Dry-Season Water Level Shift Induced by Channel Change of the River–Lake System in China’s Largest Freshwater Lake, Poyang Lake

As a large open lake that connects to the middle reaches of Yangtze River, Poyang Lake, China, has experienced dramatic water level changes, especially during the dry periods (November–March) in recent years. For better understanding the shift of dry-season lake level, we quantified the variations induced by channel changes of the river–lake system (CCORLS). Attempts were achieved through establishing two Long Short-Term Memory (LSTM) based water level prediction models, combined hydrological data in the periods of 1956–2000 and 2006–2016. The results indicated that: (1) With reference to earlier (1956–2000) dry periods, the recent (2006–2016) mean water levels had an overall decreasing trend, with values of 0.04–1.26 m across the lake. (2) The response of water level to CCORLS was spatially heterogeneous, triggering a sharp reduction in the central to northern lake, an obvious reduction in the lake outlet, and a relatively smaller reduction in the southern lake. (3) The bottom topography alteration of Poyang Lake made a dominating contribution to decline the lake water levels, while the Yangtze River’s mainstream channel near the lake played a minor role in the decrease of water level. Results shed light on the impact of systematic watercourse changes on the water level of a lake connected a river and enhanced understanding of the characteristics and mechanisms behind the hydrological changes of large freshwater lakes.


Introduction
Freshwater resources, which are critical to human wellbeing, are highly unevenly distributed on the Earth, with nearly 90% of the planet's liquid, surficial freshwater stored in lakes (Sterner et al. 2020;Zhang et al. 2022). A lake is both a crucial carrier of physicochemical processes in the Earth's water cycle and an essential component of the global aquatic ecosystem (Huang et al. 2022;Messager et al. 2016). Large lakes provide a wealth of ecosystem services globally (Li et al. 2022), spanning biotic or abiotic cultural services (e.g. habitat conservation, recreational tourism), and provisioning services (e.g. fish stocks and nutrient transport) (Sterner et al. 2020). As a critical indicator representing its hydrological regime, the water level of the lake is of great importance for lake management. Its fluctuation is closely related to the structural stability and functional integrity of lake ecosystems (Gownaris et al. 2018;Coops et al. 2003). The timely fluctuation of water levels is necessary to maintain lake biodiversity. But, significant changes in lake level regime may break the original ecosystem balance, cause undesirable ecological consequences (Wantzen et al. 2008), and even trigger the security issues in drinking water and food supply (Wheater and Gober 2015;Ebi and Bowen 2016).
Influenced by climate change and intensive human activities, many large lakes have undergone dramatic changes. Poyang Lake, the largest freshwater lake in China, which is connected downstream with the Yangtze River, has formed a unique regime in water level fluctuation due to the strong river-lake interaction (Dai et al. 2015;Lai et al. 2014a). The regime drives the inundation and exposure process of a large area of shoals in Poyang Lake. It fosters a wetland of the global importance that provides the critical habitats for a large number of rare and endangered species (Tang et al. 2016). Unfortunately, significant changes in hydrological regimes have been observed in recent decades, which may break the longformed ecological equilibrium of river-lake ecosystem .
Numerous studies have been conducted to understand hydrological changes, e.g. when the water level variation occurred, how the water level fluctuated, what the reason and potential impact of variation were, and how to cope with it (Dai et al. 2015;Feng et al. 2012;Guo et al. 2018;Yang et al. 2016). These studies agree that Poyang Lake has experienced extreme low water levels, with earlier time and longer duration of dry water level, recently, due to climate change and human activities (Lai et al. 2014b;Min and Zhan 2012). But due to the complex hydrological regime in Poyang Lake, there are different views and research conclusions on the causes of the lake level changes, especially its changes in the dry season. Feng et al. (2012) suggested that precipitation dominated the recent drought. Contradicting Feng's argument, Zhang et al. (2011) found that the winter rainfall increased. Guo et al. (2012) stated that the river-lake interaction had been altering along with the river discharge variation because of the Three Gorges Dams (TGD). But the numerical investigations showed that the TGD can compensate for the amount of water in dry season and raises the water level in the middle Yangtze River (Lai et al. 2014c(Lai et al. , 2016. Liu et al. (2013) explored the water budget and size of the lake and ascribed the lake decline to weakened blocking effect of the Yangtze River. These studies have explained the river-lake interaction and its possible impact on the lake water level change from a water budget perspective. However, water budget analysis cannot fully reveal the major cause that alters the lake hydrological regime. The Poyang river-lake system has complex topographies, upon which the hydrological regime is strongly dependent. Lai et al. (2014d) first explored the effect of channel change on the lake discharge capacity and found that the extremely low lake level can be attributed to the outflow channel change caused by sand mining. Yao et al. (2018Yao et al. ( , 2019 further investigated the spatiotemporal changes of hydrodynamic processes and hydrological regime caused by bathymetric changes of Poyang Lake using a hydrodynamic model. These studies confirmed that sand dredging boosted around the northern part of Poyang Lake, leading to a wider and deeper outflow channel and a stronger discharge ability into the Yangtze River. But, the Yangtze River channel, on which the river-lake system depends, was not considered. To gain a better understanding of the hydrological regime change in this dynamic river-lake system, we quantified the contributions of channel changes including the Yangtze and the lake outflow channel to the dry-season lake water level variation. First, two Long Short-Term Memory (LSTM) based water level prediction models were developed for two main stages with relatively stable hydrological regimes. Then, characteristics of dry-season water level changes of Poyang Lake and spatiotemporal response of low lake levels to CCORLS were derived and discussed.

Study Area
Located on the south bank of the middle reaches of the Yangtze River (Fig. 1b), Poyang Lake (28°22′-29°45′N, 115°47′-116°45′E) maintains a natural connection with the river and is the largest freshwater lake in China. The lake has a water volume of 2.76 × 10 9 m 3 and its catchment area is 1.62 × 10 5 km 2 , which nourishes about 10 million residents and of 10 thousand km 2 of farmland in the basin (Shankman et al. 2006;Shankman and Liang 2003). Poyang Lake has a complex flow regime featured by strong river-lake interaction. The lake mainly receives water from the southern five major tributaries including the Ganjiang River, the Fuhe River, the Xinjiang River, the Raohe River and the Xiushui River (Fig. 1c). Then the lake discharges into the Yangtze River through a narrow and deep outflow channel in the north. It yields backflow from the river into the lake while the river water level is higher. Dominated by the subtropical monsoon climate, the region is characterized by a wet season between April and September while a dry season between November and March of next year, resulting in unevenly inter-distributed precipitation (Zhang et al. 2011). As a typical floodplain lake, Poyang Lake accordingly exhibits drastic seasonal and inter-annual water level fluctuations and inundation coverage changes. The largest water surface area exceeds 3000 km 2 in the wet season while it covers less than 1000 km 2 , effectively dwindling to a river channel, during the dry season.

Data Collection
Observations of daily water level at four water level stations and daily river discharge at eight streamflow stations are available for the period of 1956-2016 (Fig. 1). The stations of Kangshan, Duchang, Xingzi and Hukou, located from the (c) southern upstream to the northern downstream outlet sequentially, were selected to explore spatial and temporal water level alterations with the river-lake interactions. The sum of the discharges at six gauging stations (Waizhou, Lijiadu, Meigang, Dufengkeng, Hushan, Wanjiabu) of the southern five major tributaries were adopted to represent catchment inflow. For the sake of later description, the catchment inflow was marked as Q W . The discharges at Hankou were used to indicate the flow components of the Yangtze River, denoted as Q Y . And the discharges at Datong were applied to detect the river channel changes in the middle reaches of the Yangtze River. All daily hydrological data were obtained from the Hydrological Bureau of Jiangxi Province and the Changjiang Water Resources Commission of the Ministry of Water Resources, in China. It is widely believed that the river-lake interaction has changed before and after the TGD operation in 2003 (Guo et al. 2012;Lai et al. 2014b, c).

A LSTM Based Prediction Model for Poyang Lake Water Level
LSTM has been successfully applied to rainfall-runoff and runoff-water level simulations with high performance for both large and small watersheds (Kratzert et al. 2018;Lv et al. 2020;Xiang et al. 2020). Kratzert et al. (2019) pointed out the similarities and differences between LSTM and traditional hydrological models and given a possible hydrological interpretation of LSTM applied the rainfall-runoff problems. LSTM is a specifically designed temporal recurrent neural network to solve the long-term dependency problem (Hochreiter and Schmidhuber 1997) that existing in a general recurrent neural network (RNN). In contrast to RNNs, a LSTM possesses a more complicated chain-like repeating structural module (Gers et al. 2000;Zhang et al. 2018a). The memory block is comprised of a forget gate ( f ), an input gate ( i ), an output gate ( o ) and a memory cell ( C ), as shown in Fig. 2. Firstly, the forget gate controls how much information of cell state will be transferred from the previous moment to the current moment ( f t ), with values in the range (0,1). Secondly, the input gate and memory cell decide to what extent new information will be added in the cell state ( C t ). The sum of the two steps represents the cell memory at the current moment ( C t ). Finally, the hidden output ( h t ) is determined by resizing the C t value to -1 to 1 through tanh and multiplying it by the output gate. Given an input time series X = x 1 , x 2 , ⋯ , x t , ⋯ , x n , the memory block activation values are calculated sequentially at each time t ( 1 ≤ t ≤ n ), according to the following equations: (1) where W i , W f and W o denote the matrix of weights from the input, forget, and output gates to the input, respectively. Similarly, U i , U f and U o denote the matrix of weights from the input, forget, and output gates to the hidden, respectively. b i , b f and b o denote the input, forget, and output gate bias vectors, respectively. σ is the logistic sigmoid function with values in the range of 0-1. tanh is the hyperbolic tangent function with values in the range of -1-1. x t is the input vector. ⨂ represents the scalar product of two vectors.
To explore the lake water level changes caused by CCORLS, we built two models (Model1 and Model2), respectively, representing flow regimes in the past river-lake system and the recent river-lake system (Fig. 2). According to the channel modifications in this river-lake system (Lai et al. 2014d), two relatively stationary periods, i.e. 1956-2000 and 2006-2016, were adopted for the training of both models. With respect to the LSTM, several important hyperparameters such as neuron numbers in hidden layer and the initial learning rate can directly affect the prediction effect of models (Abbasimehr et al. 2020;Smith 2017). The optimal water level prediction model contains a single hidden layer with 50 neuronal nodes and an initial learning rate of 0.001, based on the trial-and-error method. The ADAM optimization algorithm is selected to update the gradient of the memory block (Kingma and Ba 2014). And the Dropout regularization mechanism is used to prevent overfitting (Srivastava et al. 2014). Both models have the same structure as described above. In each model, the seven flow variables (six streamflow of Poyang Lake and Hankou's streamflow) from the previous 6 days (t-6) to the current moment (t) were cast as the inputs, and the four water levels at time t + 1 were treated as the outputs. Additionally, the partitioning of each model training and testing sets followed the 80/20 rule (Zhang et al. 2018a). All observations were normalized in the range of 0-1 to eliminate the foreseeable numerical problems by utilizing the maximum and minimum scale transform principle. The root mean square error (RMSE), the Nash-Sutcliffe Efficiency (NSE) and the coefficient of determination (R 2 ) were selected as evaluation indicators to quantify the effectiveness of the established models. Interpretively, the model was more favourite when RMSE approached 0 and the other two indexes closed to 1 (Kisi et al. 2012;Nash and Sutcliffe 1970;Zhang et al. 2018b).

Model Calibration
Both models perform well by using discharges of river-lake water system to predict water levels at Hukou, Xingzi, Duchang and Kangshan stations (Table 1) (Fig. 3). That both the simulated water levels fit well with the corresponding observations indicates the prediction models yielding high accuracy. The relative errors (RE) between the observed and the simulated water levels vary from 0.01% to 0.62% in Model1 and from 0.66% to 1.69% in Model2. These results mean that both models can fairly capture the stage-discharge relationship in the connected river-lake interaction and can be used for further scenario analysis.

Scenario Analysis
Model1 constructed based on data from 1956-2000 and Model2 built based on data from 2006-2016 has been used to delineate the old and new CCORLS, respectively. To determine the impact of CCORLS on the dry-season lake water level, four computational scenarios were designed: (1) Lake water levels with the past flow regime in the past river-lake system (Eq. 7), (2) Lake water levels with the past flow regime in the recent river-lake system (Eq. 8), (3) Lake water levels with the recent flow regime in the past river-lake system (Eq. 9), (4) using Lake water levels with the recent flow regime in the recent river-lake system (Eq. 10). The difference between the two simulated water levels under the same flow regimes was ascribed to the impact of CCORLS on the water level changes (Eq. 11 and Eq. 12).  where f 1 and f 2 represent the stage-discharge relationship simulated by Model1 and Model2, respectively. ΔZ CCORLS1 is the water level variation induced by CCORLS under the past river-lake flow condition . ΔZ CCORLS2 is the water level variation induced by CCORLS under the recent river-lake flow condition (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016).

Water Level and Flow Changes
The mean lake water levels during dry season have altered significantly in recent years (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016) at four stations compared with the earlier years , while (12) ΔZ CCORLS2 = Z sim 22 − Z sim 12 both the lake inflow and Yangtze River discharge show an obvious increasing trend (Fig. 4). In the dry season, the mean water levels at Xingzi and Duchang, the central and northern lake areas, decreased by 0.63 m and 1.26 m (Fig. 4a) in recent years, respectively. But, the lake levels at Hukou, the lake outlet, and Duchang, the southern lake area, do not show significant declines. The recent mean water level at the river-lake connection varies insignificantly, declining by only 0.04 m at Hukou, the lake outlet. Further, the mean water level at Kangshan, which is in the upstream of Poyang Lake, reduces slightly by 0.24 m. Monthly water level changes during the dry period present various spatiotemporal characteristics, from 1956-2000 to 2006-2016 (Fig. 4b). Compared to the earlier monthly water levels, the recent water levels at Xingzi, Duchang, and Kangshan generally decline in the periods, except for Kangshan where the water level slightly increases in December. In particular, the magnitude of water level reduction at Duchang is the biggest. From the radar chart we can see clearly that the water  (Fig. 4c). And water discharges of the Yangtze River increase from 11,001 m 3 /s to 12,514 m 3 /s (Fig. 4d).

Relationship among Water Level, Lake Inflow and the Yangtze River Flow
Compared to the earlier decades , the relationships among the lake water level, lake inflow and Yangtze River discharge in recent dozen years (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016) changed in various degrees across the lake (Fig. 5). Similar to the earlier, the recent water levels are all mainly controlled by the river discharges, when the Yangtze River discharges are greater than 30,000 m 3 /s. This is because the excessive river water level enhances the blocking effect, limits the lake's discharging capacity, and even forces the river water backflow into the lake. With reductions of flow regime of the river-lake system, the water level-Q Y -Q W relationship at Hukou remains nearly unchanged, which means that the water level at Hukou is enormously controlled by the Yangtze River discharges. Contrary to Hukou, the water level at Kangshan is dominated by the catchment streamflow of Poyang Lake. Under the knowledge that the lake inflow impact on the water level shows an increasing spatial distribution pattern from the north to the south, while the influence of river discharges reveals an opposite spatial distribution characteristic, the water levels at Xingzi and Duchang are controlled by the flow regime of the river-lake system. Unlike the earlier smooth relationship curves, the current relationship curves are complex, especially during the dry period. Given the same flow regime of the river-lake system, the water level changed insignificantly as there was a large river-lake discharge, while the recent (

Impact of CCORLS on Water Levels
Regardless of the earlier  or recent (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016) river-lake flow condition, CCORLS significantly brought about a decrease in the water level of Poyang Lake. Furthermore, the spatial distribution of the effect of CCORLS on the water level of Poyang Lake was heterogeneous, triggering a sharp reduction in the central lake, an obvious reduction in the lake outlet, and a relatively smaller reduction in the southern lake (Fig. 7). Under the past flow regime , the mean water levels most significantly decreased by 1.62 m at Duchang, followed by 1.24 m at Xingzi, then by 0.93 m at Hukou, and by 0.38 m at Kangshan, during the dry season. In the light of monthly time scale, the water level changes exhibited a prominent decreasing trend during the dry period. The CCORLS yielded lower monthly water levels

Discussion
Statistical results show that the mean lake inflow increases by 21%, during the dry period, and the mean Yangtze River discharge rises by 14% ( Fig. 4c and d). It is puzzling why the water level of Poyang Lake has decreased ( Fig. 4a and b) even though the river-lake discharge has increased. The seemingly contradictory phenomena virtually implied that the dry-season water level reduction of Poyang Lake was considered to be caused by the river-lake system's channel change. Yao et al. (2018) applied a physically based hydrodynamic model to calculate the response of water level decline to the bathymetric changes of Poyang Lake during the dry season. This study further develops previous work by involving the effect of the Yangtze River channel change on the dry-season water level reduction and simplifies the analysis tools by using a data-driven LSTM model. The unevenly spatiotemporal distribution of water level changes during the dry periods are the results of CCORLS involving lake bottom topography and river channel changes owing to natural evolution and anthropogenic activities. It is known from previous research that Poyang Lake is being in a relatively stable tectonic development stage (Wu et al. 2015). Therefore, the main factors affecting bottom topography of the area must be attributed to scouring and siltation alterations of the river-lake waterway system under the complex interaction among hydrodynamic, sediment transport and human activities (Gao et al. 2014;Wu et al. 2015).
The bottom topography change of Poyang Lake plays a leading role in declining water level, especially in the northern outlet channel. On the one hand, the decreasing sediment load into Poyang Lake slowed down deposition of five rivers' alluvial cones in recent years. The implementation of the policy of converting cultivated farmland to lake has weakened sediment concentration, controlled soil erosion and intercepted sediment load (Feng et al. 2011;Min et al. 2011). Gao et al. (2014) stated that the average annual sediment load of five rivers decreased by 64% during 2003-2010, compared to the reference during 1956-2002. On the other hand, severe scouring from central lake to northern outlet increases the sediment load entering into Yangtze River and decreases the central-to-northern elevation of lake bottom topography. The lake scouring can mainly be attributed to human activities, but not natural dynamics . Sand mining is thought as main contribution to this scouring (Wu et al. 2007;Lai et al. 2014d). Since the government banned sand mining in the Fig. 7 Comparison of the water level changes due to CCORLS at four stations during the dry season in 1956-2000 (a1-d1) and in 2006-2016 (a2-d2). Z sim 11 is the water level simulated by Model1 during 1956-2000 andZ sim 22 is the water level simulated by Model2 during [2006][2007][2008][2009][2010][2011][2012][2013][2014][2015][2016] Yangtze River, sand mining was mainly concentrated in the northern river-lake connected channel during 2001-2007, then expanded to the lake's central part recently (Wu et al. 2007). According to statistics, the average dredging depth was 4.95 m and the accumulated sand mining volume was up to 1.29 × 10 9 m 3 by 2010 (Jiang et al. 2015). The intensification of sand mining strengthened the hydrodynamic conditions in the central lake, deepen and widened the lake outflow channel, and increased suspended sediment content at Hukou (Lai et al. 2014d;Yao et al. 2019). Yao et al. (2018) resulted that the bed erosion of the northern outlet channel averaged 3 m, decreasing the water level by 1.2-2 m in the northern channels and increasing the average annual outflow by 6.8%.
Meanwhile, the channel morphology of Yangtze River has been subject to considerable down cutting since the TGD operation, further accelerating flow and sediment discharge capacities of the lake into the river. As the results of sediment retention and clear water releasing by the TGD, the riverbed turned from depositional before the dam construction to erosional afterwards, leading to varying degrees of channel sinking along the thalweg throughout the main river course (Dai and Lu 2014;Lai et al. 2014b;Yang et al. 2014). It is worth mentioning that the degree of erosion was related to the distance to the TGD. The sediment carried by the upstream streamflow has been heavily trapped as it passed through the Three Gorge Dams, only transporting 21%, 47% and 56% of the original sediment load at Yichang, Hankou and Datong, respectively (Zhao et al. 2017). Dai and Liu (2013) defined the rate of erosion as the amount of erosion per unit length of the river, quantified the average erosion rate from Hankou to Hukou (0.1-0.2 × 10 6 t/km) and from Hukou to Nanjing (smaller than 0.1 × 10 6 t/km), and classified the former as intermediately erosional and the latter as weakly erosional. This was supported by our findings in this study (Fig. 6).
Up to now, we have known that the dry-season water level variation, caused by CCORLS, showed an unevenly decreasing trend across the Poyang Lake, no matter under which water condition. In 1956-2000, the mean dry-season waler levels at Duchang, Xingzi and Hukou decreased by 1.70 m, 1.12 m and 0.36 m, respectively. In 2006-2016, the mean dry-season waler levels at Duchang, Xingzi and Hukou decreased by 1.62 m, 1.24 m and 0.38 m, respectively. The channel morphology variation of Yangtze River led to a water level reduction of about 0.6 m at the lake outlet. The hydraulic connection between the Yangtze River and the lake decreased with the increasing distance from the lake outlet, thus the influence of the Yangtze River channel morphology variation on the water level of Poyang Lake weakened correspondingly . Yao et al. (2018) deemed that the bathymetric changes of the lake caused mean water level to decrease by 1.2-2.0 m in the northern lake, during the low water level period. In a word, the bottom topography alteration of Poyang Lake makes a dominating contribution to decline the central and northern lake water levels, while the impact of the Yangtze River's channel change on the water level reduction cannot be ignored.

Conclusions
This study integrates an LSTM-based water level prediction model and hydrological data to quantify the impact of the channel change of the river-lake system on the water level variation of Poyang Lake, during the dry periods (November-March). An uneven lake water level reduction with values of 0.04-1.26 m is recognized across Poyang Lake. The significant decline of water level in Poyang Lake is ascribed to the CCORLS, which contains the bottom topography of Poyang Lake and the Yangtze River's mainstream channel near the lake. The response of water level to the CCORLS is spatially heterogeneous, triggering a sharp reduction in the central to northern lake, an obvious reduction in the lake outlet, and a relatively smaller reduction in the southern lake. Reduction of lake water levels are mainly controlled by the bottom topography change of Poyang Lake, while are further enhanced by the river channel change of the Yangtze River. However, it is important to note that the impact of the TGD on downstream rivers and lakes may be exacerbated because of potential riverbed scouring of the Yangtze River mainstream in the future. The water level of the main stream could further drop for the same flow rate and then drive lake level down.