The CSHC region covers the area between the Toudaoguai and Longmen hydrological stations in the main stream of the YR (Fig. 1). The main stream that flows through the CSHC region is 733 km long and its drainage catchment covers 12.97 × 104 km2, which accounts for 14.8 % of the entire YR Basin. The CSHC region is characterized by arid to semi-arid climate conditions. The annual precipitation in the region during 1961–2011 was 437 mm on average, and varied from lower than 300 mm in the northwest to 580 mm in the southeast (McVicar et al., 2007). The precipitation that occurs during the flood season (June–September) is usually in the form of rainstorms with high intensity and accounts for 72 % of the annual rainfall total. Correspondingly, about 45 % of the annual runoff and 88 % of the annual sediment yield within the region are produced during the flood season. The northwestern part of the CSHC is relatively flat, while the southeastern part is more finely dissected (Rustomji et al., 2008).

Fourteen main catchments along a north–south transect within the CSHC study area were chosen for the study (Fig. 1). These catchments account for 57.4 % of the CSHC area, and contribute about 70 and 72 % of the streamflow and sediment load of the overall CSHC, respectively, based on observed hydrological data during 1961–2011 (Rustomji et al., 2008; Yao et al., 2011). Characteristics of these catchments are presented in Table 1 and Fig. 2, showing that the catchments present strong climate and land surface gradients. The catchments in the northwestern part (nos. 1–6) had relatively lower mean annual precipitation (380 mm < P‾ < 445 mm, where P‾ is mean annual precipitation over 1961–2011) and a low growing season (April–October) LAI (0.41 < LAI < 0.48, where LAI is the leaf area index), while the corresponding values for catchments in the southeastern part (nos. 7–14) were 470–570 mm and 0.63 < LAI < 3.26, respectively.

The entire CSHC region is considered as an additional “catchment”, and it is also examined independently. The streamflow and sediment load for the whole region were taken to be equal to the differences of corresponding measurements between the Toudaoguai and Longmen gauging stations. The average annual precipitation, streamflow, and sediment load of the region during 1961–2011 were 437.27 mm, 33.30 mm and 5.17 Gt, respectively. Both the annual river discharge and sediment load across the region showed significant decreasing trends (-0.82 mm yr-1, p < 0.001 and -0.19 Gt yr-1, p < 0.001, respectively) over the past 5 decades, whereas precipitation decreased only slightly (-0.93 mm yr-1, p=0.25) (Fig. 3).

Monthly streamflow and sediment load data during 1961–2011 were provided by the Yellow River Conservancy Commission of China. Daily rainfall data from 1961 to 2011 at 66 meteorological stations in and around the region (Fig. 1) were obtained from the National Meteorological Information Center of China. The spatial average of rainfall data was calculated using the co-kriging interpolation algorithm with the DEM as an additional input. The hydro-meteorological data (including annual precipitation, P (mm), streamflow, Q (mm), and sediment load, S (t)), specific sediment yield defined as SSY =S/A (t km-2), where A is the drainage area of the hydrological station (km2), sediment concentration defined as SC =S/(Q×A) (kg m-3), and the sediment coefficient defined as Cs= SSY /P (t km-2 mm-1) were estimated for each catchment.

The mean catchment slope gradient based on the ASTER GDEM data with a resolution of 30 m and soil data (scale 1:500000) were provided by the National Earth System Science Data Sharing Infrastructure (http://www.geodata.cn). The land use information as at 1975, 1990, 2000, and 2010 was determined with Landsat MSS and TM remote sensing images at a spatial resolution of 30 m. Six land use types were classified, i.e., forestland, cropland, grassland, construction land, water body, and barren land. The LAI data during 1982–2011 were obtained from the Global Land Surface Satellite (GLASS) NDVI Series with a spatial resolution of 1 km (www.landcover.org, Zhao et al., 2013). The total areas impacted by various SWCMs (i.e., afforestation, grass plantation, terraces, and check-dams) in each catchment during the 1960s–2000s were obtained from Yao et al. (2011).

The non-parametric Mann–Kendall (M–K) test method proposed by Mann (1945) and Kendall (1975) was used to determine the significance of the trends in annual meteorological and hydrological time series. A precondition for using the MK test is to remove the serial correlation of climatic and hydrological series. In this study, the trend-tree pre-whitening (TFPW) method of Yue and Wang (2002) was used to remove the auto-correlations before the trend test. There was no residual autocorrelation remaining after performing the TFPW. A Z-statistic was obtained from the M–K test on the whitened series. A negative value of Z indicates a decreasing trend, and vice versa. The magnitude of the slope of the trend (β) was estimated by (Sen, 1968; Hirsch et al., 1982) β=medianxj-xij-iforalli<j, where xi and xj are the sequential data values in periods i and j, respectively.

The time-trend analysis method was used to determine the quantitative contributions of LUCC and precipitation variability to sediment yield changes. This method is primarily designed to determine the differences in hydrological time series between different periods (reference and validation periods) with different LUCC conditions (Zhang et al., 2011). In this method, a regression equation between precipitation and sediment yield is developed and evaluated during the reference period, and the established equation is then used to estimate sediment yield during the validation period. The difference between measured and predicted sediment yields during the validation period represents the effects of LUCC, and the residual changes are caused by precipitation variability. The governing equations of the time-trend analysis method can be expressed as SSY1=f(P1),SSY2′=f(P2),ΔSSYLUCC=SSY2‾-SSY2′‾,ΔSSYPre=SSY2‾-SSY1‾-ΔSSYLUCC, where SSY′ is the predicted sediment yield, and subscripts 1 and 2 indicate the reference and validation periods, respectively. SSY1‾ and SSY2‾ represent mean measured sediment yield during the reference and validation periods, respectively, and SSY2′‾ represents mean predicted sediment yield during the validation period. ΔSSYLUCC and ΔSSYPre are sediment yield changes during the validation period associated with LUCC and precipitation variability, respectively. Rustomji et al. (2008) found that the square root of annual sediment yield in the catchments of the Loess Plateau was linearly related to annual precipitation. This was, therefore, used in this study as the motivation to develop the precipitation–sediment yield relationship during the reference period: SSY=aP+b. In this study, the full data period of 1961–2011 was divided into three phases (1961–1969, 1970–1999, and 2000–2011). The first period was considered the reference period as the effects of human activities were slight and could be ignored (Wang et al., 2016). During the second stage, numerous SWCMs were implemented. For the third stage, a large ecological restoration campaign (the GFG project) was launched in 1999.

The CSHC region has undergone extensive LUCC caused by the implementation of SWCM and vegetation restoration projects (e.g., the GFG project). Figure 4 shows the distribution of land use types of the region in 1975, 1990, 2000, and 2010. More than 90 % of the whole area was occupied by the cropland, forestland, and grassland. The area of cropland decreased by 26.72 % and forestland increased by 53.15 %, and there was no significant change for the area of grassland (increase of 4.21 %) in the CSHC region from 1975 to 2010. The majority of changes occurred during 2000–2010 due to the GFG (reforestation) project (26.67 % decrease and 36.21 % increase for cropland and forestland, respectively). The transition from cropland to forestland was greater in the catchments of the southeastern part (especially in catchment nos. 7–9) than that in the northwestern part (Fig. 4). In the period 1975–2000, the increase in forestland was 26.34 and 4.55 % in the southeastern and northwestern parts, respectively, and the change in cropland was negligible (only -0.39 and 0.22 %, respectively). During 2000–2010, the forestland increased by 47.79 and 18.30 %, and the cropland decreased by 44.84 and 21.04 % in the southeastern and northwestern parts, respectively.

The SWCMs implemented in the LP included both biotic treatments (e.g., afforestation and grass-planting) and engineering measures (e.g., construction of terrace and check-dam and gully control projects). Afforestation, grass-planting, and construction of terraces were seen as the slope measures, while building of check-dams and gully control projects were the measures on the river channel. Although the area utilized for engineering measures was much smaller than the biotic treatments, they immediately and substantially trap streamflow and sediment load. The fraction of the treated area (area treated by erosion control measures relative to total catchment area) increased from 3.95 % in the 1960s to 28.61 % in the 2000s (Fig. 5). The increase in the treated area was greatest during the 1980s as a result of comprehensive management of small watersheds and the 2000s due to the GFG project since 1999. Some decreases in SWCM areas (i.e., afforestation and check-dams) occurred during the 1990s (Fig. 5) as some planted trees died due to drought and some small and medium check-dams were fully deposited by sediment and then subsequently destroyed by floods.

The growing season LAI of the whole region changed from 0.74 during 1982–1999 to 0.81 during 2000–2011, an increase of 10.16 % (Fig. 5). The LAI did not show a significant increase during 1982–1999 (0.003 yr-1, p=0.11), and it increased significantly during 2000–2011 (0.024 yr-1, p<0.01). The increase in growing season LAI during 1982–2011 was greater for the catchments in the southeastern part (0.009 yr-1) compared to the northwestern part (0.004 yr-1), especially after 2000 (Fig. 6). In the period from 1982–1999 to 2000–2011, the average increase in growing LAI of the 14 sub-catchments was 0.088 yr-1 (0.010–0.183 yr-1), with increases of 0.114 and 0.053 yr-1 in the southeastern and northwestern parts, respectively.

Table 2 shows the trends in annual P, Q, SSY, SC, and Cs of the 15 catchments during the period 1961–2011. The annual P showed a declining trend in all catchments except the Jialu catchment, but the changing trend is only significant in the Xinshui and Zhouchuan catchments (p<0.05). The annual Q, SSY, SC, and Cs showed significant decreasing trends in all the catchments, and most of the decreases were at the 0.001 significance level. For the 14 sub-catchments, the average decrease rates of annual values of Q, SSY, SC, and Cs were 0.86 mm yr-1 (0.24–1.66 mm yr-1), 190.06 t km-2 yr-1 (26.47–398.82 t km-2 yr-1), 2.73 kg m-3 yr-1 (0.69–4.70 kg m-3 yr-1), and 0.38 t km-2 mm-1 yr-1 (0.04–0.87 t km-2 mm-1 yr-1), respectively. The changing rates of Q, SSY, SC, and Cs for the whole region were -0.85 mm yr-1, -131.52 t km-2 yr-1, -2.06 kg m-3 yr-1, and -0.27 t km-2 mm-1 yr-1, respectively. The annual average reductions in the whole region were equivalent to 2.56, 3.30, 2.01, and 3.07 % of the mean annual values of Q, SSY, SC, and Cs, respectively.

The mean and the coefficient of variation, Cv, representing inter-annual variability of annual values of P, Q, SSY, SC, and Cs for the 15 catchments during the three phases (reference period-1, period-2, and period-3) are shown in Fig. 7. Compared to standard deviation, the Cv value was better able to indicate the inter-annual variability of precipitation, streamflow, and sediment load among the catchments with distinctly different average values. Compared to the reference period, the mean annual precipitation decreased by 11.73 % (6.36–15.69 %) and 10.64 % (5.88–16.7 %) on average in period-2 and period-3, respectively. From period-2 to period-3, the change in mean annual precipitation was slight (increased by 1.32 % on average) with a decrease of 2.45–5.87 % in four catchments and an increase in the remaining catchments (0.35–8.29 %). The variability of annual P also decreased as indicated by the reductions of Cv values during period-2 and period-3 (Fig. 7a). In contrast to annual P, the reductions of mean annual Q, SSY, SC, and Cs were clearly more evident. With respect to the reference period, the reduction was 34.41 % (9.45–54.72 %), 48.02 % (17.98–67.61 %), 24.20 % (-9.93–47.77 %), and 39.31 % (4.64–63.5 %) for Q, SSY, SC, and Cs during period-2, and the decreasing rate was even more in period-3, with values of 64.82 % (36.72–84.19 %), 88.23 % (64.94–97.64 %), 67.81 % (17.28–91.12 %), and 85.85 % (63.51–96.97 %), respectively. Cv of annual Q increased in eight catchments, with the remaining ones showing decreasing trends (Fig. 7b), while Cv values for SSY, SC, and Cs increased in all catchments (Fig. 7c–e). The above results indicate substantially different behaviors of the changes among precipitation, streamflow, and sediment load.

The effects of precipitation change and LUCC on sediment yield reductions in period-2 and period-3 were quantified using Eqs. (2)–(6) and the results are shown in Fig. 8. The form of Eq. (5) during the reference period is shown in Table 3. The analysis showed that both decreased precipitation and increased area treated with erosion control measures contributed to the observed sediment load reduction, and that LUCC played the major role. On average, LUCC and precipitation change contributed 74.39 and 25.61 %, respectively, to sediment load reduction from the reference period to period-2, with their respective contributions to sediment load reduction from the reference period to period-3 being 88.67 and 11.33 %. The effect of LUCC in period-3 was greater than in period-2 as the land use/cover (see Figs. 4–5) and vegetation coverage (see Fig. 6) had undergone substantial changes due to the ecological restoration campaigns launched during period-3. From period-2 to period-3, the contribution of precipitation was negative for sediment yield reduction in 11 catchments where the annual precipitation slightly increased, and thus the contribution of LUCC was larger than 100 % (Fig. 8c). In the remaining four catchments, the average contribution of LUCC increased to 83.96 %.

In broad terms there are two factors that govern the annual sediment yield of a catchment: precipitation and landscape properties (soil, topography, and vegetation). Precipitation is the primary driver of runoff and, therefore, directly influences the sediment transport capacity of streamflow and sediment yield at the catchment scale. Higher precipitation means higher streamflow, which is the immediate driver of erosion and sediment transport. Landscape properties not only have an impact on the volume or intensity of streamflow, but also determine the erodibility of the soil. Correlations between the potential factors (precipitation, percentage area of afforestation, pasture plantation, terracing, check-dams and construction land, and LAI) and sediment yield change between different stages (see Table 4) showed that check-dam construction was the dominant factor for sediment yield reduction from the reference period to period-2. Pasture plantation and check-dam construction acted as the dominant factors for sediment yield from the reference period to period-3. The increase in precipitation mitigated the reduction of sediment yield to some degree from period-2 to period-3.

Based on the above results, the variation of SSY mainly depended on precipitation in the reference period before LUCC took effect and any spatial patterns of SSY in the catchments were controlled by differences in annual precipitation and land surface conditions. During the validation period (period-2 and period-3) when increased LUCC had taken effect, SSY decreased considerably. The decrease in precipitation was insignificant and LUCC contributed over 70 % of the sediment yield reduction. In this case, the temporal changes in SSY depended more on the fraction of treated surface area, and precipitation possibly played a secondary role. The spatial pattern of the impacts of precipitation on sediment yield was dependent on the landscape properties among catchments. Guided by this framework, data were next analyzed to generate separate spatial and temporal patterns constituting respective components of the spatio-temporal patterns.

The regression equations of SSY=aP+b are shown in Table 3. The spatial distributions of precipitation–sediment relationships during the three stages are shown in Fig. 9. During the reference period, the correlation between precipitation and sediment yield was significant in 11 catchments (p<0.05) and the coefficient of determination (R2) ranged from 0.48 to 0.87 (Table 3). Furthermore, the precipitation–sediment yield relationship varied from catchment to catchment and showed a spatial pattern. The correlation coefficient between precipitation and sediment yield was greater for catchments in the northwestern part, with an average R2 value of 0.75 and a p value of 0.007 compared to those in the southeastern part, where the average R2 and p values were 0.48 and 0.059, respectively (Table 3). Based on the slopes of the regression equations between annual precipitation and sediment yield, the 14 catchments were classified into four groups (Group-1: a > 0.3; Group-2: 0.2 < a < 0.3; Group-3: 0.1 < a < 0.2; and Group-4: 0 < a < 0.1), which indicate that the sediment production capability of annual precipitation is different among the catchments (Fig. 9a). The four catchments in the northwestern part (nos. 1–3 and 5) had the greatest regression slopes of a > 0.3 (Group-1) and the Shiwang catchment had the lowest regression slope of 0.07 (Group-4). Most of the catchments in the southeastern part were in the second group of 0.2 < a < 0.3. Overall, the regressed equations were significant for most of the catchments, and were suitable for estimating the relative contributions of LUCC and precipitation variability to sediment yield changes.

Compared to the reference period, the correlation between precipitation and sediment yield during period-2 decreased in the catchments, as indicated by lower R2 values in Table 3. The slopes of the regression lines in the period-2 decreased in most of the catchments with respect to the reference period, except in Huangfu, Gushan, and Kuye catchments, which increased slightly. Furthermore, the spatial patterns of the precipitation–sediment yield relationship during these two periods were somewhat different (Fig. 9a and b). From the reference period to period-2, Jialu catchment moved from Group-1 to Group-2 and five catchments moved from Group-2 to Group-3.

During period-3, the correlation between precipitation and sediment yield was weaker compared to the reference period and period-2 (Table 3). The relationships between precipitation and sediment yield were not significant in all the catchments (Table 3). The slopes of the regression lines during period-3 decreased sharply (Table 3). Six catchments (five in the northwestern part and one in the southeastern part) had negative regression slopes (Fig. 9c). This result indicates that the sediment production capability of annual precipitation decreased greatly during period-3, and the increase in precipitation amount in some catchments did not lead to increased sediment yield. Furthermore, the spatial patterns of the precipitation–sediment relationship during period-3 were clearly different from those during the reference period and period-2 (compare Fig. 9c against Fig. 9a–b). There were only three groups, with two catchments having regression slopes of 0.1 < a < 0.2, six catchments having regression slopes of 0.1 < a < 0.2, and six catchments having negative regression slopes.

The aforementioned analysis of the precipitation–sediment yield relationship in different periods clearly indicates that the impacts of precipitation on sediment yield declined with time. The impacts were different among catchments, with a clear spatial pattern. The effects of precipitation on the sediment yield were greater in the northwestern part compared to those in the southeastern part. The decreased effects of precipitation on sediment yield with time were consistent with the significant reductions of sediment coefficient (Table 2) and the decreased contribution of precipitation to sediment load reduction (25.61 and 11.33 % in period-2 and period-3, respectively). During period-2, the LUCC were mainly induced by SWCM, especially engineering measures. During period-3, the combined effects of substantial vegetation cover and conservation measures further weakened the effects of precipitation on sediment load reduction.

In order to quantify the effects of SWCM on sediment load reduction, the relationships between the decadal sediment coefficient and the fraction of area treated with erosion control measures in the 15 catchments were analyzed and the results are presented in Table 5. The decadal sediment coefficient (SC‾) decreased linearly with the fraction of treated land surface area (Ac) in all catchments: SC‾=-mAc+n. The correlations were significant in 11 catchments (p < 0.05), with R2 ranging from 0.78 to 0.99 (Table 5). The effects of SWCM on sediment load change show a spatial pattern. The correlation between sediment coefficients and conservation measures were stronger in catchments located in the northwestern part compared to that in the southeastern part (Table 5). Based on the slope of the regression equation between the sediment coefficient and fraction of the treated area, the catchments were classified into three groups in Fig. 10 (Group-1: 0.8 < m < 1.2; Group-2: 0.4 < m < 0.8; and Group-3: 0 < m < 0.4), which indicated that the degree of sediment load impacted by conservation measures was different among the catchments. The average m value was 0.73 and 0.37 for the catchments in the northwestern and southeastern parts, respectively. Half of the catchments in the northwestern part were in Group-1 and the other half were in Group-2, whereas six of the eight catchments in the southeastern part were in Group-3 with the lowest regression slope.

Differences in catchment characteristics, including land use/cover, soil properties, and topography, as well as precipitation characteristics, are clearly the reason for the spatial patterns in the precipitation–sediment yield relationship (Morera et al., 2013; Mutema et al., 2015). The lower vegetation cover was the main reason for the greater effects of precipitation on sediment yield in the northwestern part. In order to fully explore this, the mapping of information of catchment characteristics into sediment yield models and simulations under different climate scenarios would be needed (Ma et al., 2014; Achete et al., 2015). In this context, the inter-annual and intra-annual patterns of variability of precipitation, including the distribution of storm events, may also contribute to the observed spatial patterns of the precipitation–sediment yield relationship.

As LUCC took effect during period-2 and period-3, and despite the much reduced role of precipitation in driving changes in sediment yield, within-year temporal rainfall patterns did play an important role in the observed changes in sediment yield, given that most of the sediment yield was produced during a few key storm events. The correlation between sediment yield and storm events with a daily precipitation amount larger than 20 mm (including storm numbers, precipitation amount of storms) in the CSHC region during different decades was investigated (see Table 6). The analysis showed that the sediment yield was significantly correlated with storm numbers in the 1960s, 1970s, and 1980s (p < 0.05), and precipitation amount of storms in the 1960s and 1970s (p < 0.05). This result indicated the critical role of storm events in sediment yield, especially during the periods before substantial LUCC took effect.

Looking into this in more detail and taking the Yanhe catchment as an example, the precipitation amount during the rainy season (May–October when sediment load was measured) in 2003 and 2004 was 514.31 and 389.05 mm, respectively, whereas the sediment load in 2004 (2427.37 × 104 t) was about over 4 times that in 2004 (590.04 × 104 t). As shown in Fig. 11, there were 6 days with precipitation amounts over 20 mm and the maximum daily precipitation amount on 25 August was 27.85 mm in 2003, and the values in 2004 were 5 days and 46.34 mm on 10 August. Furthermore, heavy rainfall events were distributed in every month in 2003, whereas they were concentrated in July and August in 2004. There were five evident peaks of sediment load with the sum of 1646.24 × 104 t (67.82 % of annual total) in 2004, especially the one on 10 August which produced 784.53 × 104 t of sediment load (32.32 % of annual total) (Fig. 11b). In contrast, there were three peaks of sediment load in 2003, and the maximum value was only 139.97 × 104 t (Fig. 11a). Therefore, apart from annual precipitation amounts, within-year rainfall patterns should also be considered when investigating the effects of precipitation on temporal–spatial changes in streamflow and sediment load.

The sediment load reductions in the CSHC region were primarily caused by the LUCC and the implementation of SWCMs. The cropland area decreased by 9733.91 km2 (8.73 % of the region's area) and the forestland area increased by 7662.50 km2 (6.87 % of the region's area) in the region from 1975 to 2010. Most of the increase in forestland area was converted from cropland area induced by the GFG or reforestation project. As a result of the land use change, vegetation cover increased greatly, and it substantially contributed to the decreases in runoff and sediment production. The SWCMs, such as afforestation and engineering measures, were the major interventions in the study area to reduce the runoff-sediment generation from precipitation and to retain streamflow and sediment load within the catchment. Establishing perennial vegetation cover was considered one of the most effective measures to stabilize soils and minimize erosion (Farley et al., 2005; Liu et al., 2014). It was reported that both the runoff coefficient and sediment concentration of catchments in the LP decreased significantly and linearly with the vegetation cover (Wang et al., 2016). The engineering structures mainly included the creation of terrace and building of check-dams and reservoirs, which reduced flood peaks and stored water and sediment within the catchment. There were about 59 874 check-dams in the region which trapped about 9842 × 104 t of sediment per year (approximately 19 % of annual sediment yield) during the past 6 decades (Yao et al., 2011). Over time, the effectiveness of engineering measures decreased as they progressively filled with sediments, and vegetation restoration played a greater role in controlling soil erosion.