Vegetation indices are developed for evaluating vegetative covers using spectral measurements in the applications of remote sensing. Over forty vegetation indices have been developed in order to enhance vegetation response and minimize the effects of soil brightness, soil moisture, shadow and environmental effects [35]. NDVI is the normalized difference vegetation index and calculated by reflectance in the near-infrared and red wavelengths [36,37]. The NDVI is sensitive to green leaf area and surface vegetation coverage [4,36] and has been widely used in ecological research [37,38,39]. A few research works [4,6,7,8,36,37,38,39] have demonstrated the potential of using NDVI to detect vegetation dynamics. After the GFG project was implemented for more than ten years, the low-yield slope croplands were transformed into grasslands and woodlands, and many barren hills and wastelands turned “green". High-resolution images of remote sensing can detect these changes. MODIS (Moderate Resolution Imaging Spectroradiometer) NDVI with a 250-m spatial resolution can be adapted for the higher resolution detection. To eliminate the influence of cloud cover and atmosphere, the annual maximum value of the NDVI was chosen for use in this study by the maximum value composites methods [40]. Through that, the continuity of NDVI time series was enhanced, and the temporal-spatial change of vegetation could be identified more accurately.

According to the Millennium Ecosystem Assessment [13], more than twenty kinds of ES were listed and could be classified into provisioning services, regulating services, supporting services and cultural services. We only chose three typical and representative services, including food provision (provisioning service), hydrology regulation (regulating service) and climate regulation (regulating service) according to the availability of data, the feasibility of the method and the importance of ES. The explanation was as follows. Food provision service was directly related to human well-being. Moreover, food provision would be affected because of the reduction of cropland in our site. Hydrology regulation was a significant and key service because our site was located in a semi-arid area. Moreover, the increase of forest area might consume more water because of more intense transpiration. In our site, ecosystems were very sensitive to climate changes [4], so climate regulation also was an important service. Moreover, in recent years, carbon emission growth caused by human activities was 13.96% per year in this area [41], and the government undertook the huge responsibility of carbon emission reduction in our site.

Correlation and partial correlation analysis were used to analyze the relationships between ES and different driving factors. Both natural and anthropogenic factors have an important impact on ES [13]. In this paper, the climate fluctuations were chosen to be the main natural driving factors. The anthropogenic driving factors in our site were different for provisioning service (GP) and regulating services (WY and NPP). Temperature (T), precipitation (PPT), agricultural machinery power consumption, chemical fertilizer consumption and labor force input were chosen to be the driving factors of GP change, which was based on the study of Zhang et al. [58]. T, PPT and cumulative afforested area were chosen to be the driving factors of WY and NPP change, which agreed with the research of Xie et al. [59]. The cumulative afforested area could be understood as an indicator of GFG policy.

Partial correlation analysis is used when the two variables are related to the other variables at the same time; it will exclude the impact of the other variables and only analyze the correlation of the two variables [4,60]. For example, if there are three variables, the equation is as follows: (6) R x y - z = ( R x y − R x z R y z ) / ( 1 − R x z 2 ) ( 1 − R y z 2 ) where R_xy-z is the partial correlation coefficient between x and y excluding the impact of z; R_xy is the correlation coefficient between x and y; R_xz is the correlation coefficient between x and z; R_yz is the correlation coefficient between y and z. The calculation of the correlation coefficient R_xy (or R_xz or R_xz) is detailed in Xu [60].

Meteorological data from 32 weather stations (Table S1) within and near the NSLP were collected from the National Climatic Bureau [61], including monthly T, monthly PPT and sunshine duration during the study period.

The tile numbers of the remote sensing data in the NSLP were h26v05 and h27v05. A series of Moderate Resolution Imaging Spectroradiometer (MODIS) data for the CASA model during the 14-year period was derived from the MODIS on NASA’s Terra satellite [62], including surface albedo (MOD09A1), vegetation index (MOD13A2) and land cover type (MCD12Q1). The annual ET product (MOD16A3) was used to estimate WY, and the annual NPP product (MOD17A3) was used for comparison with the simulated NPP results of the CASA model. MOD16A3 and MOD17A3 were derived from the Numerical Terradynamic Simulation Group [63].

Data on agricultural machinery power consumption, chemical fertilizer consumption, labor force input in GP and cumulative afforested area during the 14-year period were obtained from provincial-level Bureaus of Statistics [42] in the NSLP and used for quantifying the relationships between ES and different driving factors.

According to the original meteorological station data within our site, the area exhibited a warming and drying trend during the period from 1955–2013. We divided the complete interval into two periods, 1955–1999 and 2000–2013. From 1955–1999, we found that the annual mean T was 8.97 °C with an increase of 0.019 °C/year, and the annual mean PPT was 469.84 mm with a decrease of 2.61 mm/year (Figure 2). From 2000–2013, the annual mean T was 9.78 °C with a decrease of 0.024 °C/year, and the annual mean PPT was 433.48 mm with an increase of 8.46 mm/year (Figure 2). The annual mean T was higher and PPT lower from 2000–2013, while the changes of T and PPT were opposite to those of 1955–1999. At the pixel level, T and PPT both had higher values in the south than the north (Figure S2). In the last 14-year period, the T of 59.32% of the area was decreasing, while the PPT of the whole area was increasing (Figure S2). The global surface T was increasing over the past few decades, while the situation seemed to be different in our site. GFG policy might play a part of the role, because the most T decreasing regions were distributed in the GFG regions (Figure S2).

From 2000–2013, the NDVI exhibited a significantly increasing trend of 0.01/year (R² = 0.858, p < 0.05), and thus, it improved 33.02% from 2000–2013 (Figure 3). From 2000–2002, the NDVI was in a period of rapid growth with an increase of 0.02/year, because the scale of GFG was larger in the early days of the project [4]. From 2003–2007, the NDVI was in a fluctuating period and kept decreasing from 2005–2007. Some vegetation (e.g., Hippophae rhamnoides L.) was planted, but it was not suitable for the semi-arid climate in the NSLP and needed more water when growing [64]. When PPT was low, these plants started to wither. More suitable vegetation was subsequently planted (e.g., Pinus sylvestris), and the NDVI started increasing from 2008.

In spatial terms, the NDVI in 85.62% of the area showed an increase (Figure 4b), and the NDVI in 52.96% of the area showed a significantly increasing trend (p < 0.05) (Figure 4c). Those areas were in the mid-eastern NSLP, which originally had low vegetation coverage (Figure 4a) and was therefore heavily involved in the GFG program from 2000–2013. At the same time, the NDVI in 14.38% of the area showed a slight decrease (Figure 4b), and these decreases were mostly not significant. The decreased areas were located in the southern NSLP (i.e., Ziwu Mountain and Huanglong Mountain) with high forest coverage. Some other researchers have found similar results either on larger scales [4,6] or at the same scale [64,65].

Response of regulating services to vegetation change is an important topic with the GFG project proceeding, but it has not been fully understood in current research [25,26,27,28,57]. At the pixel scale in our site, response relationships between regulating services (WY and NPP) and NDVI were analyzed and identified for the period from 2000–2013 (Figure 7). The response of WY change to NDVI change was negative (Figure 7a) (R² = 0.012, R = −0.111, p < 0.01), which showed a negative relationship between WY and vegetation restoration. Vegetation restoration may have increased evapotranspiration (ET) and reduced regional WY, as has been found by other researchers [12,69,70]. The response of NPP change to NDVI change was positive (Figure 7b), which showed a positive relationship between NPP and vegetation restoration and that was consistent with previous research [2,59,68]. Moreover, Figure 7 showed that the response levels of regulating services to vegetation change were NPP > WY.

From the results of correlation coefficients and partial correlation coefficients, both chemical fertilizer and agricultural machinery power were significant positive with GP from 2000–2013 in our site (Table 2). The GP correlation level of chemical fertilizer was higher than that of agricultural machinery power. The correlation between natural factors (T and PPT) and GP both were not significant (Table 2). A negative correlation was identified between labor force and GP from 2000–2013. During the study period, the labor force input in GP kept decreasing. However, other investments in GP including chemical fertilizer and agricultural machinery power were increasing, so the GP still kept increasing in our site. We divided the whole period into two stages: a GFG-extensive period (2000–2006) and a GFG-weakened period (2007–2013). It seemed that the correlation between different factors (natural or anthropogenic factors) and GP was stronger in the GFG-weakened period than in the GFG-extensive period (Table 2).

Over the whole regional scale, neither T nor cumulative afforested area were significantly correlated with WY from 2000–2013, but a significant positive correlation was identified between WY and PPT (Table 3). The strong response of WY change to PPT change at the pixel scale also showed that WY was significantly correlated with PPT (Figure 8a). The correlation between natural factors and WY was stronger in the GFG-weakened period than in the GFG-extensive period (Table 3). Through the correlation coefficients between WY and cumulative afforested area in different periods, it seemed that we could not judge that WY was increased or decreased with the increasing of cumulative afforested area. However, the partial coefficients in different periods were consistent and indicated that WY was decreased with the increasing of cumulative afforested area.

Over the whole regional scale, neither T nor PPT were significantly correlated with NPP (Table 4). NPP was decreased with the increasing in T, and the response of NPP change to T change at the pixel scale could indicate that (Figure 8b). The NSLP is a semi-arid area, so higher T would intensify soil evaporation and restrain plant photosynthesis [4]. The response of NPP change to T change at the pixel scale indicated that NPP was increased with the increasing in PPT (Figure 8c). Vegetation growth was restrained by moisture in the NSLP, so higher PPT enhanced plant activity and vegetation growth. During the 14-year period, the correlation between NPP and cumulative afforested area was significant and positive (Table 4). A stronger correlation was identified between NPP and cumulative afforested area during the GFG-extensive period than the GFG-weakened period (Table 4).

NPP and ET are crucial to the process of assessing ES change. NPP was estimated by the CASA model using multi-source data. ET was derived from the MODIS product with strong variability and it was a key and uncertain parameter in simulating WY. NPP and ET are affected by vegetation physiological factors and complex environment factors, so it is necessary to evaluate the credibility of the NPP and ET values.

The several simulated NPP values in the Loess Plateau have differences, and the main factors responsible include the simulation method, the study period and the scope of the Loess Plateau (Table 5) [71]. With respect to the scope of the Loess Plateau, annual NPP values simulated by Biome-BGC (Biogeochemical Cycles) were 10.08% lower for our site than for the whole area of the Loess Plateau. The simulated NPP value from the CASA model was lower by 11.69%–12.26% for our site than for the whole area of the Loess Plateau. This comparison showed that the CASA model used for our site performed well. Overall, the NPP values simulated by Biome-BGC were lower than the CASA values, partly because CASA simulates NPP at monthly intervals, and its estimates of the fraction of photosynthetic active radiation may therefore be larger [59].

The average NPP estimated by the CASA model (272 g C/m²·year) was higher than the MOD17-NPP (214 g C/m²·year) (Table 5). In the 25 counties of our site, simulated NPP values ranged from 141 g C/m²·year–695 g C/m²·year, and MOD17-NPP values ranged from 148 g C/m²·year–357 g C/m²·year. The correlation between simulated NPP and MOD17-NPP is significant in all 25 counties (Figure S3a). However, the simulated NPP was much higher than the MOD17-NPP in Fu County, Yichuan County, and especially in Huanglong County and Huangling County (Figure 9 and Figure S3). The above four counties are primarily located near Ziwu Mountain and Huanglong Mountain, which have high vegetation coverage (Figure 4a), so the NPP value should be higher there than in other counties, and some research corroborates this [3,68]. MOD17-NPP values were in middle range in the four counties, which shows that NPP simulated by the CASA model is more reasonable in those areas. Figure 9 shows that MOD17-NPP had small fluctuations in the 25 counties, which was mainly due to the simulation of MOD17-NPP at a global scale, while the NPP as estimated by CASA exhibited reasonable variation. During the 14-year period over the whole area, the annual simulated NPP and MOD17-NPP also had a significant correlation (R² = 0.61, p < 0.01) (Figure S3b). Overall, the NPP simulated by CASA in this study is reliable.

In order to demonstrate the reliability of the ET dataset in our site, we offer the following points. First, the validity of MOD16-ET has been verified on different scales, as has been shown for Asia [73], Korea [74], China [75], Shaanxi Provence [76] and the Wei River Basin [77]. Second, the synergy between NPP and ET, which has been found in our site by Jia et al. [33], means that ET can be enhanced with the increases in NPP. At the pixel scale, the correlation between annual NPP and ET from 2000–2013 was analyzed and found to be significantly positive (R² > 0.8, p < 0.01) (Figure 10a). The response of the ET change trend to the NPP change trend was also analyzed, and it too was significantly positive (R² = 0.282, p < 0.01) (Figure 10b). As a result of this positive correlation between NPP and ET, the reliability of MOD16-ET in the NSLP is further enhanced.

According to the relationships between GP and different factors (Table 2), neither T nor PPT were significantly correlated with GP during the study period, but the correlations between GP and the investments in agriculture (i.e., chemical fertilizer and agricultural machinery power) were significant (Table 2). Fu et al. [78] considered GP to be closely related to agricultural production conditions, artificial inputs and technical progress in the Loess Plateau. Our study showed that the increasing GP was mainly driven by the increasing agricultural investments in our site. Chemical fertilizer investment ranged from 155.71 kg/ha in 2000 to 330.29 kg/ha in 2013 (Figure 11a); it thus grew 1.12-times from 2000–2013 with a significantly increasing trend of 15.39 kg/ha·year (R² = 0.928, p < 0.01). Agricultural machinery power grew 2.07-times and ranged from 19.80 kW/ha in 2000 to 60.77 kW/ha in 2013, with a significantly increasing trend (R² = 0.996, p < 0.01) (Figure 11b). Moreover, our study also showed that chemical fertilizer input was more effective than agricultural machinery power input (Table 2). Other minor drivers of GP improvement in our site may include farm management and the construction of high quality basic croplands [2,65,66].

According to the relationships between WY and different factors (Table 3), both T and cumulative afforested area had a weak connection with WY change from 2000–2013 (Table 3). However, the negative effect of afforestation was significant in the GFG-extensive period (Table 3). The response of WY change to NDVI change showed there was a negative relationship between vegetation restoration and WY (Figure 7a). In this study, WY was calculated by subtracting ET from PPT. During the study period, ET increased with the growth of woodlands (Figure 11c), which can partly explain the negative relationship between afforestation or vegetation restoration and WY. However, the increase in ET was slight, and the annual growth in PPT (8.46 mm/year) was much larger than the growth in ET (1.90 mm/year), which suggests that the slight improvement of WY was mainly driven by PPT during the study period. The strong and positive response of WY change to PPT change (Figure 8a), significant correlation (Table 3) and linear fitting (Figure 12a) between WY and PPT can enhance the credibility of that conclusion. However, the afforestation producing the negative effect on WY still cannot be ignored.

According the relationships between NPP and different factors (Table 4), neither T nor PPT were significantly correlated with NPP during the study period, but the correlation between NPP and cumulative afforested area was significant. Xie et al. [59] consider land use change caused by GFG to be the main factor in increasing NPP in the Loess Plateau. Driven by GFG policy (the input of labor and funding), the decreasing croplands and increasing woodlands (including scrublands) were the main types of land use change in our site between 2000 and 2013. That change brought about the recovery of natural vegetation systems, which played a major role in the increase in NPP. Additionally, NPP was in fact significantly increasing in the mid-eastern NSLP, including Jia County, Hengshan County, Mizhi County, Wubao County, Zizhou County, Suide County, Zichang County, Qingjian County, Yanchuan County and Yan’an County, which correspond with the main GFG regions. The significant correlation (Table 4) and linear fitting (Figure 12b) between NPP and cumulative afforested area showed that the sustained growth of NPP was mainly caused by afforestation (or GFG policy).

The drivers involved in ES changes in our site from 2000–2013 are complex. For different ES, there are different dominant factors, some of them natural and other anthropogenic. The change in WY is attributed to the PPT belonging to the climate factor. The improvement in GP can be explained by the increasing in chemical fertilizer and agricultural machinery power, and the increasing in NPP can be explained by afforestation. Actually, these drivers (i.e., chemical fertilizer, agricultural machinery power and afforestation) can be attributed to the investments in natural capital. Ecosystems (e.g., forest ecosystem, farmland ecosystem) have been seen as natural capital assets, with the potential to generate a stream of vital life-support services [79]. The stable and persistent investments in natural capital are crucial to improve some ES. Recent research at a large scale [1] have found that the investments in natural capital starting in 2000 have significantly improved the key ES in China. On a relatively small scale, our analysis showed that the investments in the farmland system improved food provision and that the afforestation increased climate regulation service.

GFG policies increased GP and NPP in the NSLP, and those policies provide a natural foundation for maintaining ES. However, the ecosystem and socioeconomic system are interactive, and during this interaction, farmers’ attitudes and strategies are key to sustainable developments in ecological restoration [80]. Close cooperation between government and local stakeholders can maximize the effects of GFG projects [2]. The benefits produced by ecological restoration happen on a large scale (e.g., regional and national), while the costs of ecological restoration are undertaken by the local farmers. When the incomes of local farmers are temporarily reduced by ecological restoration projects, they typically re-plant later in the restored regions [81]. The conflicts caused by the differing scales of the costs and benefits have an impact on the sustainability of ecological restoration. After a dozen years of GFG in the NSLP, ecological restoration has shifted focus from ecological benefit to reconciling ecological and social benefits. However, according to the data derived from the National Poverty Alleviation Office in 2015, 44% of the counties in the GFG region are still poverty-stricken (i.e., Qingjian County, Zizhou County, Suide County, Mizhi County, Jia County, Wubao County, Hengshan County, Zichang County, Ansai County and Wuqi County). Achieving a win-win situation involving both ES and poverty alleviation is the primary objective for the sustainable development of ecological restoration.

Ecological restoration improves regulating services, but it may at the same time cause a drop in provision services and, thus, have a negative impact on local farmers. A few studies have found that there is a trade-off between provision services and regulating services [82,83,84]. However, our case study in the NSLP shows that the GP does not decrease, which mainly benefiting from sustaining investments in agriculture. This suggests that configuring effective compensation mechanisms (e.g., technological and economic) would contribute to the coordinated growth of both provision services and regulating services. Moreover, following the laws of ecological succession by choosing native tree species and developing more ecologically-informed GFG programs are both conducive to achieving stable and sustainable ecological restoration.