《1. Introduction》

1. Introduction

Changes in climatic conditions and land use patterns have been confirmed to have a significant influence on the spatial and temporal distribution of water resources within specific watersheds [1– 3]. Such changes affect relevant hydrologic cycles by altering precipitation, evapotranspiration (ET), and the corresponding streamflow regimes [4], which potentially causes a reduction in water availability and an increase in water conflicts [5,6]. As noted above, many studies have been conducted to understand and analyze the hydrological cycles under climate change and human disturbance [1–7]. However, many challenges arise because of the complexities associated with certain hydrological cycles, particularly for the core process of the cycle, that is, streamflow. Generally, the identification and projection of relevant streamflow is considered an essential prerequisite for hydrologic cycle analysis. Streamflow projections under changing conditions are thus regarded as an important process that can contribute to water facility management, reservoir regulation, and water resource planning. Consequently, identifying the dynamic characteristics of hydrological processes under climatic and land use variations are critical for evaluating the usability of water resources and developing adaptation strategies [8].

Previously, hydrologic models integrated with climatic and land use projections were adopted to evaluate the possible impacts of climate change on hydrological processes under multiple scenarios [9–11]. Wilby and Harris [12] evaluated the contribution of climatic factors to the dry seasonal phenomenon of the Thames watershed using two sets of parameters with four sets of global climate model (GCM) outputs and two hydrologic models under two emission scenarios. Ahiablame et al. [13] designed several scenarios of land use and climatic patterns to analyze the corresponding possible streamflow responses. Morán-Tejeda et al. [14] combined a regional climate model and multiple hypothetical land use scenarios with hydrologic models in a mountainous watershed. Several studies have also revealed that climate change is one of the dominant factors leading to streamflow variations [15]. The hydrologic models were commonly calibrated with historical data and then driven by future predictions of climate and land use variations for hydrologic studies. As the environmental changes are increasing significantly, it is predicted that future environmental factors will lead to dynamic changes in hydrological model parameters. Compared with indirect impacts of climate change on soil characteristics, land use variation has a much more significant impact by directly and quickly altering the soil porosity and connectivity [16]. As for land use variation, Merz et al. [17] pointed out that model parameters can be considered as fixed within five years, and Luo et al. [18] stated that hydrologic simulation should consider the different model parameters in different seasons, and then the indirect impact of land use change by changing the soil properties on the hydrologic process should be considered. 

Generally, GCMs are widely used as primary tools for predicting future climatic changes. However, various uncertainties are associated with climate projections generated by GCMs owing to the complexity of climatic systems. The downscaling of GCMs to a suitable scale is frequently used to represent spatial heterogeneity [19]. Two common methods of dynamic and statistical downscaling are typically used for GCM projections [20]. Statistical downscaling aims to build statistical relations between historical GCMs and observations while there are adequate data, which have wide application in the related fields of hydrologic research [21– 23]. The uncertainties [23,24] in hydrologic forecasting will increase because of the cumulative impacts of various options of multiple downscaling methods and GCMs. Multi-ensemble GCMs have been proven to be superior to an individual GCM, two methods such as long short-term memory (LSTM) and multi-model ensemble mean (MMEM) were usually adopted for ensembling GCMs [25–28]. There is little agreement on metrics to distinguish between ‘‘good” and ‘‘bad” models, and an ensemble of multiple climate models generally outperformed any individual model; this can improve the reliability and resolution of future climate projections [29–31]. Using machine learning models to ensemble or average multi-GCMs can reduce bias in projecting future hydrological responses [32].

Furthermore, there exist only a few models to project the land use variation because of the spatial characteristics and diversity of land use types. The current common methods include the cellular automata (CA) model, the Markov model, and the conversion of land use and its effects at small region extent (CLUE-S) model. The CLUE-S model has been widely used in small-scale applications, but there are limits on the number of types and grids [33]. The Markov model was primarily used to predict the temporal resolution, but it cannot present the spatial distribution of land use variations. This model is usually used along with other models to develop its advantages [34]. The CA model can identify the spatial dynamic evolution process for complex systems, but it is difficult to effectively reflect macro influences, such as social and economic factors. The CA–Markov model has a great advantage as it combines the powerful spatial computing ability of the CA model and the long-term prediction of the Markov model for spatial and temporal pattern stimulation, and is widely used for land use projections [35]. Moreover, the CA–Markov model contains the effects of natural, social, economic, administrative centers, transportation, and climate conditions on land use projections. However, most studies of the CA–Markov model have focused on large cities, and there are few studies on land use projection at the watershed scale, especially in mountainous areas.

The Anning River Basin (ARB) is a tributary of the upper Yangtze River and an ecotone of agricultural and animal husbandry. Because of its unique topography characterized by undulating terrain, high altitude, steep slopes, and valleys; the local ecosystem is inherently fragile. There exists an uneven seasonal and regional distribution of precipitation, strong storm intensity in the rainy period, little precipitation, and large ET in the dry season [36]. The climatic variation in the Yangtze River Basin has been analyzed in numerous previous studies using multiple methods and technologies [37]. An assessment of the hydrologic process under the combined influence of climate and land use should be conducted, particularly for fragile arid or semi-arid regions in the Yangtze River.

With the increase in complexity and uncertainty of changing conditions, the investigation and understanding of hydrological progress in ARB are challenging because of the following: ① the scarcity of hydrological modeling in a fragile region with complex terrains, ② finding an applicable technique for climatic projections among various scenarios, ③ spatial and temporal variations of land use for the future, and ④ diversity between simplified hydrologic models and nonlinear hydrologic processes [30]. It is a critical scientific issue that must be addressed immediately to consider the influencing factors of climate and land use under multiple spatial and temporal scales for long-term hydrologic forecasting in the dry–warm valley of the Yangtze River.

Therefore, the objectives of this research are ① to evaluate the feasibility of MIKE SHE in a mountainous area, ② to predict precipitation and temperature in the ARB with an ensemble of 31 GCMs to improve accuracy through LSTM and MMEM; ③ to explore land use projections in consideration of natural, social, and economic conditions, and ④ to investigate the underlying influence of climatic and land use variation on streamflow and actual ET. The developed model incorporated an ensemble of 31 GCMs for climate projections and a CA–Markov-based model for land use projections, which can not only resolve meteorological and land use projection uncertainties, but also investigate the dynamic variation condition and range of hydrologic processes. Consequently, this is of great significance in dealing with stochastic uncertainty and the optimization management of water resources under complex changing conditions. This work provides benchmarked information on adaptive measures for water resources and environmental sustainability in the upper Yangtze River for local and downstream areas.

《2. Overview of the study site》

2. Overview of the study site

The dry–warm valley in the Hengduan Mountain has a unique climatic, geographical, and hydrological features. It has a typical non-zonal climate, not only based on global climate change, but also closely related to local physical conditions. Zhang et al. [38] found that the climate in the dry–warm valley is inconsistent with global climate change; the temperature has declined since the 1950s, while the precipitation increased. As a typical potential ecological degradation region and an important agricultural production base in China, hydrologic analysis under changing conditions is helpful for decision-making regarding development in the cropping system, biological resources, and ecological restoration.

The Anning River watershed, located in the dry–warm valley of southwest China, is a secondary tributary of the upper Yangtze River (Fig. 1). The range in latitude is between 102°06' E and 102°10' E and in longitude between 26°38' N and 29°02' N, with an elevation of 900–4750 m. It has a total length of 337 km and flows through Mianning, Xichang, Dechang, and Miyi, with a valley area of approximately 11 150 km2 . For several years, the average precipitation was 1240 mm, while the average temperature was 17– 19 °C. There were distinct wet and dry seasons; the rainy season was from June to October and accounted for more than 90% of the annual rainfall. Intense rainfall, along with steep slopes and valleys, usually leads to flash floods. Alternatively, there is usually little precipitation and large evaporation during dry seasons. In both wet and dry seasons, precipitation exhibits significant spatial variability, with more rainfall in mountainous areas and less rainfall in the valleys.

《Fig. 1》

Fig. 1. The Anning River watershed. Jinsha River is the upper stream of Yangtze River. DEM: digital elevation model.

As the mother river of Liangshan State, ARB is a gathering place for 14 national minorities. It is also the second-largest valley plain in Sichuan Province and an ecologically fragile region in the upper Yangtze River, which is a vital area for western economic development in China. Over the years, because of the influence of global climatic and land use variations, many eco-environmental problems, such as water shortages, land desertification, and water quality deterioration, have become serious obstacles to the development of ecological economics in southwest China. There are relatively concentrated social and economic activities, planning and construction of water conservancy and hydropower projects, large irrigation districts, and major cities, which have brought increasing uncertainties and difficulties to the system management of the watershed. Therefore, it is an urgent issue to analyze hydrologic responses of climate and land use variations for the rational use of water resources, constructing ecological barriers, and sustainable development for the Yangtze River.

《3. Methodology》

3. Methodology

《3.1. Integrated modeling framework》

3.1. Integrated modeling framework

The main research objective is to project the hydrologic response to climate and land use variation using a hydrologic model. A flowchart of the hydrologic response projection is illustrated in Fig. 2. First, according to the availability and continuity of the meteorological and hydrological data, MIKE SHE was calibrated from 1977 to 1983 and validated from 1984 to 1986. Land use data from 1980 and the daily streamflow data at Wantan station were used for both calibration and validation. Subsequently, MIKE SHE was further validated from 2006 to 2007 with the land use data of 2005, and the sensitivity parameters remained unchanged, the performance of MIKE SHE remained well. Then, two climate ensembles of 31 GCMs based on LSTM and MMEM were developed to predict future climatic changes. The LSTMbased ensemble was trained using historical data from 1961 to 2000 and was tested using data from 2001 to 2010. Subsequently, the CA–Markov model combined with multi-criteria evaluation (MCE) and spatial multiple regression (LogisticReg) modules were used to project the three types of land use projections. Finally, we employed four sets of climate data projected by LSTM and MMEM and three sets of land use projections to drive MIKE SHE to evaluate future hydrologic responses in ARB.

《Fig. 2》

Fig. 2. Diagram for hydrologic response projection under climate and land use changes. NSE: Nash–Sutcliffe efficiency; RCP: representative concentration pathway; MCE: multi-criteria evaluation; LogisticReg: LogisticRegression.

《3.2. Hydrologic modeling 》

3.2. Hydrologic modeling 

A typical representative of the distributed hydrologic model based on the physical mechanism, MIKE SHE, was selected to explore the impacts of climate and land use changes on the hydrological process [39]. The model mainly considers hydrological processes such as plant interception, ET, overland flow, unsaturated and saturated zones, surface water, groundwater exchange, and snowmelt runoff, which can be estimated from the measured data as the parameters have physical significance [40]. The MIKE SHE model is mainly applied in the following four aspects: ① exploring the applicability of the model and the model construction process [41], ② parameter sensitivity analysis [42], calibration and verification [43], and model uncertainty analysis [44];③feasibility analysis of constructing distributed hydrologic models in areas with insufficient data [45]; and ④ hydrologic responses to climatic and land use variation under the background of global warming [46]. MIKE SHE is a hydrologic model that has been universally used in multiple fields in numerous scales and has great application potential in humid and arid regions [47]. Additionally, researchers can choose suitable modules for modeling along with different climatic conditions, geological characteristics, and available data in the study region, which are different from the distributed hydrologic models such as the soil and water assessment tool (SWAT), variable infiltration capacity (VIC), and TOPMOEL [48].

3.2.1. Input data of hydrological model

The digital elevation model (DEM), meteorological, land use, soil, and hydrological data were used to characterize the spatial heterogeneity for hydrological modeling and analysis (Table 1). Three meteorological stations (Mianning, Xichang, and Miyi) located upstream of the outlet station were selected. The National Meteorological Center acquired downscale precipitation, high temperature, and low temperature-based stations from 1961 to 2010. The available Food and Agriculture Organization (FAO) Penman– Monteith method was used to evaluate the reference ET in this study [49]. Daily discharge observation values at the Wantan hydrologic station from the People’s Republic of China Hydrologic Yearbook were used to calibrate and validate the MIKE SHE. DEM and land use data were acquired from the data platforms of the Chinese Academy of Sciences, whereas soil data were obtained from the Harmonized World Soil Database↑ . The Geographic Information System (GIS) tool was used to process and obtain spatial data information such as flow direction, flow concentration, drainage network, soil, and land use data in the study area.

↑ http://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonizedworld-soil-database-v12/en/.

《Table 1》

Table 1 The input data for MIKE SHE model.

Additionally, it is difficult for the static parameters of the hydrological model to reflect the changes in the hydrologic characteristics of the future environment. Vegetation coverage is one of the most critical environmental factors in the watershed, and it directly affects the distribution of rainfall between evaporation and runoff. For this study, the leaf area index (LAI) and root depth (RD) of vegetation were used as the characterization factors for the dynamic changes in model parameters, and the vegetation growth period and non-growth period were divided. Subsequently, based on climate change conditions such as precipitation and temperature, the functional relationship between LAI and RD was established and analyzed, and the dynamic parameters of LAI and RD for every decade in the future were estimated. The dynamic parameters of the hydrological model established the response relationship between the model parameters and vegetation changes.

3.2.2. Calibration and validation

The parameter calibration process adjusts the model to match the simulation results with the observed data. Overparameterization occurs if all the parameters are freely adjusted during the calibration process. The primary sensitivity parameters are listed in Table 2, based on a literature review [50,51]. The auto calibration tool of the MIKE SHE model was adopted to calibrate the parameters (Table 2).

《Table 2》

Table 2 Sensitive parameters of MIKE SHE for calibration.

The Nash–Sutcliffe efficiency (NSE) coefficient and coefficient of determination (R) were used as statistical criteria to assess the MIKE SHE model performance in this study. The Nash–Sutcliffe coefficient is sensitive to both the mean and variance of the simulated and measured values. The coefficients of the NSE and R are as follows:

where Qm is the observation value, Qs is the simulation value, and  are the average value of Qm and Qs, respectively. n represents the number of observation values (simulation values), and i represents the nth value. 

The NSE value ranges from 0 to 1; the closer it is to 1, the better the simulation effect, and the higher the simulation accuracy. Generally, the model can be used for calculation when R of the simulated value and the measured value is greater than 0.6, and the simulation results reach a better level when the NSE value is between 0.50 and 0.95.

《3.3. Multi-ensemble GCMs for climate projection》

3.3. Multi-ensemble GCMs for climate projection

In this research, 31 GCMs at a monthly scale covering the time period 1961–2050 were downloaded from the Coupled Model Intercomparison Project Phase 5 (CMIP5) website (Table 3). The statistical downscaling model of a weather generator (NWAI-WG) developed by Liu and Zuo [52] was used to obtain data every day for each meteorological station. The dataset covers three variables: precipitation, maximum temperature, and minimum temperature. The Intergovernmental Panel on Climate Change 5 (IPCC5) report stated that greenhouse gas emissions are the main source of uncertainty for future temperature and precipitation. Representative concentration pathway 4.5 (RCP4.5) and RCP8.5 are moderate and higher mitigation scenarios, respectively, which are appropriate for research on the influence of climate [53]. 

 https://www.cesm.ucar.edu/experiments/cmip5.html.

《Table 3》

Table 3 Basic information of 31 GCMs in CMIP5.

Climate projections generated by GCMs are subject to significant uncertainties at large spatial scales and deviations in simulation. Multi-ensemble GCMs are generally considered superior to single models [54]. LSTM and MMEM methods were adopted for climate change projections to improve the prediction accuracy of the simulated outputs from multiple GCMs. MMEM is a traditional method that is evaluated using mean values from the multi-model ensemble. LSTM is a special type of recurrent neural network (RNN), capable of learning long-term dependencies and remembering information for long periods and has been proven to capture the temporal relationship among time series data better than traditional RNNs [26]. The introduction of LSTM was used to detect temporal signals and build the nonlinearity correlation between downscaled data from the 31 GCMs and the observed data of the reference period, and then propagate the long-term information to a future period. The primary purpose of the LSTM was to integrate the GCM ensembles into high-quality and highly accurate precipitation and temperature.

The special structure of LSTM is that it has a memory cell, which is used to store information and propagate outputs for different time steps. A structure known as gates, which includes a forget gate, input gate, and output gate, manages the cell states and outputs. The gate consists of a sigmoid function with a multiplication operation. The sigmoid function varies from 0 to 1, indicating how much information can be passed through or discarded. Specifically, first, a sigmoid layer called the ‘‘forget gate layer” determines how much information to be discarded from the previous cell state, for example, 0 represents all of the information that would be abandoned; second, a sigmoid layer called the ‘‘input gate” combined with a hyperbolic tangent (i.e., an activation function in neural network to push the values to be between –1 and 1) layer created a new cell state; finally, the ‘‘output gate” will output determined parts based on the cell state. 

In this research, LSTM models were built for temperature and precipitation projections with 31 GCMs. First, the inputs of the LSTM model for temperature contain 31 mean temperatures calculated by the maximum and minimum daily temperatures of the 31 GCMs. The paired dataset between the calculated mean daily temperature of 31 GCMs and the observed daily temperature was obtained from 1961 to 2010. The paired dataset for precipitation prediction was formulated based on the monthly precipitation of 31 GCMs and the monthly observed precipitation from 1961 to 2010. Then, temporal signals were detected, and nonlinearity correlations were developed between the downscaled historical and observation data. The period for training the model was determined to be 40 years from 1961 to 2000, and for testing, it was ten years from 2001 to 2010. The evaluation criteria NSE and R were used to represent the prediction accuracy in the reference period. Finally, the temperature (or precipitation) of the 31 GCMs from 2021 to 2050 was fed into the aforementioned LSTM models during 1961–2010 to predict the corresponding future temperature and precipitation, respectively.

Better climatic projections from 2021 to 2050 would be achieved if the temporal signals between the downscaled data from 31 GCMs and the observed data were detected from 1961 to 2010. Deep learning and remembering through LSTM were used to improve the prediction accuracy using observed and downscaled climate data from 1961 to 2010. Using multi-ensemble GCMs and taking 1961 to 2010 as the reference period, the projection of precipitation and temperature from 2021 to 2050 would propagate outputs based on the memory cell of the LSTM. The precipitation and the temperature from 2021 to 2050 were directly averaged from the 31 GCMs using MMEM. 

《3.4. CA–Markov model》

3.4. CA–Markov model

The CA–Markov model, which is formed by integrating the CA model with the Markov model, was adopted to reflect the dynamics of land use patterns. The CA–Markov model synthesizes the advantages of applying multi-standard evaluation and multiobjective decision support systems to define the transfer rules between land use types.

The Markov model is a stochastic process and optimal control theory method for predicting the types and transfer rates of land use [35]. The land use projections based on Bayes are as follows: 

where C(t) and C(t + 1) are the states at time t and t + 1, respectively, and Pij is the shift probability matrix of a status, which can be calculated as follows:

The CA model is a grid dynamics model with discrete-time, space, and state, and can comprehensively consider historical trends, including natural, social, and economic conditions for land use projections. The CA model can be calculated as follows [54]:

where t and t + 1 represent time, C is the set of cellular status, D represents all the cells in a certain neighborhood, and is the transformation rule of the cellular status. For this research, D contains the influence factors (i.e., elevation, slope, precipitation, temperature, river, road, and population density), represents the combined rule of each cell in the neighborhood, and C(t + 1) represent the land use status at the latter moment, which is affected by the status of the previous moment C(t).

Specifically, the CA–Markov model in the IDRISI software was used to project the long-term land use variation at ten-year intervals of 2025, 2035, and 2045 based on land use maps of 1995, 2005, and 2015. The detailed processes are as follows: ① calculation of the transfer area matrix and transfer probability matrix of land use types in ARB from 1995 to 2005 and 2005 to 2015; ② selection of natural factors such as elevation, slope, precipitation, temperature, river, and human factors of road and population density as driving factors for land use projection. The MCE and LogisticReg in the module were used to create transition suitability images for each land use type. ③ Based on the spatial distribution and driving factors in 1995 and 2005, the spatial variation of land use for 2015 was simulated, and the actual interpretation data for 2015 were used to verify the simulation results to ensure the reliability of the model. ④ Based on the links between the change and various driving factors, three sets of land use projections for 2025, 2035, and 2045 were projected using three methods.

《4. Results and discussion》

4. Results and discussion

《4.1. Performance of MIKE SHE》

4.1. Performance of MIKE SHE

The DEM, land use map, soil data, rainfall, wind speed, sunshine duration, and daily maximum and minimum temperatures were used to calibrate the MIKE SHE. The daily streamflow at the watershed outlet was compared with the measured data to assess the model performance under climatic changes. First, MIKE SHE was calibrated during the calibration period (1977–1983) and was validated during the validation period (1984–1986) with data from the daily streamflow at the Wantan station and the 1980 land use map. The model was further validated from 2006 to 2007 using the 2005 land use map (Fig. 3). For the daily simulation, the R and the NSE values were 0.85 and 0.72 in the calibration period, 0.87 and 0.73 in the first validation period, and 0.84 and 0.70 in the second validation period, respectively. There was no obvious systematic shift of the relevant points in each period and there were consistent trends, indicating that the Wantan hydrologic station has good consistency for the watershed streamflow. The R and NSE values in the calibration period were 0.88 and 0.79, respectively, compared to the monthly streamflow, which was 0.86 and 0.74. The results suggest that MIKE SHE can effectively simulate daily streamflow of ARB. The simulation effect of the monthly scale is better than that of the daily scale, which shows good applicability in mountainous areas as a whole.

We can see from Fig. 3 that the streamflow during spring and winter was relatively small, while it was relatively large during the summer and autumn. From the peak simulation, the simulated value was smaller than the observed value at a certain time. The hydrologic model did not perform well during the second validation period as there were changes in the input vegetation types, illustrating that land use change also affects the hydrological process. Additionally, precipitation and other meteorological data cannot fully represent each calculation unit as the meteorological stations in this study area are sparse, which will have a certain impact on the simulation accuracy of the model.

《Fig. 3》

Fig. 3. MIKE SHE simulated and observed daily streamflow at Wantan station during (a) the calibration period (1977–1983) and (b) the validation period (1984–1986 and 2006–2007).

To validate the applicability of the MIKE SHE model to other hydrologic elements, the actual ET from 1977 to 1986 was simulated and translated into the average value of the study area. The annual mean actual ET was approximately 580–720 mm·a–1 in the study area, which was consistent with the range of the latest research from the National Tibetan Plateau Data Center, indicating that the actual ET in the southwest river basin (including the Yangtze River) should be significantly higher than 400 mm·a–1 [55]. The actual ET from 2021 to 2050 will be further analyzed and discussed in the following sections. 

《4.2. Performance of multi-ensemble GCMs for climate projection》

4.2. Performance of multi-ensemble GCMs for climate projection

The future climatic conditions were obtained from 31 GCMs using the LSTM and MMEM methods under scenarios of RCP4.5 and RCP8.5, from 2021 to 2050. The downscaled historical data and measured climate data for the period from 1961 to 2000 were used to train the model, whereas data from the period of 2001 to 2010 were used to test the model performance. Table 4 shows the daily temperature and monthly precipitation of meteorological stations from 1961 to 2010 using the two methods. The coefficients of R and NSE suggest that the LSTM method has a strong ability for an ensemble of GCMs and can better reflect the daily temperature in the ARB over the reference period. However, certain extreme values still existed that could not be captured perfectly. As 90% of the daily precipitation was zero or extremely low, the LSTM could not predict the extreme values well for capturing most of the data. Certain studies have also indicated that it is usually difficult for RNNs to predict high floating and low regular data impeccably, and daily precipitation has irregular changes and large fluctuations [56]. Thus, daily precipitation with large fluctuations was transformed into monthly precipitation to eliminate zero values, and the results suggested that the LSTM had a strong ability to project monthly precipitation. Furthermore, the coefficients of R and NSE indicate that the projected monthly precipitation has a better performance than the daily temperature. 

《Table 4》

Table 4 Accuracy test of daily temperature and monthly precipitation projection at the three meteorological stations from 1961 to 2010.

The performance of the LSTM method was evaluated by comparing the projections to the observed precipitation, and it was found that the dataset projected from 31 GCMs of the LSTM method can fit better than the MMEM method (Fig. 4). The monthly average precipitation and daily temperature from the LSTM projection had a stronger correlation with the observed data as a whole and could better project the temperature and precipitation over the reference period. For the multi-year monthly precipitation and temperature of Xichang, the precipitation and temperature curves estimated by MMEM were closer to the observations, especially from May to September. However, for monthly precipitation and temperature from 1961 to 2010, the coefficients of R and NSE indicate that the climate data estimated by LSTM have a better performance than MMEM at the three stations. The advanced deep learning method of LSTM proved to be more efficient at dealing with temporal signals than the conventional average method of MMEM. 

《Fig. 4》

Fig. 4. Multi-year monthly average precipitation and temperature obtained from MMEM and LSTM method (1961–2010). Pre_O: observed precipitation; Pre_M: precipitation projected by MMEM; Pre_L: precipitation projected by LSTM; Tem_O: observed temperature; Tem_M: temperature projected by MMEM; Tem_L: temperature projected by LSTM.

《4.3. Future climate projections》

4.3. Future climate projections

Considering the historical period (1961–2010) as the baseline period, the trend of temperature variation in the time period 2021 to 2050 was analyzed using the LSTM and MMEM methods (Fig. 5). Therefore, four sets of climate change scenarios were generated for the influence study from 2021 to 2050 (i.e., RCP4.5_L, RCP8.5_L, RCP4.5_M, and RCP8.5_M). Different variation trends were observed under different ensemble methods from the perspective of the monthly average temperature. The projected values from the LSTM in different scenarios were generally higher than the historical values from October to April, but lower from May to September. The temperature from the MMEM value was higher in the meteorological stations of Mianning and Xichang, but lower in the Miyi station compared to the historical temperature. It also showed different levels of the increasing trend under the two greenhouse emission scenarios of RCP4.5 and RCP8.5, of which the latter had a slightly higher temperature. For the Mianning and Xichang stations, the average annual temperature under RCP4.5 and RCP8.5 projected from LSTM both increased compared with the historical period (1961–2010), and the latter increased approximately 1.4  °C than the former, which is consistent with the results predicted in the IPCC5 report. The increase in watershed temperature will inevitably lead to the rapid melting of glaciers and snow in the ARB headwaters, and a change in the runoff generation mechanism, which will bring new opportunities and challenges to local water resource management and use. 

《Fig. 5》

Fig. 5. Monthly average temperature for projected period (2021–2050). Hist: historical period from 1961 to 2010; RCP4.5_L: RCP4.5 projected from LSTM; RCP8.5_L: RCP8.5 projected from LSTM; RCP4.5_M: RCP4.5 projected from MMEM; RCP8.5_M: RCP8.5 projected from MMEM.

Fig. 6 illustrates the changing trend of projected monthly average precipitation in the future period (2021–2050) under two scenarios (RCP4.5 and RCP8.5) from the MMEM and LSTM values. The precipitation projected from the LSTM was higher than the MMEM at the Mianning and Xichang stations. For the future period (2021– 2050), the monthly precipitation under multiple scenarios fluctuated significantly compared to the historical precipitation from May to September, while changing minutely in the other months. The results also indicate that the average annual precipitation under RCP4.5 and RCP8.5 increased relative to the historical period, and the former had an increase of 1.9 mm, especially from June to October.

《Fig. 6》

Fig. 6. Monthly average precipitation for the projected period (2021–2050).

《4.4. Land use projections》

4.4. Land use projections

The land use during the years 1980, 1995, 2005, and 2015 was classified into five major types, including cropland, forest, grassland, water, and built-up land (Table 5). Specifically, according to the available remote sensing image data, the 1980 land use map in the study area was acquired, and was adopted for calibration during the calibration period (1977–1983) and for validation during the validation period (1984–1986). We mainly obtained the future land use projections based on the land use data from 1995, 2005, and 2015 with transition matrix per decade; thus, the changes from 1980 to 1985 had little influence on land use projections from 2021 to 2050. As presented in Table 5, the dominant land use was forest, followed by cropland and grassland in ARB. This demonstrated that the land use in different types of cropland have sustained reductions over the past decades, which may be attributed to national conservation programs such as the ‘‘Grain for Green Project” since the 1990s. The change in the watershed may have been caused by the construction and operation of the Daqiao reservoir in the upper basin since the 1990s and may have been affected by water conservancy projects.

《Table 5》

Table 5 Various land use types and changing rates over 1980–2015 in the watershed.

With a comprehensive consideration of the driving factors of land use change, the elevation, slope, precipitation, temperature, river, road, and population density were selected to generate transition suitability images. The representative modules of MCE and LogisticReg combined with the CA–Markov model were adopted to evaluate the relationship between the land use and the driving factors for three sets of land use projections. To test the accuracy of the CA–Markov model’s prediction of the quantity and spatial distribution for each land type, the land use of 2015 was projected based on the land use data of 1995 and 2005. The Crosstab module in IDRISI software was used to compare three sets of projections and actual land use maps, the kappa coefficients obtained were 0.96, 0.89, and 0.91, respectively. The results indicate that the CA–Markov model is a dependable estimator and can be used to conduct future land use prediction simulations.

Because of the satisfaction of the calibrated module for 2015, three Markov-based modules (i.e., scenario 1 from MCE; scenario 2 from LogisticReg; scenario 3 from Markov; Fig. 7) were executed to project the land use change of 2025, 2035, and 2045; each land use projection represents the land use map of ten years (i.e., 2025 represents the period of 2021–2030; 2035 represents the period of 2031–2040; and 2045 represents the period of 2041–2050). As can be seen in Fig. 7, the forest area is the largest, followed by cropland and grassland. Built-up land has generally expanded in three scenarios, mainly concentrated in the residential and industrial areas along with the transportation land, especially for the midstream and downstream. The primary trend was the conversion from natural land type to artificial land use type, which seems reasonably feasible to happen because of the regional economic development, increase in population, and other factors. There is no specific methodology to assess the results, and all the future scenarios can occur as no evidence of future conditions exists. As land use varies and spatial variation becomes more complex, it is necessary to strengthen the protection of cropland, forest, and water bodies for the rational use of land.

《Fig. 7》

Fig. 7. Three scenarios of land use map projections in 2025, 2035, and 2045. (a) MCE method; (b) LogisticReg method; (c) Markov method.

《4.5. Evaluation of hydrologic responses》

4.5. Evaluation of hydrologic responses

4.5.1. Streamflow variation to climate and land use changes

Subsequently, based on the above four sets of climate projections and three sets of land use projections, the MIKE SHE model was used to assess the hydrological process in the corresponding period. Cross combination of 12 scenarios were used as inputs to project the process and the trend of streamflow as well as the temporal and spatial variability of the actual ET under climate and land use variations, respectively. The simulated average annual stream flow under different scenarios is illustrated in Fig. 8. The streamflow with climate projections from LSTM was higher than that from MMEM. We can see that there was no obvious change in the streamflow under the three land use projections, but it was more sensitive to changes in climate variables. The climatic change is the principal cause of streamflow variations during the historical period, which is consistent with previous research. From the perspective of the multi-year average discharge, the simulated value under the RCP4.5 scenario was higher than that under the RCP8.5, particularly for the period from 2031 to 2040. Even though there was no significant difference in temperature under RCP4.5 and RCP8.5, the future streamflow of the former was higher because of precipitation variations. As precipitation had a great influence on the hydrologic change, the streamflow prediction of low precipitation projections from MMEM was much lower than that from LSTM. We can conclude that the annual runoff increased significantly with an increase in precipitation and decreased with an increase in temperature.

《Fig. 8》

Fig. 8. Streamflow under different scenarios in the future (2021–2050). (a) Climate change projected from LSTM; (b) climate change projected from MMEM. LU1, LU2, and LU3: three land use scenarios.

To further analyze the multi-year monthly average distribution of future streamflow variations, box–whisker plots of the change rate under RCP4.5 and RCP8.5 in the future period (2021–2050) were compared with the baseline period (1977–1986) (Fig. 9). The monthly streamflow had an obvious increase from January to April, while it had a dramatic decline from August to November. The reason for the significant increase may be because of the slight increase in temperature, which will inevitably lead to the melting of glaciers and snow in the headwater of ARB. As illustrated in Fig. 9, there was a higher variation in the monthly flows from May to September than in the other months, which indicates a higher uncertainty in the streamflow predictions. Two reasons can illustrate the slight decreasing trend in streamflow under different emission scenarios for the flood period. First, the calibrated hydrologic model did not perform well for the peak value in the flood period because of the deep slopes, valleys, and geographical conditions. Second, although the precipitation had a slight increase during this period, the hydrologic process may be affected by other factors, such as the increased temperature, surface type, cloud cover, sunshine hours, and wind speed. Runoff is likely to increase in the rainy season, especially from May to August; therefore, attention should be paid to prevent the occurrence of flood disasters. The runoff is likely to increase in the dry season compared with the reference period, especially from February to April, while a decrease in runoff during November indicates an increased risk of drought.

《Fig. 9》

Fig. 9. Projected change of streamflow at the monthly scale in the future period (2021–2050) compared with the baseline period (1977–1986). (a) RCP4.5 scenario; (b) RCP8.5 scenario.

4.5.2. Temporal and spatial variability of actual ET

For a better assessment of the hydrologic process under changing conditions, an analysis of actual ET was conducted. For this study, the temporal and spatial actual ET was acquired from the output result of MIKE SHE using the two-layer unsaturated zone/ evapotranspiration (UZ/ET) model based on a formulation presented by Yan and Smith [57]. Fig. 10 illustrates the spatial variability of the actual annual ET for different climate and land use scenarios. For the simulation results of MIKE SHE, dynamic changes in the actual ET under two climate scenarios and three land use scenarios from 2021 to 2050 were presented compared to the reference period (1977–1986). The actual ET increased with an increase in temperature, and the increase under RCP8.5 (Figs. 10(c) and (d)) was higher than that under RCP4.5 (Figs. 10(a) and (b)), owing to the higher temperature. The actual ET was more sensitive to climate variation than to land use variation according to the influence study.

《Fig. 10》

Fig. 10. Spatial variability of annual average ET in 2021–2030, 2031–2040, and 2041–2050 compared with the baseline period of 1977–1986 under RCP4.5 and RCP8.5. (a) climate data projected from LSTM under RCP4.5; (b) climate data projected from MMEM under RCP4.5; (c) climate data projected from LSTM under RCP8.5; (d) climate data projected from MMEM under RCP8.5.

Fig. 11 illustrates the comparison of streamflow and actual mean ET per decade for each of these scenarios. The annual average streamflow under RCP4.5 was higher than that under RCP 8.5, while the actual ET exhibited a consistent variation trend under different scenarios except for the first land use scenarios under RCP8.5. The components for water balance include precipitation, canopy interception, surface runoff, unsaturated and saturated flow, and so forth. Clearly, precipitation is closely related to the changes in streamflow and the actual ET. Moreover, the increase in temperature can influence ET and reduce canopy resistance in a certain temperature range, of which soil moisture in arid and semi-arid areas positively impacts the actual ET [11]. There was no consistent or opposite correlation between the actual ET and the streamflow, which may be attributed to the uncertainties of future climate and land use projections, as well as the complex topographic conditions.

《Fig. 11》

Fig. 11. Comparison of streamflow and actual mean ET per decade. (a) RCP4.5; (b) RCP8.5. C1 and C2: two climate scenarios; L1, L2, and L3: three land use scenarios; P1, P2, and P3: 2021–2030, 2031–2040, and 2041–2050, respectively.

《5. Conclusions》

5. Conclusions

This research revealed the influences of climatic and land use changes on the hydrological process in the dry–warm valley of the Anning River watershed. To address the uncertainty issue that arises from a single GCM, multi-ensemble GCMs using an advanced method of LSTM and a traditional method of MMEM were used in this study to obtain four climatic projections. Three sets of land use projections were generated using a comprehensive consideration of driving factors such as elevation, slope, precipitation, temperature, river, road, and population density. With the constraints of data availability and the large-scale heterogeneity of mountainous areas, this study evaluates the dynamic variation condition and range of hydrologic processes in ARB under complex changing conditions.

The results showed that the hydrologic model could be effectively applied to streamflow simulation in the dry–warm valley of the ARB. The LSTM method had a better performance for daily temperature and monthly precipitation projections than the traditional MMEM method. However, it was difficult for the LSTM to predict daily precipitation with zero values. The future temperature projection under RCP8.5 was slightly higher than that under RCP4.5, while the precipitation projection was slightly lower from June to October. The annual streamflow increased significantly with an increase in precipitation and decreased with an increase in temperature. Runoff and actual ET are more sensitive to climate change than land use changes. Runoff is likely to increase in the rainy season from May to August and decrease during the dry season of November. The actual ET increased with an increase in temperature and had a higher increase under RCP8.5. In future changing environment, effective measures should be taken to deal with flood disasters from May to August, and drought disasters in the dry season, especially in November. As an essential branch of the upper Yangtze River with a fragile ecological environment, it is essential to investigate complicated spatial and temporal variability and stochastic hydrologic processes for complex terrain regions. Consequently, this work could provide an effective guidance for dealing with the risk of flooding and drought in a changing environment, contributing to the sustainable development of the upper Yangtze River’s local and downstream areas.

The changing environment presents new challenges to water resources and environmentally sustainable development for the eco-fragile and downstream regions. A dynamic hydrologic model with more updated parameters is expected to explore the complex and uncertain future changing environment.

《Acknowledgments》

Acknowledgments

This study was supported by the National Key Research Program of China (2016YFC0502209), Beijing Municipal Natural Science Foundation (JQ18028), and the National Natural Science Foundation of China (51879007 and U20A20117). We acknowledge the modeling groups in making the available CMIP5 multi-model dataset and AGRIVY for providing the data for this study. The authors would also like to express their gratitude to the editor and reviewers for their suggestions to improve the manuscript and references.

《Compliance with ethics guidelines》

Compliance with ethics guidelines

Congcong Li, Yanpeng Cai, Zhong Li, Qianqian Zhang, Lian Sun, Xinyi Li, and Pengxiao Zhou declare that they have no conflict of interest or financial conflicts to disclose.