Study area

This study was conducted at the Luoyugou watershed in the third subplot of the Hilly-Gully region at the Loess Plateau of China. Luoyugou is a tributary of the Ji River that flow into the Wei River, and is located at east longitude of 105º30’ – 105º45’ and northern latitude 34º34’ –34º40’ in Tianshui of Gansu Province, China (Figure 1). Luoyugou watershed is relatively small and pinnate, with a drainage area of 72.79 km², a high symmetry coefficient of 0.9, and a low watershed development index of 1.49. The 21.81 km long main channel has 293 branches, with 96 on one side and 97 on the other side. The total length of the branches at all levels above the medium gully is 395.35 km. The cutting density is 3.54 km per km², and channels with a length of more than 1 km reach a total length of 107.28 km. The gully density at this watershed is 1.47 km per km², and the average gradient of the main channel is 2.5% (Han, 2011; Wang et al., 2008; Zhang, 2012).

Luoyugou watershed is located at beam-shaped hilly terrain, has a high elevation and steep slopes, and is cross-cut by numerous ravines and gullies. Given a high surface runoff potential with an annual runoff modulus of 3.069×104 m³/km², gully and gravity erosions are serious, with an annual soil erosion modulus of 3300–9500 t/km² (Zhang, 2012). It is one of most vulnerable areas on the Loess Plateau for soil loss and one of the main sources for sediment loads to the lower Yellow River, thus eliciting serious consequences in downstream ecosystems. As a result, the Luoyugou watershed was managed through 3-year planning of agriculture and forestry production and water and soil conservation under the guidance of China’s Ministry of Water Resources as early as 1956. In August 1983, it was listed as a pilot watershed at the middle Yellow River and a key comprehensively managed watershed in Gansu Province. Detailed hydrological, soil, climate and land use data have been collected for the Luoyugou watershed since 1950s, therefore it is a suitable small watershed for the research of the relationship between land use change and water and sediment transportation.

Input data

In order to optimize the land use pattern and subsequently its impact on surface runoff and sediment loss in the watershed, we have collected the following input data, including rainfall, stream flow, sediment data, climate data and Digital elevation model (DEM) data. Annual runoff and sediment data from 1986 to 2000 were recorded at the outlet station of the watershed (i.e., the Luoyugoukou station, Figure 1), and accessed from the monitoring data Tianshui hydrology and water resources survey. Rainfall data from 1986 to 2000 were obtained from nine stations within and adjacent to the Luoyugou watershed (Figure 1) from Tianshui hydrology and water resources survey. Tianshui county weather station provided detailed relative humanity, wind speed and temperature data at the same time (Wang et al., 2008). DEM data for the area was downloaded from the Service Platform of the International Science Data. The slope map of the watershed was obtained from the DEM data (Liu et al., 2011). Briefly, the DEM of the watershed was analyzed through the gradient tool in the ArcGIS9.3 (ESRI, America) to generate the slope data. The resolution of the DEM at 30 m×30 m was transformed into 5 m×5 m overlaid by remote sensing data through the data conversion tool of ArcGIS9.3 after projection conversion. According to the grading index of soil erosion intensity in the Luoyugou watershed (Zhang, 2012), the slopes in the watershed were reclassified into six classes, including 0º–5º, 5º–8º, 8º–15º, 15º–25º, 25º–35º, and > 35º. This slope data is critical for generating the map of land use gradient distribution.

Recently, vegetation coverage has been widely obtained from the remote sensing images through the correlation analysis of vegetation coverage and the spectral index (Los et al., 2000; Sellers et al., 1994) or using a regression model (Asrar et al., 1992; Graetz et al., 1988; North, 2002). Nonetheless, these methods are generally difficult to be implemented, require longtime observations, and have many limitations such as operation difficulty, longtime observation and low efficiency. Thus, the study chose normalized difference vegetation index (NDVI) to evaluate vegetation:

Vegetation coverage in the watershed was estimated by a Landsat Thematic Mapper (TM) remote sensing image with a 30-meter resolution taken in late August, 2001, when the cloud cover was almost zero. Radiometric, atmospheric and geometric correction were conducted on the image first, then the RS image of Luoyugou watershed was obtained by cropping the image along the watershed boundary. It is used to calculate NDVI through ENVI RS processing software. According to dimidiate pixel model, NDVI of a pixel can be expressed as NDVI_veg (contribution information of vegetation) and NDVI_soil (contribution information of bare soil), so vegetation coverage of any pixel can be express as:

The vegetation coverage map was collected in ArcGIS combining NDVI_veg, NDVI_soil and function (2). The map use in this study was reclassification to 0-0.3, 0.3-0.45, 0.45-0.6, 0.6-0.75 and above 0.75 five levels.

The land use data of Luoyugou watershed were obtained from the Landsat images taken by the Thematic Mapper (TM) in 1986 and 1995 and Enhanced Thematic Mapper (ETM+) in 2001, as well as the SPOT image in 2008. Image preprocessing was conducted on the platform of ERDAS9.2 and ArcGIS 9.3. Supervised classification and artificial visual interpretation (Gong et al., 2009) were used to classify the forth, third and second wave synthesis image and vectorization, generating the original land use map. The overall accuracy was determined around 85% through examining, repairing and establishing land use spatial database after field investigation and literature search. The land use was classified into seven types according to the national classification standard of China (GB/T21010—2007, 2007), namely, sloping farmland, terraces, forests, grassland, water, orchards and villages. At the same time, slope map and land use map overlaid to generate soil erosion intensity grading map according to different weight.

Overall, we have assembled yearly average rainfall, temperature, evaporation and wind speed in dBase as well as stream flow and sediment loads at the watershed outlet from 1986 to 2000. The vegetation coverage map, and the land use map and associated gradient distributions in 1986, 1995, 2001, and 2008 were also prepared. Considering our dataset, we have designed our experiment in the following manner: the period of 1986–2000 was used to establish our SWAT model; the land use of 2001 was used as the baseline land use on which the land use optimization was implemented as described below; the land use of 2008 was considered the future land use when no land use optimization was performed prior.

Land use optimization

Optimization of the land use pattern was mainly based on landform features, soil types and fertility, and severity of soil erosion in this area. When combined with the local social and economic development, this allows assessment of the key impact factors, and adoption of tillage measures, engineering measures and biological and other measures, and adjustment of the land use pattern in the watershed, with the purpose of reducing soil erosion and soil nutrient loss. In this study we will examine if a more elaborated scheme of land use optimization targeted at the most vulnerable areas to soil and water losses could reduce surface runoff and sediment transport in the watershed. This should enhance the sustainable use of soil natural resources, and promote sustainable development of the regional economy and the ecological environment. The land use in 2001 was intentionally chosen for the land use optimization, because severe soil erosion and surface runoff have occurred during this period. In order to optimize the land use in 2001, we prepared our own land use map (Figure 2), slope map, and gradient distribution of land use, soil erosion intensity, as well as obtained the spatial distribution of rainfall erosivity from Zhang et al. (2012), provided in the supporting information (Figures S1, S2, S3, S4, S5, S6 and S7 [online supplement]). As the main factors influencing surface runoff and soil erosion include the slope and vegetation cover conditions in the area (Li & Wang, 2007; Xu, 2002), we use land use gradient distribution map, rainfall erosivity map, and soil erosion intensity map combining with the national policy to optimize the land use with the following scheme: (1) the sloping farmland and terraces in the area with a slope above 25º are returned to forest; (2) the sloping farmland with a slope between 5º and 15º are adjusted to terraces; (3) the orchards with slopes above 25º are converted to forest; and (4) the areas with severe soil erosions on the slopes above 8º are changed to forest, and other area of severe soil erosions on the slopes less than 8º to terraces. Our primary focus is to make sure that wherever possible, the farmlands are changed to the terraces so that agricultural food production and the livelihood of rural communities could be protected. In responding to the China’s national policy of “returning farmland to forest” (1999), in our scheme only the farmland on the slopes of greater than 25º is changed to forests. In essence, the principal goal of the land use optimization is to reduce soil erosion in the watershed, which was verified by watershed-scale modeling using the Soil and Water Assessment Tool (SWAT) as described in detail in the next.

SWAT modeling

The SWAT model (version 2009, USDA-ARS, USA) was chosen for this study (Figure 3). We first verified the applicability of the SWAT model for the Luoyugou watershed. The calibration period was 1986–1994 and the input data for the model were included the land use in 1986, and observed rainfall, runoff, and sediment loads for the same period (Figures 4, 9 and S8 [online supplement]). The validation period was 1995–2000, and the model input included the land use in 1995, and observed rainfall, stream flow, and sediment loads for the same period. Other yearly climatic data such as temperature, potential evapotranspiration and wind speed in the dBase format were inputted in advance. 41 parameters such as ESCO (Soil evaporation compensation factor), EPCO (Plant uptake compensation factor), ALPHA_BF (Base flow alpha coefficient), (Table S1 [online supplement]) were adjusted automatically in the calibration period. A detailed list of calibrated parameters is provided in the supporting information.

Once the SWAT model for the watershed was calibrated and validated, three scenario analyses were performed to assess the influence of land use change on the surface runoff and sediment yields in the watershed, using the original land use in 2001, the optimized land use, and the land use in 2008, respectively. In order to assess the impact in a long period, these three scenario analyses were performed over a 15-year span using the input data from 1986–2000 while replacing the land use data with one of three above-mentioned land uses. The simulated surface runoff and sediment load in the SWAT calibration and validation were used a baseline values for comparison with the results of the above three scenario analysis. The objective was to assess the responses of surface runoff and sediment loads to the scenarios whether or not the land use optimization was implemented.

Because the change of surface runoff and sediment loads could also be influenced by climate change in addition to land use change, the last exercise dealt with the simulations with invariant climate input so as to exclude the effect of climate forcing. Because the similarity of annual rainfall in 1997 and 2008, the climate input in 1997, including rainfall, air temperature, potential evapotranspiration, etc., was used to simulate surface runoff and sediment loads using our developed SWAT model for the watershed with either the optimized land use or the actual land use in 2008. The simulated results were compared with each other and with the observed surface runoff and sediment load in 2008.

Land use change and optimization

The land use patterns in 1986, 1995, 2001, and 2008 in the Luoyugou watershed are shown in Figure 2, and the slope map of the watershed is shown in Figure S2 [online supplement] at the supporting information. Given that no major earthquakes or landslides able to significantly alter the terrain slopes were known to have occurred in the past 15 years, we overlaid the derived slope map (Figure S2 [online supplement]) and the land use map in each year to generate the land use gradient distribution maps in order to determine the terrain slope of a particular land use type (e.g., sloping farmland, terraces, orchards, and forests). A representative land use gradient distribution map and associated distribution in 2001 is shown in Figures 4 and 5, respectively, and similar maps for other years are provided in the supporting information.

From Figure 2, it was clear that a dramatic change in the land use patterns occurred during the period of 1986–1995 (Figure 2a and b). In 1986 the sloping farmland constituted the largest area of 49.28 km2, representing 68.82% of the watershed area (Table S2 [online supplement]). Its main slope distribution was in the range of 8–15º, which covered 28.24 km², representing 39.44% of the watershed area. The area with a slope greater than 25º was 0.96 km², representing 1.35% of the watershed area. The combined area of terrace, forest and grassland was relatively small at 17.34 km² (i.e., 24.22% of the watershed area), which was only equivalent to 35.2% of the sloping farmland area. Consequently, the area was prone to severe soil erosion and rapid surface runoff. It concluded in previous study that runoff and sediment from 1960s to 1990s in Jing River basin that both loads reduced in 1970s, increased from 1981 then decreased in 1990s (Liu, 2011). The results were consistent with the available data in this study. The reason was probably the national police. Before 1980s, the people’s commune collective ownership was executed in this region that developed large-scale construction of water conservancy works such as terraces; farmland contracted responsibility system took the place after 1981, many old terraces were out of repair or influenced by graze and farming which led to terraces reduction (Chen & Yang, 2001). After that, as a soil and water conservation practice, the conversion of sloping farmland to terraces (termed as terrace engineering) took place from the middle of 1980s (Jing, 2001; Jing & Jiao, 2010). As a result, the terraces became the largest land use type by 1995 (Figure 2b), with an area of 39.04 km2 (i.e., 54.52% of the total watershed area) (Table S3 [online supplement]). Its main slope distribution was still in the range of 8–15º, which covered 23.74 km², representing 33.16% of the watershed area. The area with a slope greater than 25° was 0.19 km², representing 0.27% of the watershed area. The combined area of sloping farmland, forest and grassland was relatively small at 27.4 km² (i.e., 24.22% of the watershed area).

From 1995 to 2001, the area of terraces remained relatively unchanged. In 2001, the terraces were mainly distributed in the slope range of 8–15º, which covered an area of 23.73 km² or 33.14% of the total watershed. The area with a slope above 25º was 0.19 km², covering 0.27% of the total watershed. The combined area of forest, orchards and grassland was 18.35 km², representing 25.63% of the watershed area, and less than half of the terraced area.

Nonetheless, by 2008 there were still some mismanaged areas of land use in 2008, with 0.68km2 sloping farmland and 0.38km² terraces in regions where the slopes are greater than 25°, 0.96% and 0.54% of the total area. (Table S4 [online supplement])The area of sloping farmland was still large with an area of 9.83km², 13.45% of total area. Thus, despite the enormous improvement from the terrace engineering since the middle of 1980s (Jing, 2001; Jing & Jiao, 2010), in order to curtail the surface water and sediment loss, more targeted optimization of land use patterns is needed. According to the optimization scheme described previously, the land use pattern in 2001 was optimized, as shown in Table 1 and Figure 6.

Table 1. Land use gradient distribution after optimization of the land use pattern

After optimization, the slope farmland had a scope gradient of 0º–8º with a size of 2.10 km², i.e., 2.93% of the watershed area. All terraces with a total area of 47.08 km² were located at the regions with a slope less than 25º, which not only ensures food production but also reduces soil and water loss through small terrain changes of terraces. Compared with the land use in 2001 prior to optimization, the sloping farmland was reduced by 12.55%, the terraces increased by 11.29%, and the forests increased by 1.32%. The increased forest areas were mainly located mainly in the regions with a slope greater than 25º (Table 1 and Table S5 [online supplement]). There were minor changes in the sizes of orchards and grassland, whereas the sizes of villages and water remained unchanged (Figure 7). Obviously, the optimization was mainly involved the sloping farmland, the terraces, and the forests as these land uses are the major types in the watershed.

For the pure purpose of scenario analysis, we have chosen the land use pattern in 2001 as the current land use, while the land use in 2008 as the future land use. There was little land use change from 2001 and 2008, unless the optimization of the land use was implemented. Next we will examine how the optimization of land use patterns influences the surface runoff and sediment loads.

The optimization also corresponded to former researchers’ results. It is indicated that land use changes had a deep influence on runoff and sediment (Xie, 2009). Tang studied sediment and runoff of a small watershed of Loess Plateau by using non-parametric Mann-Kendall test, Moving t-test technique, and hopped parameter analysis, which had a result of land use/cover changes accounting for 90.11% on sediment (Tang et al., 2010). Moreover, expect for vegetation coverage, slope also played an important role in water and soil conservation (Li & Wang, 2007). Terrace had a significant advantage over sloping farmland on water and sediment reduction, which lead to slope-to-terrace taking an important part in land use optimization.

SWAT development in the Luoyugou watershed

First, we developed the SWAT model for the Luoyugou watershed. As previously described, the calibration period was 1986–1994 using the land use in 1986 (Figure 2a), whereas the validation period was 1995–2000, using the land use in 1995 (Figure 8b).

As shown in Figure 8 and Table 2, for surface runoff, the certainty coefficient (Ens) was 0.88, the relative error (Re) was 5.10%, and the correlation coefficient (R²) was 0.94 in the calibration period, whereas in the validation period, the Ens value was 0.87, the Re value was 7.20%, and the R² value was 0.98. Similarly, the sediment simulation was also satisfactory, showing decent Ens, Re, and R² values for both the calibration and validation periods. Therefore, we assumed that the SWAT model has captured the underlying hydrological processes in the watershed, which was in agreement with the study on a similar nearby watershed (i.e., the Zhifanggou watershed) (Qiu et al., 2012). Nevertheless, it may possibly misestimate the extreme event. In the study of the application of SWAT in the River Grote Laak in Belgium, the SWAT model underestimated the base flow during an extreme dry period in 2003 (Nossent & Bauwens, 2007); while the research in Zhifanggou gave another result that the model might underestimate runoff in wet season and overestimate it in dry period because it is difficult to determine the soil moisture conditions before and during high-flow events (Qiu et al., 2012). Hence, it is necessary to improve the parameters and corresponding performance in extreme event simulation in the future study of SWAT model. And as our calibration and validation were performed on the yearly basis, new calibration and validation will be needed for simulations in daily or monthly time-steps.

Table 2. Comparison of observed and simulated surface runoff and sediment loads in the Luoyugou watershed ^a

Impact of land use optimization on surface runoff and sediment loads

In order to evaluate the impact of land use optimization on surface runoff and sediment loads over a long term, we first performed the simulations for the period of 1986−2000 using the land use in 2001 while keeping other input unchanged for each year. Then, we replaced the 2001 land use by the optimized land use and the 2008 land use, respectively, so as to assess the responses of surface runoff and sediment loads to the land use optimization and to no designed land use optimization (i.e., the 2008 land use).

As shown in Figure 9, the runoff in each year was decreased significantly after the optimization of the land use pattern in the Luoyugou watershed. In the SWAT model development for the Luoyugou watershed, the average simulated annual runoff for the period of 1986–2000 was about 0.074 m³/s, which serves as the baseline value for comparison. When the 2001 land use data or the 2008 land use data were used for the period of 1986–2000, the simulated average annual runoff was 0.072 m³/s or 0.070 m³/s, respectively (Figure 9), suggesting that little reduction in surface runoff would occur without more targeted land use optimization. Conversely, when the optimized land use pattern was used, the simulated average annual runoff was 0.061 m³/s, i.e., a reduction of 17.6% from the baseline value (i.e., 0.074 m³/s). Compared with the scenario that the land use optimization was not implemented (i.e., the 2008 land use), the runoff was decreased by 12.9% (i.e., 0.009 m³/s).

Similarly, the sediment loads were remarkably reduced with the optimization of the land use pattern compared with the simulated results using the 2001 land use (Figure 9). Again the simulated baseline sediment load was 32.01 million ton per year averaged over the period of 1986–2000. If the land use optimization was not implemented, the simulated sediment load was 31.35 million ton per year or 30.42 million ton per year for the 2001 land use or the 2008 land use, respectively. This observation suggests that little reduction in the sediment transport would occur if the current practice remains unchanged. However, if the land use optimization was implemented (i.e., using the optimized land use pattern for the simulation over 1986–2000), the average simulated sediment load was decreased to 26.33 million ton per year, i.e., a reduction of 17.74% from the baseline value. Our optimization scheme of land use pattern mainly focuses on converting the sloping farmland to the terraces, while other researchers proposed more aggressive strategies by converting the farmland to the forests. Expectedly, the reduction in surface runoff and sediment loads in this study were lower than the results of Wang et al. (2006). Another assumption in the Luoyugou watershed gave a result that when forest coverage greater than 80% of the total area, the sediment would be similar with our reduction (Xie, 2009). Nonetheless, given that farming is the livelihood of many rural communities and food security is of paramount importance, our approach balances the need of conserving water and soil resources while maintain the food production area.

Further validate of land use optimization

Given the significant driving forces of rainfall, and possibly other meteorological and climatic factors in impacting surface runoff and sediment transport, we further performed the scenario analysis using the invariant meteorological-climatic input. Because the annual rainfall was 378.6 mm in 1997 and 380.1 mm in 2008, the similarity of rainfall between the ends of this decadal period provides a unique opportunity to pinpoint the dominating role of land use change by excluding the climatic variations. Subsequently we used the rainfall and other climatic input from 1997 to simulate surface runoff and sediment loads using our developed SWAT model for the watershed with the optimized land use. According to the 2008 communique data, the observed runoff and sediment load was 681700 m³ per year and 73090 t, respectively (Table 3). When using the optimized land use, the runoff and sediment loads were decreased by 68.59% and 31.13% compared with the observed values. The results suggest that: first, the land use optimization was responsible for the reduction in surface runoff and sediment loads despite the potential role of climate forcing; second, the runoff decreased by a large proportion makes water conservation easier inside the watershed.

Table 3. Comparison of the simulation data for the adjusted and optimized land use with the observed data from 2008