysis model richer and more powerful than the general regression model in describing the actual causal relationship. Thus, path analysis is not a multiple regression equation, but rather a set of multiple regression equations. For example, the equations are as follows:

$$\mathrm{SOS}={ \mathsf{\beta }}_1·SFC+{\mathsf{\beta }}_2·TEM+{\mathsf{\beta }}_3·SW+{\mathsf{\beta }}_4·SMD+{\epsilon }_1$$

(15)

$$\mathrm{SMD}={ \mathsf{\beta }}_5·SFC+{\mathsf{\beta }}_6·TEM+{\epsilon }_2$$

(16)

$$\mathrm{SW}={ \mathsf{\beta }}_7·SFC+{\mathsf{\beta }}_8·TEM+{\mathsf{\beta }}_9·SMD+{\epsilon }_3$$

(17)

where ${\mathsf{\beta }}_1$…${\mathsf{\beta }}_9$ are path coefficients, and ${\epsilon }_1$…${\epsilon }_3$ are residuals. The main purpose of path analysis is to test the accuracy and reliability of a hypothetical causal model and to measure the strength of the causal relationship between variables. If a correlation exists between the two variables X and Y in the model, further study is required to determine if a causal relationship exists between them. This may be a direct or indirect relationship, or both. In this study, we assumed that a direct or indirect relationship exists between SMD and SOS. Path analysis can be used to more comprehensively understand the mechanism by which snow removal affects the SOS. For example, SCF and SMD affect SOS directly or through some variables, or some variables affect SOS through SMD. All path designs and path coefficients were calculated using IBM SPSS Amos 25 software.

2.3.6. Geographical-Detector Model

The geographical-detector model can analyze the relationship between factors based on spatial heterogeneity. In this study, the dependent variable Y was the phenological parameter of the Mongolian Plateau, and the independent variable X included natural factors and human factors. The geographical-detector model consists of four detectors: a factor detector, an interaction detector, a risk detector and an ecological detector [57]. In this study, we mainly used the factor detector to determine the explanatory degree of the independent variable to the dependent variable. The scope of interpretation is measured by the metric q [58], which is expressed as follows:

$$q=1-\frac{{{\displaystyle \sum }}_{h=1}^L{N}_h{\sigma }_h^2}{N{\sigma }^2}=1-\frac{SSW}{SST}, SSW={{\displaystyle \sum }}_{h=1}^L{N}_h{\sigma }_h^2, SST=N{\sigma }^2$$

(18)

where q is the degree of interpretation of factor X to Y, and the value range of q is [0, 1]. h = 1…, L represents the grade of the independent variable Y or the dependent variable X; ${\sigma }_h^2$ and ${\sigma }^2$ are the variance of Y in layer h and the variance in the entire region, respectively; N_h and N represent the number of layers, h, and the number of cells in the entire area, respectively.

3. Results

3.1. Comparison of Vegetation Trends Based on SIF, NDVI and NIRv

SIF, NDVI and NIRv all have the ability to restore vegetation changes, but their results may vary for vegetation dynamics. To explore the differences in the vegetation dynamics retrieved from the three types of remote-sensing data on the Mongolian Plateau, we used the BFAST package to analyze the SIF, NDVI and NIRv trends, and to explore the breakpoints. Additionally, we analyzed the trends and breakpoints in the meadow steppe, typical steppe, broadleaf forest and coniferous forest, over a large area of the Mongolian Plateau (Figure 2). The results show that the SIF, NDVI and NIRv trends and breakpoints on the Mongolian Plateau were obviously different.

SIF breakpoints were found on day 225 in 2003, day 145 in 2007, day 225 in 2008 and day 209 in 2014. Moreover, SIF decreased significantly from 2003 to 2007, increased sharply from 2007 to 2008, and increased significantly from 2008 to 2014 and from 2014 to 2020 (Figure 2a). NDVI breakpoints were found on day 161 in 2009, day 305 in 2012, day 193 in 2014 and day 177 in 2018 (Figure 2b). NDVI experienced a continuous surge from 2009, 2012 and 2012 to 2014 and a decline from 2018 to 2020. However, no NIRv breakpoint was detected on the Mongolian Plateau. The comparison showed that the SIF and NDVI breakpoint years overlapped greatly (2007 to 2014), showing significant growth, and the SIF breakpoint years appeared earlier.

In addition, our detection results for different vegetation types were very different. SIF of meadow steppe experienced a sharp increase from 2006 to 2007 and a sharp decrease from 2013 to 2014. Both meadow steppe and forest experienced a sharp increase or a decrease from 2002 to 2007, but NDVI did not detect these changes. The drastic changes in NDVI mostly occurred from 2008 to 2011. During this period, SIF also underwent drastic changes. Therefore, SIF may capture more mutation details than NDVI, especially for mutations within a short period of time that may be caused by human activities. NIRv and SIF were similar, especially in forest areas. Both could detect breakpoints in earlier years, but the NIRv breakpoints occurred later than those for SIF. This result may indicate that SIF and NIRv have more advantages in detecting short-term upheaval.

3.2. Temporal and Spatial Variability of Vegetation Phenological Parameters

In the past 20 years, the SOS_NDVI and SOS_NIRv values of the Mongolian Plateau have been relatively close, and have not increased significantly, with rates of 0.143 days/year and 0.192 days/year, respectively (Figure 3a). The average values over the past 20 years were 103.6 day and 97.4 day, respectively. The earliest SOS occurred in 2016 and 2013, and the latest SOS occurred in 2006 and 2008. In the past 20 years, SOS_SIF was late, with an insignificant increase of 0.038 days/year, and the annual average value was the 132.3nd day. The earliest and latest SOS occurred in 2001 and 2005, respectively, indicating that this difference may be due to the low chlorophyll concentration and weak SIF of the returning vegetation at higher latitudes on the Mongolian Plateau. Concurrently, the earliest and latest years of SOS are breakpoint years.

The EOS_NDVI and EOS_NIRv values on the Mongolian Plateau are also close, and EOS_NDVI is slightly later than EOS_NIRv. EOS_NDVI and EOS_NIRv decreased at the rates of 0.387 days/year and 0.173 days/year, respectively, but not significantly, with mean values of the 303.9nd day and the 297nd day over 20 years. The earliest and latest SOS were in 2001 and 2019, respectively. However, EOS_SIF increased at a rate of 0.145 days/year, but not significantly. The average value from the past 20 years was 274nd day, and the earliest and latest EOSs were in 2001 and 2006, respectively. LOS_SIF was the shortest, and LOS_NDVI was the longest. The three LOSs were 112.3 days, 198.6 days and 195.6 days, respectively.

As shown in Figure 4, the spatial distributions of the three SOS trends in central and eastern Mongolia were similar. Central and eastern Mongolia showed an upward trend, while western Mongolia and central and eastern Inner Mongolia showed a downward trend. However, SOS_SIF had a higher significance; 10.29% of pixels passed the significance test (p < 0.05), and the agricultural land in south-central Inner Mongolia and the broadleaf forest in the northeast showed a significant downward trend (Figure 4a).

Obvious differences were observed among the three EOSs. Approximately 58.57% of the EOS_SIF pixels showed an increasing trend that was mainly distributed in the middle of Mongolia and the middle of Inner Mongolia. Approximately 10.38% of the pixels passed the significance test (p < 0.05), and these pixels were distributed in the coniferous forest belt in the middle of Mongolia and on agricultural land and desert steppe in the middle and south of Inner Mongolia (Figure 4d). The growth of EOS_NDVI and EOS_NIRv in this region was also detected, but it was not significant. Approximately 69.05% of the EOS_NDVI pixels showed a decreasing trend, and these pixels, of which 13.71% passed the significance test (p < 0.05), were mainly distributed in the typical steppe belt in the southeast and the meadow steppe belt in the northeast of Mongolia, as well as some coniferous and broadleaf forest-areas in Inner Mongolia. The spatial distribution trend of EOS_NIRv was similar to that of EOS_NDVI, but the decreasing trend of the coniferous and broadleaf forest belts detected by EOS_NIRv was not significant (Figure 4f). However, EOS_SIF did not detect significant changes in the forest region. In general, the spatial distribution of the three EOS trends on the Mongolian Plateau was generally consistent, and the difference was mainly reflected in the sensitivity of the trend to different vegetation types.

Figure 5 shows that SOS_SIF tends to be significantly delayed by approximately 2 days/100 m of elevation. The difference is that SOS_NDVI and SOS_NIRv significantly advance at rates of 0.6 days/100 m and 1.3 days/100 m, respectively. EOS_SIF and EOS_NDVI advanced by approximately 0.9 days/100 m and 0.04 days/100 m, respectively, although the trend for EOS_NDVI was not significant. EOS_NIRv was significantly delayed by approximately 0.5 days/100 m. As the altitude increased, the change trends of the three phenological parameters varied. Therefore, as the altitude increased, LOS_SIF was significantly shorter, LOS_NDVI slowly increased, and LOS_NIRv was significantly longer.

This may be due to the difference in chlorophyll concentration of different vegetation types, which affects the observation of phenology by NDVI and NIRv, especially in high altitude areas, where the greenness reflectance of forest vegetation is high. SIF can directly express the photosynthesis inflection-point, and may better show the relationship between vegetation phenology and altitude, which also highlights the advantages of SIF in retrieving phenological parameters.

We analyzed the occurrence of the change trend of the vegetation phenological parameters on the Mongolian Plateau over the past 20 years (Figure 6 and Figure 7). The results show that the years that SOS_SIF experienced the most changes were 2001 to 2003 and 2016 to 2019. A total of 43.55% and 24.41% of the changes occurred in these two periods, respectively. The years with the least change were from 2011 to 2013, and only 2.24% of the change occurred during this period. The change points were mainly distributed in the broadleaf forest area in northeastern Inner Mongolia, the desert steppe belt in southern Mongolia, and the coniferous forest areas in northern and middle Mongolia (Figure 6a). The decreasing trends of SOS_SIF before and after the change accounted for 70.95% and 48.28%, respectively. The increasing trends before and after the change accounted for 21.97% and 51.65%, respectively (Figure 6b,c). This result shows that the change was mainly reflected in the change from the original decreasing trend to the increasing trend, especially in the central and western regions of Mongolia and some regions in southern Inner Mongolia. In northeastern Inner Mongolia, the original increasing trend changed to a decreasing trend.

The years that SOS_NDVI experienced the most changes were 2004 to 2007 and 2016 to 2019. A total of 25.02% and 24.91% of the changes occurred in these two periods. The years with the least change were from 2011 to 2013, and only 4.54% of the changes occurred then (Figure 6d). The change points were mainly distributed in the coniferous forest area in northern Mongolia and in central and southern Inner Mongolia. The decreasing trend of SOS_NDVI before and after the change accounted for 48.82% and 72.07%, respectively. The proportion of the increasing trend before and after the change was 53.55% and 27.58%, respectively (Figure 6e,f). This result shows that the original increasing trend changed into a decreasing trend, which was mainly distributed in the coniferous forest in northern Mongolia and in the sandy steppe and agricultural land in central Inner Mongolia.

The years that SOS_NIRv experienced the most changes were 2017 to 2019 and 2004 to 2007. A total of 24.46% and 24.25% of the changes occurred in these two periods. The years with the least change were from 2011 to 2013, and only 5.46% of the change occurred then (Figure 6g). The change years of SOS_SIF and SOS_NIRv were relatively close. The distribution was closer to SOS_NDVI, and more pixels were detected. The decreasing trend of SOS_NIRv before and after the trend-change accounted for 53.17% and 64.46%, respectively. The increasing trend before and after the change accounted for 46.64% and 35.49%, respectively (Figure 6h,i). This result shows that the change was mainly manifested as the original increasing trend changing to a decreasing trend. The change points were mainly distributed in the coniferous forest in northern Mongolia and most areas of Inner Mongolia. The number of change-point pixels detected in the forest area of northeast Inner Mongolia was greater than that of SOS_NDVI. In general, the change years of the three SOSs were similar to the BFSAT breakpoint-monitoring years for the three data sources. In terms of the years and the distribution, SOS_NDVI and SOS_NIRv were closer, and the changes in NDVI occurred later than the changes in SIF and NIRv. SOS_SIF was more sensitive to the change in the forest area, and SOS_NDVI was more sensitive to the change in the steppe area. SOS_NIRv detected the most change points.

In the past 20 years, the years that EOS_SIF experienced the most changes were 2004 to 2007 and 2001 to 2003. A total of 31.08% and 29.22% of the changes occurred in these two periods. The years with the least change were from 2014 to 2016, with only 5.47% of changes. The trend change-points were mainly distributed in the grassland belt in eastern Mongolia and the desert steppe in southern Inner Mongolia (Figure 7a). The decreasing trend of EOS_SIF before and after the change point accounted for 58.85% and 57.35%, respectively. The proportion of the increasing trend before and after the trend change was 41.52% and 51.66%, respectively (Figure 7b,c). This result shows that the change was mainly reflected in the change from the original decreasing trend to the increasing trend, especially in the central and western regions of Mongolia and in some regions in southern Inner Mongolia. In northeastern Inner Mongolia, the original increasing trend changed to a decreasing trend.

The years in which EOS_NDVI experienced the most change were from 2017 to 2019, and 44.51% of the changes occurred during this period. The years with the least change were from 2011 to 2013, with only 5.39% of the change (Figure 7d). This change mainly occurred in the coniferous forest belt in the north of Mongolia and the artificial vegetation area in the middle of Inner Mongolia. The decreasing trend of EOS_NDVI before and after the change accounted for 46.18% and 37.47%, respectively. The proportion of the increasing trend before and after the change was 48.56% and 59.26%, respectively (Figure 7e,f). This result shows that the overall trend is from decreasing to increasing. The trends mainly changed in some areas in the middle of Mongolia, and the trends in the sandy steppe and agricultural land in the middle of Inner Mongolia changed from decreasing to increasing. The slope of the decreasing trend in northeastern Inner Mongolia also changed.

The years that EOS_NIRv experienced the most changes were 2001 to 2003 and 2016 to 2019, and 39.41% and 22.1% of the changes occurred in these two periods, respectively. The years with the least change in trends were from 2014 to 2016, and only 3.38% of the change occurred during this period (Figure 7g). The year and spatial distribution of the EOS_NIRv change points were close to those of EOS_NDVI, and EOS_NIRv detected more pixels. The decreasing trend of EOS_NIRv before and after the change accounted for 63.61% and 47.55%, respectively. The increasing trend before and after the change accounted for 36.33% and 52.41%, respectively (Figure 7h,i). This result shows that the change in the trend was mainly manifested in the change from the original decreasing trend to the increasing trend. The change points were mainly distributed in the coniferous forest in northern Mongolia and in central and southern Inner Mongolia. In addition, the slope of the decreasing trend in the middle of Inner Mongolia changed. In terms of the year and location, EOS_SIF and EOS_NIRv were closer. The change in EOS_NDVI occurred later than that of EOS_NDVI and EOS_NIRv. EOS_NIRv detected the most change points.

3.3. Correlation Analysis between Vegetation Phenological Parameters and Meteorological Factors

The area of snow-cover in winter showed a significant negative correlation with SOS_NDVI and SOS_NIRv (Figure 8a–c). Approximately 41.56% and 34.87% of the pixels passed the significance test (p < 0.05), and were mainly distributed in most regions of Mongolia and the central region of Inner Mongolia. This indicated that the larger snow-cover area in winter will lead to the advance of the SOS. This is related to the supplementation of snow to soil moisture in spring. However, SOS_SIF was not sensitive to the change in the snow-cover area (Figure 8a). This result shows that snow-cover in winter had a significant effect on the advance of the SOS. The spring snow-melting period showed a more significant negative correlation with SOS_NDVI and SOS_NIRv (Figure 8d–f). Approximately 54.48% and 47.65% of the pixels, respectively, passed the significance test (p < 0.05), and were mainly distributed in most regions of Mongolia and parts of the middle and north of Inner Mongolia. Specifically, the late snowmelt-period may lead to the advance of the SOS; in contrast, if the snowmelt period is relatively early, it will lead to the delay of the SOS. However, SOS_SIF was still not sensitive to changes in the snowmelt period (Figure 8a).Therefore, the larger snow-cover in winter and late snow-melting in spring positively contributed to the advance of the SOS. This result shows that snow-melting may play an important role in supplementing soil water and in the supplementation time for the growth of vegetation in spring. However, the mechanism of the snow-cover and the snowmelt period on the SOS is complex and indirect, especially when the soil water is supplemented by melted-snow cover, which plays an important role here; thus, further analysis is needed.

To eliminate the interference between the meteorological factors in spring, we conducted a partial correlation analysis on the precipitation, temperature and soil moisture of the three SOSs (Figure 9). SOS_SIF, SOS_NDVI and SOS_NIRv all showed a positive correlation with PRE_Spring, and 5.26%, 6.09% and 6.69% of the pixels, respectively, passed the significance test (p < 0.05). Consequently, compared with the other two SOSs, SOS_NIRv and PRE were more significant. SOS_SIF, SOS_NDVI and SOS_NIRv showed a negative correlation with TEM_Spring. Approximately 11.83%, 12.8% and 8.07%, respectively, of the pixels passed the significance test (p < 0.05); thus, SOS_NDVI is highly significant, and mainly distributed in the forest areas in the northeast of Inner Mongolia and the west of Mongolia. SOS_SIF, SOS_NDVI and SOS_NIRv all showed a significant negative correlation with SW_Spring. Approximately 20.41%, 24.71% and 17.52%, respectively, of the pixels passed the significance test (p < 0.05) and were mainly distributed in the northern coniferous forest of Mongolia, the eastern meadow steppe and the northeast broadleaf forest of Inner Mongolia. After comparison, the significance of SOS_NDVI was the highest. The results show that the effect of spring precipitation on the SOS was relatively small, and due to the decreased precipitation in spring. In addition, the early precipitation in spring may lead to low temperatures, thus delaying the SOS; thus, precipitation and the SOS were positively correlated. In the meantime, the water source for vegetation in spring is mostly from the soil water formed after the melting of snow or frozen soil. Therefore, the temperature and soil moisture in spring showed a high negative correlation with the SOS.

Due to the complex impact of water and heat synergy on the vegetation in spring, we used the path-analysis method to analyze the direct or indirect effect of the snow-cover area in winter and snow-melting date on SOS under the synergistic effect of spring temperature and soil moisture (the precipitation in spring is small, and the influence is low; thus, no analysis was made). SOS_SIF is not sensitive to snow-cover and snowmelt periods. Therefore, in this study, we only analyzed SOS_NDVI and SOS_NIRv (Figure 10). The results show that a large snow area will delay the snow-melting date. The temperature in spring has a high negative correlation with snow-melting date and soil moisture. Consequently, an increase in temperature may lead to earlier snowmelt. Simultaneously, the advance of the snow-melting date means that the time of soil water supplementation is also advanced. The temperature rise will further lead to the evaporation of soil water and thus be consumed. Hence, the later the snow-melting date is, the higher the amount of soil moisture preserved, which plays an important role in advancing the vegetation rejuvenation period. However, the earlier melting of snow will lead to an increase in evaporation time, and the water eventually stored in the soil will be insufficient to supply vegetation. Moreover, according to the path coefficient, SW_Spring has little effect on the SOS, which indicates that the soil moisture content is low during the spring drought. In this period, soil moisture is more important to the supply time of vegetation, which depends on the content of snow and the influence of temperature on snow-melting date. In other words, even if snow-melting supplies much soil water, because the snow-melting date is too early, the soil water is not sufficient for vegetation growth but rather is consumed by evaporation, which means that higher spring soil water may not have the effect of advancing the SOS. This also shows that the path coefficients of SMD and SOS are higher than those of SW_Spring. The response of SOS_NDVI and SOS_NIRv to snow-melting conforms to this mechanism.

On the Mongolian Plateau, the summer precipitation is positively correlated with the three EOSs, and the spatial distribution is similar; however, EOS_SIF is more significant (Figure 11a). A total of 21.19% of the pixels passed the significance test (p < 0.05), and they were mainly distributed in the coniferous forest area in the central part of Mongolia, the eastern border area and the central area of Inner Mongolia. The correlation between the EOS and summer temperature was not significant. The positive correlation areas are mainly distributed in Mongolia and northeastern Inner Mongolia; this may be related to the low temperature at high latitudes (Figure 11b,e,h). The three EOSs showed a negative correlation with soil water in summer, with a similar spatial distribution, and EOS_SIF was more significant. A total of 16.44% of the pixels passed the significance test, and they were mainly distributed in southern Inner Mongolia, the meadow steppe belt at the border between central Inner Mongolia and Mongolia, and some forest areas in northeastern Inner Mongolia. The results show that the EOS is delayed by precipitation in summer, and the effect of temperature on the EOS is weak. In addition, too much soil water will lead to the advance of the EOS, and the significant-negative-correlation spatial distribution of EOS_SIF and SW_Summer overlaps with the positive-correlation spatial distribution of EOS_SIF and PRE_Summer, mainly distributed in the meadow steppe at the border between Inner Mongolia and Mongolia and the forest area in northeastern Inner Mongolia. However, in the coniferous forest region in western Mongolia, precipitation is significantly positively correlated with EOS_SIF; meanwhile, in this region, soil water and EOS_SIF were not significantly negatively correlated. This shows that the precipitation in western Mongolia is lower than that in central and northeastern Inner Mongolia, and will not lead to excessive soil moisture. The precipitation on the arid and semiarid Mongolian Plateau is mostly concentrated in summer; thus, excessive precipitation will lead to high soil-moisture in areas with high precipitation, which will affect the growth of vegetation. In general, EOS_SIF is more sensitive to meteorological factors than EOS_NDVI and EOS_NIRv.

The impact of autumn precipitation on the three EOSs was not significant, which may be due to the low precipitation in autumn (Figure 12a,d,g). TEM_Autumn and EOS_SIF have a significant positive correlation (Figure 12b). A total of 11.49% of the pixels passed the significance test, mainly distributed in the middle and south of Inner Mongolia, which is mostly sandy vegetation or desert grassland and agricultural land. The first two vegetation types may have strong resistance to drought, while agricultural land may be affected by human irrigation. A 65.17% positive correlation was found between autumn soil-moisture and EOS_SIF, and 14.07% of the pixels passed the significance test (p < 0.05), which indicates that soil moisture is most important for vegetation growth in autumn, with less precipitation (Figure 12c). This further explains the negative correlation between soil moisture and EOS in the rainy season. Both EOS_NDVI and EOS_NIRv are less sensitive to meteorological factors than EOS_SIF. This shows that SIF has a stronger feedback ability toward environmental changes.

3.4. Geographical-Detection Model for SOS and EOS Drivers

From the perspective of spatial differentiation, we analyzed the explanatory power of each factor for the three SOS and EOS. Furthermore, we used the interpretation of the population density on the SOS and EOS to quantify the impact of human activities on the vegetation phenology (Figure 13).

According to the results of the factor detector, the factors with the highest interpretation degree (q value) for SOS_SIF were PRE_Spring and SW_Spring, with q values of 0.0964 and 0.0699, respectively (Figure 13a). This result shows that the water factor had a high degree of interpretation of SOS_SIF in spatial differentiation, but the SFC had the lowest degree of interpretation of SOS_SIF among all factors. The factors with the highest interpretation of SOS_NDVI were SMD and SFC, and the q values were 0.3362 and 0.2462, respectively (Figure 13b). The degree of interpretation of the population density to SOS_NDVI was higher than that of spring precipitation. SFC and SMD were the factors with the highest interpretation of SOS_NIRv, with q values of 0.3425 and 0.2442, respectively (Figure 13c). Unlike SOS_NDVI, SFC had a higher interpretation of SOS_NIRv than SMD. The degree of interpretation of the population density to SOS_NDVI was higher than that of spring precipitation, temperature and soil moisture. The results show that all influencing factors had the best explanatory power for SOS_NDVI (the highest q value), human activities had the highest explanatory power for SOS_NIRv, and SOS_SIF had the lowest explanatory power.

The factors with the highest degree of interpretation (q value) for EOS_SIF were TEM_Summer and TEM_Autumn. The q values were 0.2496 and 0.2393, respectively. This result shows that the temperature factor had a high degree of interpretation of EOS_SIF in terms of spatial differentiation, but the population density had the lowest degree of interpretation of EOS_SIF among all factors. The factors with the highest interpretation of SOS_NDVI were PRE_Summer and PD, and the q values were 0.0751 and 0.0562, respectively. The factors with the highest interpretation of EOS_NIRv were PRE_Summer and SW_Summer, and the q values were 0.08 and 0.0704, respectively. The interpretation of the population density to EOS_NDVI was higher than that of SW_Autumn, TEM_Summer, PRE_Autumn and TEM_Autumn. The results show that each influencing factor had the best explanatory power for EOS_SIF (the highest q value), human activities had the highest explanatory power for EOS_NIRv, and EOS_NDVI had the lowest explanatory power.

4. Discussion

At the beginning of the study, we analyzed the SIF, NDVI and NIRv trends and breakpoints. The difference in the three types of data is mainly shown in the sensitivity difference of the breakpoint occurrence time to different vegetation types. The breakpoints detected by SIF and NIRv were later than those detected by NDVI. SIF and NIRv can detect mutations in a short time and add more details of changes in trend. This finding shows that reflecting the advantages and disadvantages of vegetation growth based on the data may lead to some vegetation physiological information being neglected. Therefore, the retrieval of vegetation phenological parameters from singular data may have limitations.

From the perspective of time differentiation, the phenological parameters estimated by NDVI and NIRv were closer, while those estimated by SIF were quite different. In addition, SIF is related to the chlorophyll concentration and directly characterizes photosynthesis. At higher latitudes on the Mongolian Plateau, SIF signals are weak in spring and autumn; thus, the estimated SOS is later, and the EOS is earlier. The difference between NDVI and NIRv is small, and the time trend of EOS_SIF is also opposite to that of EOS_NDVI and EOS_NIRv. This result may be because the phenological changes of different vegetation types are different, and SIF is strongly affected by some vegetation types; thus, the time-trend is different from that of NDVI and NIRv. This finding is related to the difference in the spatial distribution of the phenological trend changes (Figure 4d–f) and the difference in the spatial distribution of the change-points (Figure 7). The relationships between the three phenological parameters and altitude were also different. SOS_NDVI and SOS_NIRv decrease significantly as the altitude increases. The opposite is true for SOS_SIF. At the same time, EOS_NIRv increases significantly with elevation, EOS_SIF decreases significantly, and NDVI decreases slightly. These results shows that photosynthesis is negatively correlated with altitude, and the duration of vegetation photosynthesis is also shortened with increasing altitude. However, this situation is not necessarily reflected in vegetation greenness. The NDVI and NIRv which were obtained based on spectral reflection do not have this obvious feature. In contrast, the LOS estimated by NIRv increases with elevation, which may be caused by higher chlorophyll concentrations and deeper vegetation colors in high-altitude areas. However, the change in color alone may not directly represent the change in vegetation photosynthesis. For example, the lag in NDVI changes caused by precipitation or temperature has been widely reported [59,60,61]. Vegetation with high greenness may still retain a certain chlorophyll concentration during the EOS. At this moment in time, the activity of photosynthesis-related enzymes is inhibited, due to the decrease in temperature. This may make some difference. SIF directly represents the change in photosynthesis. When photosynthesis is affected by environmental stress, it will immediately show a change in chlorophyll fluorescence. The study of this part still needs to compare and analyze the relationship between phenology and the elevation gradient of different types of vegetation.

In terms of time, the change-point of SOS_SIF is earlier than that of SOS_NDVI and SOS_NIRv (the most change-point is from 2001 to 2003). However, the time of SIF and NIRv is closer (mainly from 2001 to 2003). In addition, in the analysis of EOS change-points, NDVI and NIRv are more similar in time and space. The reason may be that SIF is usually sensitive to changes in forest areas, desert vegetation and agricultural land. NDVI is more sensitive to changes in grassland ecosystems. In addition, the results based on the change-point detection showed that SOS_SIF mainly showed a reversal of the decreasing trend to an increasing trend. The opposite was true for SOS_NDVI and SOS_NIRv. The greenness of the vegetation returned earlier, but photosynthesis was not enhanced and the inflection-point based on the SIF curve was delayed. The early greening of vegetation may not be beneficial to vegetation growth, but lead to a decrease in photosynthetic capacity. Meanwhile, the trend-change of SOS_SIF was earlier than that of SOS_NDVI and SOS_NIRv, which may indicate that SIF can signal the trend-change of vegetation phenology in advance, in terms of the change of phenology.

Additionally, phenological parameters extracted from different observation angles may represent different meanings, and their responses to climate factors may also be different. Because SOS_SIF is too late, the correlation between SOS_SIF and snow-cover is weak. However, the correlation between the SOS and EOS which were estimated based on SIF and meteorological factors in the spring, summer and autumn, was the most significant. This result shows that the phenological parameters estimated by SIF have the highest sensitivity to environmental changes within a comparison of the three phenological parameters. In general, the impact of environmental factors on vegetation physiology is significant. This may be due to the rapid response of SIF to environmental stress and the advantage of direct characterization of photosynthesis. Reflectivity-based NDVI and NIRv have a certain time-lag in response to environmental changes, which may also be reflected in the extraction of vegetation phenology.

The results also show that the snowmelt period and soil moisture were the most important factors in spring, but their effects on the SOS were very complex. Based on the path-analysis results, the change in the SOS is directly or indirectly caused by multiple factors, or even superposition. The increase in SCF will delay the SMD but also increase the soil water content. The effect of precipitation in spring is less. Therefore, the soil moisture in relatively dry springs may mainly come from the melting supply of snow. The soil moisture content in spring may still not be high overall. Therefore, for the advance of the SOS, the soil moisture must be fully supplied to the spring vegetation instead of being consumed by evaporation. Because the temperature in spring has a strong negative correlation with the snow-melting period and soil moisture, the effect of temperature on the SOS is indirect. This result mainly shows that the temperature causes the SMD to advance and the melted snow to supply soil water prematurely. Simultaneously, the temperature will further lead to the consumption of soil water. During the time interval between the SOS and the snow-melting period, a large amount of soil water will be consumed, and the soil water will not be fully used by growing vegetation. Thus, in a dry spring, the right time for soil water to supply vegetation may be even more important than the total amount of soil water. The delay of the SMD can have critical effects. Of course, soil water is not only supplied by snow; frozen soil is also very important [62,63]. In addition, the positive effect of snow-cover on vegetation is not limited to water supplementation; nitrogen from melted-snow supplementation is also important [64,65,66,67]. The SMD showed a significant correlation with SOS_NDVI and SOS_NIRv. Additionally, from the perspective of regional differentiation, the interpretation of winter snow cover to SOS_NIRv is the highest. The interpretation of EOS_NDVI by natural factors in summer and autumn is low, and, except for summer precipitation, the interpretation of SOS_NDVI by population-density is high. Moreover, the q value of natural factors and EOS_SIF is the highest. Accordingly, NDVI in the Mongolian Plateau may be more suitable for extracting the SOS, and SIF is more suitable for extracting the EOS.

5. Conclusions

We used SIF, NDVI and NIRv to retrieve vegetation changes and detect discontinuities. Based on three types of remote-sensing data, we estimated the phenological parameters of the Mongolian Plateau for 20 years, and conducted a trend analysis and change-point detection. Then, we analyzed the potential impact of meteorological factors in each season and the human factors, on phenological parameters.

The main conclusions of this study are as follows:

• SIF and NIRv can retrieve more details of vegetation dynamics and detect breakpoints earlier than NDVI;
• On the Mongolian Plateau, SOS_SIF is the latest, and EOS_SIF is the earliest. The time-trends of the phenological parameters estimated based on NIRv and NDVI were similar, and SIF can better express the relationship between altitude and vegetation phenology. The year of main change-points of SOS_NDVI and SOS_NIRv is later than that of SOS_SIF; the change-point of EOS_SIF and EOS_NIRv is earlier than that of EOS_NDVI. The trend change-points of EOS_SIF and EOS_NIRv were mostly concentrated from 2001 to 2007, and the change-points of EOS_NDVI occurred later. The difference in different phenological parameters is mostly due to the difference in the sensitivity of different data to different types of vegetation changes, as well as the different data characteristics;
• The contribution of snow-cover to SOS is mainly shown in the late snow-melting date, which can reduce the evaporation loss of melted snow-cover, and fully supplement the vegetation in the green-returning period as soil water;
• The correlation between spring climate factors and SOS_NDVI is most significant. The correlation between summer and autumn meteorological factors and EOS_SIF is most significant. From the perspective of spatial differentiation, each influencing factor has the highest degree of interpretation for SOS_NDVI and EOS_SIF. The population density has the highest interpretation of SOS_NDVI and EOS_NIRv. On the Mongolian Plateau, NDVI may be more suitable for extracting the SOS, while SIF is more suitable for extracting the EOS. According to the analysis results from different angles, NIRv did not show the most prominent advantages, and SIF is more predictable in the trend analysis of phenology.