**3. Results**

#### *3.1. The Performances of Satellite-Based SOS and EOS*

This study first verified the performances of the SOS and EOS that were produced from the 250 m MODIS NDVI and GIMMS3g NDVI against the ground-observed data from 2001 to 2013. The results showed that a good agreemen<sup>t</sup> was observed between the MODISderived phenology and ground-observed data, where the r values of the MODIS SOS and EOS were 0.83 and 0.79, respectively (*p* < 0.01, Figure 2). However, the phenology that was estimated from GIMMS3g NDVI showed poor consistency. The r values of the GIMMS3g SOS and EOS were only 0.49 and 0.21, and both the *p*-values were more than 0.05.

**Figure 2.** Validations of the (**a**) SOS and (**b**) EOS that were derived from the 250 m MODIS NDVI and the (**c**) SOS and (**d**) EOS that were derived from the GIMMS3g NDVI.

#### *3.2. Spatial Patterns of Vegetation Phenology*

Figure 3 presents the SOS and EOS of the Loess Plateau in 2001. The spatial patterns of the vegetation phenology that were produced from the GIMMS3g data were quite different from those of the MODIS data. However, three sets of MODIS phenology results with different resolutions gave a similar spatial distribution (Figure A1). The spring phenology of croplands in the Weihe Plain was the earliest, with an average SOS of DOY 46. Most SOSs in the grassland region of the central Loess Plateau were later than DOY 170. In particular, the SOS that was derived from MODIS NDVI in the Mu Us Sandy Land region was earlier than that in the surrounding areas. The early EOS in the Loess Plateau was mainly concentrated in the southern region, while the late EOS was mainly concentrated in the western region and the Lvliang mountains. Among the land cover types, the EOS of almost 80% of the croplands was between DOY 260 and 290, the EOS date of the forests was the earliest, and that of the grasslands was the latest.

**Figure 3.** Spatial patterns of the SOSs were estimated from (**a**) 250 m MODIS and (**b**) GIMMS3g data from 2001; spatial patterns of the EOSs were estimated from (**c**) 250 m MODIS and (**d**) GIMMS3g data from 2001.

Compared with the MODIS-derived phenology, the spatial distribution of the phenology period that was identified by the GIMMS3g data was more concentrated. Moreover, the SOS values that were derived from the GIMMS3g product were largely consistent with the phenology that was derived from the MODIS data, which was mainly distributed in the Weihe Plain. The SOS that was produced by the GIMMS3g NDVI data in the northern part of the Loess Plateau was concentrated in DOY 90–110, and the spatial details were greatly neglected. Additionally, there was a significant difference between the GIMMS3g EOS and the MODIS EOS in the south–central region of the Loess Plateau. The difference ranged from 20 days to more than 60 days.

Figure 4a,b presents the raw and smoothed NDVI in heterogeneous areas, where the land use types are grassland and forest, respectively. Figure 4c shows that the time series of NDVI in the Weihe Plain, where the land use type was cropland. Based on the time-series data, we found that no matter what year it was, the inflection point of the time series of the original GIMMS3g data was always concentrated in the seventh or eighth period's data in heterogeneous areas. However, this phenomenon did not occur in flat areas. This indicates that the problem of GIMMS3g data is one of the important reasons for the spatially aggregated distribution of GIMMS3g phenology.

**Figure 4.** Time series of (**a**) grassland, (**b**) forest, and (**c**) cropland NDVIs from the GIMMS3g and MODIS products.

In order to compare the differences between the GIMMS3g-derived and MODISderived phenology in flat areas, we calculated the differences of the two datasets in the Weihe Plain during the period 2001–2015 (Figure 5). The findings showed that the differences in SOS (GIMMS3g SOS–250 m MODIS SOS) were mainly less than 5 days and 10 days in 43.49% and 70.66%. The GIMMS3g data performed well at monitoring SOS over the flat areas. In addition, we found that the frequency with which the differences in SOS were greater than 20 days decreased significantly and became close to 0. In the results showing the SOS differences greater than 25 days, the value of the GIMMS3g SOS was always greater than that of the 250 m MODIS SOS. This shows that GIMMS3g SOS tended

to be later than MODIS SOS. However, differences in the EOS (GIMMS3g EOS–250 m MODIS EOS) that were less than 5 days or 10 days were only 13.85% and 29.88%. In addition, for the entire Loess Plateau, the proportion of differences in EOS between the GIMMS3g and 250 m MODIS data that were less than 5 days was still less than 20%. This indicated that the consistency of the GIMMS3g EOS and the MODIS EOS was poor, even in flat areas.

**Figure 5.** The distribution of the (**a**) SOS and (**b**) EOS differences between the 250 m MODIS and GIMMS3g phenology during the period 2001–2015.

#### *3.3. Temporal Variation in Vegetation Phenology*

The inter-annual trends of the SOS and EOS during the period 1982–2020 are presented in Figure 6. The results showed that the inter-annual trends of the SOS and EOS from GIMMS3g were the reverse of those of the MODIS data. The SOS showed a trend of postponing and the EOS presented a trend of advancing from 1982 to 2015 based on the GIMMS3g data. In addition, the SOS (EOS) of the GIMMS3g data delay (advance) trend during the period 2001–2015 was more significant, and the trend lines K of the SOS and EOS were 0.34 and −0.14, respectively. In contrast, the SOS (EOS) showed an advanced (delayed) trend based on the MODIS data, and the trend line K of the MODIS SOS and EOS was −0.63 and 0.19, respectively, during the period 2001–2015.

In the comparison of MODIS products, our findings showed that the phenological periods that were derived from MODIS products with various spatial resolutions gave only small differences. The average difference between the 500 m MODIS SOS (EOS) and 250 m MODIS SOS (EOS) was only 1.2 (0.3) days. The correlation coefficient and RMSE between the 500 m MODIS and 250 m MODIS results were greater than 0.99 and less than 0.60, respectively (Table 2). Moreover, the average difference between the 1000 m MODIS SOS (EOS) and 250 m MODIS slightly increased to 1.7 (1.4) days. The correlation coefficient and RMSE between the 1000 m MODIS and 250 m MODIS results were greater than 0.95 and approximately equal to 1.0, respectively. This demonstrated that there was little difference between the 1000 m MODIS NDVI and the 250 m MODIS NDVI. Therefore, the 1000 m MODIS NDVI could be used to monitor the LSP of the Loess Plateau. However, the GIMMS3g product may not be able to accurately monitor heterogeneous areas, such as the Loess Plateau.

Figure 7 shows the spatial trend of the phenology over the Loess Plateau that was calculated from the MODIS data during the period 2001–2020, which passed the significance test of α = 0.05. The area with a significantly advanced SOS was about 16.7 × 10<sup>4</sup> km2, which was about a quarter of the area of the Loess Plateau. Areas with a significantly advanced SOS were mainly in the central and northeastern regions of the Loess Plateau. The area of the delayed SOS was only one-third that of the advanced SOS. Meanwhile, the delayed SOS was mainly distributed across the croplands of the Weihe and Hetao Plains. Moreover, the area with a significantly delayed EOS was about 9.3 × 10<sup>4</sup> km2, which was

more than three times that of the area with an advanced EOS. More than two-thirds of the advanced SOS and delayed EOS occurred in grasslands, which determined the overall phenological changes in the Loess Plateau. Although the land cover types in the Loess Plateau changed dramatically, the phenological changes were still dominated by the areas where the land use did not change. Only about 20% of the SOS or EOS significant trends occurred in areas with changes in the land cover type.

**Figure 6.** Inter-annual trends of the (**a**) SOS and (**b**) EOS that were estimated from the GIMMS3g and MODIS products during the period 1982–2020. The yellow, red, green, and blue solid lines represent the SOS and EOS values that were inferred from the GIMMS3g NDVI (1982–2015) and MODIS NDVI (2001–2020). The yellow dashed line shows the GIMMS3g-based SOS and EOS trends during the period 1982–2015. The pink dashed line shows the GIMMS3g-based SOS and EOS trends during the period 2001–2015. The black and red dashed line shows the 250 m MODIS-based SOS and EOS trends during the periods 2001–2015 and 2001–2020, respectively.

**Table 2.** The correlation coefficient and RMSE that were used for the comparison of the vegetation phenology between different MODIS products.


The symbol \*\* indicates significance at the 0.01 level.

**Figure 7.** Spatial trends of the (**a**) SOS and (**b**) EOS from the 250 m MODIS time-series data during the period 2001–2020.

#### *3.4. Impact Factors on MODIS Products*

#### 3.4.1. Influences of Vegetation on MODIS Products

To examine the possible causes of the difference in LSP from the MODIS data, we investigated the differences in the SOSs and EOSs between MODIS products that were used on different land cover types (Figure 8). Our findings showed that the largest difference between the 250 m MODIS SOS and the 1000 m MODIS SOS (1000 m MODIS SOS–250 m MODIS SOS) in forests was 3.5 days, which was larger than the difference found in grasslands (1.9 days) and croplands (0.6 days). Additionally, the standard deviation of the inter-annual difference between the 250 m and 1000 m MODIS SOS in forests was 1.1 days, which was the largest value that was obtained among all vegetation types. The differences that were obtained between the 250 m MODIS EOS and the 1000 m MODIS EOS were 0.9 days (forests), 1.2 days (grasslands), and 1.1 days (croplands). The standard deviation of the inter-annual difference was the largest in forests. In addition, we found that the differences in the SOS between the 250 m MODIS and 500 m MODIS (500 m MODIS SOS–250 m MODIS SOS) were both less than one day.

#### 3.4.2. Influences of AT10 on the Phenology

Figure 9 presents the variations in the differences that were obtained between multiple sets of the SOS and EOS with an AT10 from 1 January to 30 April and from 1 September to 31 October, respectively. As the AT10 from January to April increased, the difference between each MODIS SOS gradually increased. In particular, the difference between the 1000 m MODIS SOS and the 250 m MODIS SOS (1000 m MODIS SOS–250 m MODIS SOS) was greater than the difference between the 500 m MODIS SOS and the 250 m MODIS SOS (500 m MODIS SOS–250 m MODIS SOS). The standard deviation of the 1000 m MODIS SOS was greater than that of the 500 m MODIS SOS. The relationship between the EOS and AT10 was found to be opposite to that of the SOS and AT10. Our results showed that with the increase in AT10 from September to October, the difference between each MODIS EOS gradually decreased. No matter whether the SOS or EOS were used, the difference in vegetation phenology between the 250 m and 1000 m products was less than 3 days. Additionally, we found that there was almost no difference between the 250 m and 500 m vegetation phenology.

**Figure 8.** Inter-annual trends of the differences in the (**a**) SOS and (**b**) EOS between different land cover types over the Loess Plateau. The dashed line represents the difference in the SOS between the 250 m SOS and 500 m SOS for different land cover types during the period 2001–2019. The solid line represents the difference in SOS between the 250 m SOS and 1000 m SOS in different land cover types during the period 2001–2019.

**Figure 9.** The variation in the differences between multiple sets of the (**a**) SOS and (**b**) EOS with an AT10 from January to April and from September to October, respectively. The red line shows the difference between the 500 m MODIS SOS (EOS) and the 250 m MODIS SOS (EOS). The black line shows the difference between the 1000 m MODIS SOS (EOS) and the 250 m MODIS SOS (EOS). The red and gray shadows indicate the standard deviations of the red and black lines, respectively.
