Multiscale attribution analysis for assessing effects of changing environment on runoff: case study of the Upstream Yangtze River in China

Evaluating the changes in runoff and analyzing its attribution under the changing environment is of great significance to water resources management. In this study, eight hydrological stations at the outlets of tributaries of the Upstream Yangtze River are selected. Based on the observed runoff data from 1951 to 2013, the spatial-temporal characteristics in runoff change are identified from time series analysis. Our results show that runoff in the Upstream Yangtze River decreases significantly with a rate of 7.6 km per ten years in general. The most significant declines in runoff are observed in the mainstream, Minjiang River, Tuojiang River, and Jialing River, while slight increase in runoff is found in the source area of the Yangtze River. Furthermore, the effects on runoff change from climate change and human activities are evaluated using the Soil and Water Assessment Tool (SWAT) and modified Fixing-Changing (MFC) method at multiple scales. Our results suggest that the main contributions to runoff change are from climate change variabilities (70%), land use/cover change (LUCC, 10%), and other human influence (20%). When examined at different spatial and temporal scales, climate change always appears to be the main cause of runoff change, although its contribution decreases over time.


GRAPHICAL ABSTRACT INTRODUCTION
Affected by the changing climate and intensive human activities, uncertainty of water resources is deepened. The impact of changing environment on the safety of water resources has become a major issue of global concern.
Assessing the changes in water resources under the changing environment and identifying the causes is critical to water resources planning and management, which have become a hotspot in water science (Hall et al. ).
Among different components in the hydrological cycle, runoff is considered as the most important one for water resources management. Its variability significantly affects the water use pattern in different sectors (Dey & Mishra ).
Runoff is a highly nonlinear process and closely related to many factors, such as atmospheric circulation, climate change, underlying surface, and human influence. As the result of multiple factors in the changing environment, changes in runoff are complex, uncertain, and hard to predict. Generally, researchers use long-term data to quantify its variability and analyze the attribution of runoff change, as a function of climate change and human activities.
Methods such as hydrological time series analysis are usually applied in such studies, to examine the discrepancies of the hydrological cycle in different regions and to understand the underlying physics (Burn & Elnur ). and other disciplines increasing significantly, and with more attention paid to the interactions and feedback effects of climate, land, and human influence (Wang ). Generally, the impacts of climate change on runoff are mainly illustrated by: (1) impacts of precipitation including its form, amount, intensity, process, and spatial distribution; (2) glacier and snow melting caused by rising temperature; (3) evapotranspiration changes caused by changes in climatic factors (e.g., temperature, wind, and humidity) (Wang et al. ; Yang et al. ).
In recent years, studies on the impacts of human activities on the hydrological process have also been conducted.
Impacts of human activities on runoff are mainly illustrated by: (1) impact on condition of runoff generation and confluence (e.g., land use/cover change (LUCC), water and soil conservation, river regulation); (2) direct influence by humans (e.g., irrigation, water diversion projects, groundwater exploitation); (3) impact on the process of runoff confluence (e.g., reservoir impoundment) (Zhang et al. ).
Most studies focus on the hydrological effects caused by LUCC. With the rapid development of the global economy, urbanization and accelerated agricultural development result in atrophy of wetlands, and decreasing area of forest and grass. LUCC has affected many hydrological processes including vegetation interception, evaporation, and infiltration. It has influenced the conditions of runoff generation and confluence, and has significantly affected the runoff process, resulting in a series of eco-environmental problems (Yang et al. ; Zhang et al. ). Techniques for studying LUCC effect include the experimental catchment method, the characteristic variable time series method, and the hydrologic model method.
Hydrological models aim to characterize the hydrological processes based on the understanding of physical mechanisms in the hydrologic cycle. A collection of bottom-up or physically based models have been applied in hydrological studies, which have detailed and high resolution descriptions of small-scale processes that are numerically integrated to larger scales (e.g., catchments) (Hrachowitz & Clark ). With the information obtained from geographic information systems (GIS) and remote sensing (RS), bottom-up models can characterize rainfall, evapotranspiration, topography, soil, vegetation, and other underlying surface conditions spatially. They not only describe the general pattern but also provide detailed information on hydrological processes within a catchment and, therefore, have been used for evaluating the hydrological effects of climate change and LUCC. These models are often used to simulate and predict runoff processes, including extreme events such as floods and droughts. Researchers drive the models by observed daily maximum and minimum air temperature and precipitation to generate daily soil moisture values, then calculate the Soil Moisture Anomaly Percentage Index (SMAPI), which can be used as a measure of the severity of agricultural drought on a global basis (Wu et al. ). These models are essential for understanding the attribution of runoff change to the changing environment and, in particular, are important in light of the increasing effects of changing environment on the terrestrial water cycle.
China is one of the most prominent countries regarding the contradiction between water supply and demand (Piao et al. ). Due to the influence of the changing environment, the spatial and temporal characteristics of water resources have changed significantly. In the past five decades, runoff has decreased significantly in six major rivers in China (i.e., the Yangtze River, the Yellow River, the Pearl River, the Songhua River, the Haihe River, the Huaihe River). The Yangtze River provides the most abundant water resources and hydropower resources in China, with an average annual runoff of 996 billion m 3 , on average, making up 36% of the total runoff in China, and its theoreti-

Study area
The Yangtze River is the largest river in China, with a total length of 6,380 km and a total elevation drop of 5,400 m. As shown in Figure 1 Detailed information on these basins is listed in Table 1.     Figure 1).

Topographic data
The topography is represented by the digital elevation model

Soil type data
A soil type map with a scale of 1:1,000,000 was obtained from the RESDC (http://www.resdc.cn) ( Figure 4).
soil water content, and has been certified as an effective tool for evaluating water resources at a wide range of scales. In this study, the SWAT model is applied for the Upstream Yangtze River. Based on the DEM data, the river system is generated and 99 subbasins are divided as shown in Figure 5.
After inputting the soil type data and land use/cover data in . The equations are given as follows: where obs i is the observed value on time i, obs is the mean observed value, sim i is the simulated value on time i, sim is the mean simulated value, and n is the length of the time series.

MFC method for attribution calculation
The Fixing-Changing (FC) method is commonly used for attribution calculation. The general idea of this method is to: (1) fix the land use/cover inputs and change the meteorological inputs, to analyze the runoff change caused by climate change; (2) fix the metrological inputs and change (1) Divide the whole study period into N þ 1 (N ! 1) segments based on the previous time series analysis.
(2) Fix the land use/cover data and meteorological data in N), and simulate runoff series twice by the SWAT model, with meteorological data in segment iÀ1 and segment i , respectively. These runoff series are denoted withW(L iÀ1 , C i À1) and W(L iÀ1 , C i ), where (L iÀ1 , ) means land use/cover data in segment iÀ1 , (Á, C iÀ1 ) and (Á, C i ) means meteorological data in segment iÀ1 and i .  (4) The attribution can be calculated as follows: where ΔW C,i and ΔW L,i indicate the runoff change in segment i , caused by climate change and LUCC, respectively.
(5) The remaining runoff change is attributed to other human influence and calculated as: where ΔW O,i indicates the runoff change caused by other human influence in segment i , ΔW T,i indicates the total runoff change in segment i .

Runoff change
Historical time series of annual runoff at eight selected hydrological stations are examined in Figure 7, and the changes are assessed using the Mann-Kendall test and the Spearman's Rho test in Table 3 Table 4. The average annual runoff in the change period decreased by 28.7 km 3 (8.4%) compared to that in the base period. Statistically significant differences in the annual runoff between the two periods can be found with a confidence level of 95% using the t-test, while differences in the standard deviations are not statistically significant when evaluated from the F-test. In addition, the coefficient of variation is 0.1 in the base period, which increased to 0.13 in the change period, indicating that the runoff varies more after 1993.

SWAT model calibration and validation
Based on the sensitivity analysis, highly sensitive parameters are screened out for calibration, with their descriptions and initial ranges listed in Table 5. Since runoff in the SA is closely related to snowmelt, parameters of snow properties  Table 5.
Simulated and observed runoff during the calibration and validation periods are plotted for the control hydrological stations in Figure 9. The R 2 and the NSE values are calculated for the hydrological stations in both calibration and validation periods and are listed in

Climate change
The Mann-Kendall test and the Spearman's Rho test are also used to analyze changes in the annual meteorological series including annual precipitation, mean air temperature,   precipitation between the two periods, and difference is found in sunshine duration at a confidence level of 80%.
Using the F-test, a significant difference at a confidence level of 99% is detected in standard deviations of maximum temperature between the two periods, as well as that in the

Details of 21 large reservoirs built in the Upstream Yangtze
River are listed in Table 9. It can be seen that 19 reservoirs began to be operated in the change period of this study, with a total capacity of 106.7 km 3 , which directly reduces runoff.
Moreover, reservoir impoundment can also lead to an increase in human water consumption, and the loss is considerably large, although most of them will be returned to runoff. The survey shows that the annual water loss in the Three Gorges Reservoir caused by the leakage of the dam foundation was 1.8 × 10 6 m 3 which, in general, cannot be completely converted into runoff. Moreover, most of the other reservoirs in the Upstream Yangtze River do not have such good closure conditions and geodetic structures.
According to planning, a complex water conservancy and hydropower system that contains more than 100 large reservoirs will be built in the Upstream Yangtze River.

LUCC
The land use/cover conditions are shown in Figure 3, with the areas of land use/cover types in the five different periods listed in Table 10. In the Upstream Yangtze River, the major land use/cover types are grass land, forest land and crop land, accounting for more than 90% of the total area.
Changes in the areas of different types range from À4,000 to 4,000 km 2 . Evaporation over water surfaces is generally greater than that over land surfaces due to the increased water area. Since the impoundment of the Three Gorges Another 18 large reservoirs and more middle-to-small reservoirs were also built in the change period. As shown in Table 10, water surfaces increased by 4,000 km 2 from 1990 to 2010. Mean annual actual evapotranspiration in the Yangtze River basin is 520 mm and mean annual pan evaporation is 1,400 mm (Xu et al. ). What is more, actual evaporation from an airy and open water surface is greater than the pan evaporation. Therefore, increase in actual evaporation caused by the increased water surfaces can be estimated at 1 m per year. Based on these calculations, the increased water surfaces can result in 4 km 3 more water loss in runoff.

Attribution calculation
The runoff change caused by climate change, LUCC, and other human influence is calculated and listed in         1994-1997, higher in 1998-2005, and lower again in 2006-2016. Therefore, the change period is further divided into three segments (i.e., 1994-1997, 1998-2005, and 2006-2016) for a gradual attribution.
where ΔW 1 , ΔW 2 , ΔW 3 , ΔW 4 represent runoff change caused by meteorological factor 1, 2, 3, 4; for simplicity, E i represents simulated runoff under E i . It can be extended to n-factor situation: where ΔW j represents runoff change caused by meteorological factor j; n represents the number of factors; for simplicity, E i represents simulated runoff under E i ; α i,j represents the weight coefficient (À1 or 1). Contributions of precipitation, mean temperature, mean wind speed, mean relative humidity, and sunshine duration are calculated and listed in Table 16. Precipitation is found to be the lead-

Significant runoff changes have occurred in the Upstream
Yangtze River under the changing environment. In this À1 represents using the data in base period as model input; 1 represents using the data in change period as model input.
study, the spatial-temporal characteristics of runoff change are identified. Causes of runoff change are explored and the attribution is calculated. The findings are concluded as follows: (1) During the study period, runoff decreases significantly by À7.6 km 3 /10 a in the Upstream Yangtze River.
Spatially, runoff decreases in most tributaries except in the Yangtze River source area. Among the rivers, the Minjiang River, the Tuojiang, and the Jialing River decrease significantly. In both the base period (before the year 1993) and the change period (after 1993), the mean runoff in the Upstream Yangtze River is found to be significantly different.
(2) Causes of runoff change are explored and evidence is