**4. Discussion**

#### *4.1. Difference between Satellite-Based LSP and Observations*

Based on the MODIS NDVI and the layer named "composite day of the year," we calculated three sets of MODIS phenology with different spatial resolutions from 2001 to 2020. The GIMMS3g data lacked a "composite day of the year" layer, therefore we regarded the 1st and 16th days of each month as the DOY and calculated the GIMMS3g phenology from 1982 to 2015. We used the ground-observed data during the period 2001–2013 to verify the performances of the phenology that was derived from the MODIS and GIMMS3g data. Our results showed that the MODIS-derived phenology and ground-observed data had a good agreement, but that the correlation between the GIMMS3g-derived phenology and ground-observed data was very bad. This indicates that the SOS and EOS that were identified by the MODIS NDVI were more robust, but that the GIMMS3g phenology performed badly.

Moreover, the MODIS SOS that was identified by the logistics model was earlier than the SOS that was observed on the ground, while the MODIS EOS was later than the ground-observed data. Similarly, previous studies also found that in the north of China, the SOS that was identified by the logistics model based on SPOT satellite data was earlier than the SOS observed on the ground [49]. We sugges<sup>t</sup> that this phenomenon was caused by the phenological recognition algorithm and the time resolution of the data used. The finer the temporal resolution of the image is, the more accurate the identified phenological phase will be [51–54]. However, among the existing vegetation index products, product data with a high time resolution are still limited. Therefore, an alternative suggestion is that using a proper method may make up for the lack of advanced or delayed phenological phases, such as the cumulative NDVI [49,55]. However, it is also worth noting that different methods will cause different problems.

The agreemen<sup>t</sup> between the EOS and the ground-observed data was not as good as for the SOS, especially GIMMS3g EOS. A previous study found that the NDVI at harvest time will be increased due to the noise-reduction algorithm [56]. The land use type of the Weihe Plain is mainly cropland, and the locations of the agro-meteorological stations are distributed around the Weihe Plain. This may lead to worse accuracy for the EOS in cultivated land than for the SOS. In addition, due to the limitation of data from observation stations, the phenological results that were gained in other areas cannot be verified. Compared with other studies, we found that the MODIS phenology was similar to those of other studies in terms of their spatial patterns [49,57].

#### *4.2. Comparisons of Different Product Data*

Effectiveness and using the smallest possible amount of data are matters that must be considered first in experiments. In this study, we compared four sets of remote sensing product data with different resolutions for the extraction of phenology in the Loess Plateau. Notably, the three sets of MODIS results with different resolutions showed good consistency. From the correlation coefficient and RMSE, we found that there was little difference between the 1000 m MODIS and 250 m MODIS results. This indicated that lower-resolution data could achieve the same effect as relatively higher-resolution data. Moreover, computers could process the 1000 m MODIS data much faster. Therefore, the comprehensive performance of the 1000 m MODIS data was better. If errors within the range of 1–3 days are allowed, we can use the 1000 m MODIS NDVI rather than the 250 m MODIS NDVI in future phenological studies.

In contrast, the GIMMS3g products were less effective in the Loess Plateau. From the time-series phenological results of the GIMMS3g products, we observed that the phenology was delayed in spring and advanced in autumn. However, the phenological results based on the MODIS products showed the opposite situation. As we know, many studies have confirmed that global warming advances spring phenology and delays autumn phenology [58–60]. This indicates that the phenological results obtained using the GIMMS3g data may contain errors in complex-terrain regions, such as the Loess Plateau. Additionally, we found that the strong spatial homogeneity of the GIMMS3g NDVI was the main reason for these phenological differences. Except for croplands, from the time series of the original data, no matter what year was used, the inflection point of the GIMMS3g data was always concentrated in the seventh or eighth period, and this led to the maximum curvature of the smoothed timing signal focus in this period. Therefore, the SOS that was identified by the GIMMS3g NDVI data in the northern part of the Loess Plateau was concentrated in DOY 90–110 (Figure A2).

However, previous studies showed that coarse-resolution SOS was comparable with finer-resolution SOS in homogeneous areas [5,42]. In this study, we calculated the difference between the MODIS and GIMMS3g results in the Weihe Plain during the period 2001–2015. Similarly, our findings showed that the GIMMS3g data performed well in monitoring the SOS over flat areas. Furthermore, the cropping intensity in a large area of the Weihe Plain during 1982–2013 changed from cropping twice a year to a single cropping taking place each year [61]. Due to the effectiveness of the GIMMS3g image in the Weihe Plain, the GIMMS3g SOS also showed a delayed trend that was the same as the MODIS phenological trend. Except for the flat area, the large heterogeneous areas in the Loess Plateau were affected by the original GIMMS3g data. Therefore, the original GIMMS3g data from the heterogeneous area incorrectly identified the phenological trend in the whole Loess Plateau.

Moreover, although the inflection point did not occur at the end of the vegetation growth of the raw GIMMS3g data, the proportion of the differences in EOSs between the GIMMS3g and MODIS data that were less than 5 days was still less than 20% for the entire Loess Plateau. This means that the GIMMS3g data had a weak ability to predict the SOS and EOS in areas with complex terrain, while they could better monitor the change in the SOS in relatively flat areas. If the GIMMS3g product cannot be replaced in an experiment on phenology production, other phenology estimation methods may be used, such as the dynamic threshold method [62–64]. Based on the maximum and minimum NDVI each year, the dynamic threshold method was used to determine the SOS and EOS in spring and autumn with the threshold ratio. This method could reduce the deviation in phenological estimation that is caused by the mutation of the temporal signal to a certain extent [65,66].

#### *4.3. Factors for the Differences from MODIS Products*

The differences in vegetation phenology that were determined from MODIS products with different spatial resolutions were mainly due to the land-cover types and temperatures involved. This finding can also be seen in other regions and ecosystems [67–70]. It is usually the case that when the same time resolution of the image is used, the finer the spatial resolution is, the more accurate the phenology properties will be. For example, in forest ecosystems, the vegetation phenology that is identified by MODIS images with a 1000 m resolution is more variable than that identified using images with a 500 m resolution. The main reason for this may be that the structures and functions of forest ecosystems are more complex than those of other natural ecosystems, such as cropland ecosystems [71,72]. Croplands are homogeneous and mainly affected by crop management. In contrast, forests are usually controlled by multiple environmental factors. Previous studies also suggested that the performance of coarse-resolution images over homogeneous areas is better than that over other regions [42,73,74].

Temperature is the main factor that leads to the advancement of spring phenology and the delay of autumn phenology [75–77]. Further, we found that the higher the AT10 during the early growth season was, the greater the differences and variability between the SOSs were. Meanwhile, the lower the AT10 during the late growth season was, the greater the differences and volatility of the EOSs were. Areas with a high AT10 from January to April were mainly distributed south of the Weihe Plain. We suggested that the main reason for this phenomenon was that there are more types of ecosystems in the southeast of the Loess Plateau, while the northwest of the plateau has a relatively homogenous ecosystem. Due to differences in the sensitivity of various types of vegetation to AT10, the SOS in areas with a relatively high AT10 showed greater differences. In addition, areas with a low AT10 from September to October were mainly concentrated in the southwest of the Loess Plateau. The huge elevation fluctuation in this area may be the reason for the phenological differences that were seen in products with different spatial resolutions. Although the AT10 and vegetation type or terrain had an impact on the data, the maximum averaged differences of the SOS and EOS between the 250 m MODIS products and the 1000 m MODIS products were less than three days. Additionally, the phenological difference remained within an acceptable range.
