Evaluation of river ecological flow based on baseflow separation in Xiangjiang River, China

Water is the building block of ecosystems and is crucial for controlling climate and preserving the equilibrium of ecosystems, particularly for material cycling, energy transfer, nutrient delivery, information transfer, and connectivity between bio-habitats (Westall & Brack 2018). However, since society and the economy have grown quickly, people have used river water resources more intensively, and thus, the issue of river water resources has progressively taken center stage in ecology and the environment (Piralizefrehei et al. 2022). To safeguard the ecological environment and species variety of rivers, numerous experts and academics have also conducted related research to preserve the equilibrium of river ecosystems through ecological flow regulation (Huang et al. 2018; Shan et al. 2022).

Ecological flow is the flow necessary to preserve the species diversity and stability of river ecosystems (Robinson et al. 2023). The establishment of reasonable and scientific ecological flow thresholds is not only beneficial to the efficient use of local water resources but also contributes significantly to keeping river ecosystems stable (Zhang et al. 2024). Ecological flow was first proposed in the 1940s, and since then, a lot of research has been done on the topic (Abebe et al. 2022). River ecological flow can be measured using a variety of techniques, such as hydraulic and hydrological methods. Due to its applicability, hydrological methods are the most commonly employed (Lei et al. 2023; Yang et al. 2023; Yu et al. 2023). However, the traditional studies on ecological flow thresholds still have some deficiencies and need to be further strengthened to determine ecological flow thresholds (Guo et al. 2024; Zhang et al. 2024). In addition, there are not many studies on ecological flow computation using river baseflow data, and conventional hydrological methods frequently rely on long series of runoff data. As the primary source of river flow and a major factor in preserving the stability of river ecosystems and the health of water ecosystems, baseflow is a reasonably steady component of river runoff (Cheng et al. 2021; Sun et al. 2021). Determining the ecological flow threshold from the river base flow data is a critical research subject because the river ecology will be threatened if the river flow falls below the base flow limit.

Numerous experts and researchers have developed various hydrological models, including semi-distributed models and data-driven models, to predict river flow (Alemu et al. 2023; Vidyarthi & Jain 2023; Yao et al. 2023). By segmenting the watershed into sub-watersheds and taking into account the hydrological processes and characteristics within the sub-watersheds, runoff is simulated for semi-distributed hydrological models (El-Nasr et al. 2010; Ding et al. 2023). The Soil and Water Assessment Tool (SWAT) is one of the most popular semi-distributed models due to its capacity to run continuous simulations for extended periods and evaluate the effects of different land management techniques on the hydrological cycle (Arnold et al. 1998; Yuan & Forshay 2021; Rahman et al. 2022; Mahdian et al. 2024). In contrast to conventional hydrological models that are based on physical processes, data-driven models can estimate and predict hydrological processes directly from observed data, negating the need for a thorough knowledge of the physical mechanisms underlying hydrological processes (Wu et al. 2023). Typically, this kind of model processes data and creates predictive models using methods like artificial intelligence, machine learning, and statistics. The potential of machine learning models to optimize water resources management, improve the accuracy of flood forecasts, and aid in the design of water projects remains enormous (Qian & Du 2023; Ahmed et al. 2024; Akinsoji et al. 2024). A popular data-driven hydrological model for runoff prediction is the long short-term memory (LSTM) network model, which tackles the problem of recurrent neural networks' long-term dependance and reduces the ‘gradient vanishing’ that results from these networks' backpropagation during training (Kratzert et al. 2018; Wang et al. 2022; Saravani et al. 2025). Based on earlier research, a coupled SWAT–LSTM model was built in this work to simulate runoff flow using both the SWAT and the LSTM models, which can significantly increase simulation accuracy.

Ecological flow early warning is an important element in water resources management, which aims to ensure the sustainable development of river ecosystems while satisfying human use of water resources (Lu et al. 2021). The principle of ecological flow early warning is to monitor, assess, and manage the future safety classification of river ecological flow through scientific and reasonable prediction models, to establish an early warning mechanism for river and lake ecological flow forecasting, to effectively protect and maintain the ecological environments of rivers and water bodies, and to make the river ecosystems meet the requirements of ecological sustainable development. In order to support the ecological flow scheduling of river terrace reservoirs and the best scheduling of regional water resources, five Early Warning Levels (EWLs) were categorized in this study according to the annual base flow process, and ecological flow thresholds were established. Furthermore, the river ecological flow warning system was also developed using the SWAT–LSTM model, and the effects of climate change and human activity on the changes of Ecological Flow Assurance Degree (EFAD) were measured using the separative framework (Guo et al. 2022). The results of this study fill a gap in this field.

In conclusion, this study proposed a river ecological flow early warning system, which was specifically divided into the following four steps: (1) the minimum ecological flow threshold of the river was determined according to the river's intra-annual base flow process, and the EWLs were determined according to the ecological flow threshold; (2) the EFAD of the river in different periods was assessed; (3) the ecological flow early warning system was established through the coupled SWAT–LSTM model; (4) the role of climate change and human activities in the changes of EFAD was quantified through a detached framework. Compared with previous studies, the ecological flow thresholds determined by the base flow of the river combined with the Flow Duration Curve (FDC) curve and Tennant's method are more in line with the actual water demand of the river and take into account the needs of the aquatic ecosystem at different times. A new runoff prediction system is established based on the SWAT–LSTM coupled model, which significantly improves the simulation accuracy. Still, the coupled model requires sufficient training data and is sensitive to the training parameters. In addition, we introduced the Range of Variability Approach (RVA) and a separate framework to quantitatively assess the changes in ecological flow security, and the study fills the gap in the related field. Xiangjiang River is the most important river in Hunan Province, and its watershed area accounts for 40% of Hunan; its distinct aquatic environment serves as a suitable home for a variety of creatures (Lu et al. 2022). However, owing to the quickening pace of population expansion, economic growth, and climate change, human society's demand for water resources has increased dramatically, and the imbalance between ecological water use and economic water use has led to the deterioration of the basic ecological environment of the Xiangjiang River (Guo et al. 2022). Therefore, it is necessary to establish a river and lake ecological flow forecasting and early warning system to ensure that the river flow meets the requirements of ecologically sustainable development.

The Xiangtan Hydrological Station's daily flow data from 1961 to 2019 were utilized in this study; these data were taken from the Yangtze River Basin Hydrological Yearbook. As illustrated in Figure 1(b), the temporal distribution of the hydrological processes throughout the year is not consistent. The China Meteorological Data Service Centre (http://data.cma.cn/) provided the meteorological data used in this study, including maximum temperature and precipitation at the Yongzhou and Hengyang meteorological stations from 1961 to 2019, and the Geospatial Data Cloud (https://gscloud.cn/) provided the elevation data for the Xiangjiang River Basin. The National Tibetan Plateau Science Data Center's Harmonized World Soil Database (http://data.tpdc.ac.cn) provided the soil data. Information about land use was gathered from the Resources and Environment Data Center (https://www.resdc.cn/).

The Mann–Kendall (M–K) test technique is a hydroclimatic diagnostic and forecasting technique that can be used to monitor long-term hydrometeorological series for abrupt changes and identify abrupt events (Gülü 2020; Wang 2020; Noori et al. 2022). However, it requires that the sample data are continuous and independent, which will challenge the ability of the M–K test when there is autocorrelation in the sample data, so the hydrological series were whitened and detrended (Hamed & Rao 1998; Bürger 2017). However, the trend-free pre-whitening M–K test (TFPW-MK) still has a shortcoming: it will have more than one mutation point when performing the mutation test, so it needs to be tested for more than one mutation point. In this study, the cumulative distance level and Pettitt's test were used to further test the detrended M–K results. The cumulative distance level method calculates the accumulation of distance level values and further determines the mutation year by the trend of the cumulative curve (Yi et al. 2021; Du et al. 2023). The Pettitt test is a method based on the extreme value statistic, which is based on the change of the location of the extreme value appearing in the time series to detect the mutation point (Hajani & Rahman 2018; Kannan et al. 2022). Related techniques have been applied in many studies and will not be repeated in this study.

The specific steps of baseflow separation technology are as follows:

• (1) Collect runoff data: First of all, it is necessary to collect the runoff data of the river and preprocess the collected runoff data to ensure the completeness and accuracy of the data.

• (2) Determine the filtering parameters: The filtering parameter is a key parameter in the CM method, which determines the response speed of the filter and the accuracy of baseflow segmentation. Usually, this parameter needs to be debugged and optimized according to the hydrogeological conditions of the specific basin or refer to the references of related watershed applications.

• (3) Baseflow extraction: The preprocessed runoff data will be input into the CM filtering algorithm, which will process the runoff data according to the set filtering parameters, separating the base flow and the fast response runoff part. The filtering algorithm will output two sequences: one representing the base flow portion and the other representing the fast response runoff portion.

This study used forward and reverse double filtering for the river flow series to increase the accuracy of base flow separation. When using single-parameter filtering, it is required that the separated daily base flow is not less than 0 m³/s. Therefore, half of the first day's flow was used as the first day's base flow in conjunction with previous studies.

Ecological flow in rivers determines to ensure that river ecosystems are not subjected to the stress caused by reduced flow, preserving river ecosystems' balance and vitality in the process (Chen et al. 2025). However, traditional hydrological methods often use natural flow sequences as the basis for determining eco-flow, without considering the eco-flow thresholds under extremely unfavorable conditions (Qu et al. 2022a). Moreover, traditional hydrological methods are generally used to assess the reasonable eco-flow thresholds of rivers by only one method, and few studies have investigated the eco-flow thresholds of rivers by combining multiple hydrological methods. Therefore, in this study, the Tennant method was combined with the Base Flow Relative Duration Curve (BFDC) based on the base flow of the river to derive the ecological flow threshold (Yu et al. 2023). The following are the precise steps:

• (1) Plot the BFDC of the monthly average base flow derived from the daily base flow separated by the base flow separation method and the monthly average base flow.

• (2) The Tennant technique takes into account that 10% of the average multi-year flow will satisfy the minimal flow requirements of river ecosystems throughout both times, with October–March being the General Water Use Period (GWUP) and April–September being the Fish Spawning and Fattening Period (FSFP). Furthermore, according to the Tennant technique, the majority of aquatic creatures' flow requirements would be satisfied by 30% of the average multi-year flow during the GWUP and 40% during the FSFP.

• (3) The river's ecological flow thresholds were classified as EWLs based on the upper and lower bounds of the ecological flow thresholds drawn by the Tennant method in the second step, along with the BFDC of the monthly average base flow during the GWUP and the FSFP (Zhang et al. 2012; Guo et al. 2021).

There were several sub-basins created in the basin for this study to better capture variations in hydrological conditions. The daily flow data that was collected at the Xiangtan Station was then used to calibrate the model. In this research, the construction of the ecological flow early warning system was carried out with 1991–1993 as the warm-up period, 1994–1998 as the model correction period, and 1998–2005 as the validation period, and then the daily flow of 2006–2019 was simulated based on the relevant data. In this study, model calibration was done using the SUFI-2 algorithm in SWAT-CUP (Ozdemir & Leloglu 2019; Sao et al. 2020).

Because of its superior performance during training, LSTM models are a popular kind of recurrent neural network in deep learning (Fang & Shao 2022; Wang et al. 2023). It has produced positive results in earlier studies and can be applied to build large-scale recurrent neural networks to solve complex sequential machine-learning challenges (Kim et al. 2024). The LSTM model can overcome the gradient vanishing problem by ‘backpropagation in time’ during training. In this work, the Xiangtan Station's multi-year daily flow data are used as a driver to predict future runoff using the LSTM model. Additionally, the original multidimensional vectors were normalized to hasten the model's convergence. The timeframe chosen for the model correction and validation in the construction of the ecological flow early warning system was 1991–1998 and 1998–2005, respectively. The daily flow following these selections was then simulated using pertinent data.

The formula, indicates the EFAD of EWL m in the year i (m is Ⅰ, Ⅱ, Ⅲ, Ⅳ, and Ⅴ represents five EWLs respectively), %; indicates the EFAD of EWL m in period j of the year i (j is 1, 2 represents GWUP and FSFP of the year i, respectively), %; denotes the number of days to satisfy the EWL m in the year i; denotes the entire count of days within the year i; denotes the number of days to satisfy the EWL m in period j of the year i; indicates the entire count of days within the period j of the year i; indicates the daily flow of the Xiangjiang River on day l of the k months in period j of the year i; denotes the EWL m ecological flow threshold at period j, the flow units are all m³/s; i denotes the i year, k denotes the k month, and l denotes the l day of the k month.

It was determined that 0–33% is a low level of change, 33–67% is a medium level of change, and 67–100% is a high level of change when analyzing the RVA results.

In Equation (12). and are the contributions caused by both climate change and human activities to changes in EFAD, %, respectively.

Combined with the measured runoff data from 1961 to 2019 at Xiangtan Hydrological Station, the test results of the mutation year obtained from the TFPW-MK, cumulative distance level, and Pettitt test are shown in Table 1. After careful consideration, 1991 was ultimately chosen as the abrupt change year of the hydrological situation of the Xiangjiang River. The base period, which was defined as 1961–1990, and the human activity period, which was defined as 1991–2019, were determined by the mutation year.

In this study, the flow was simulated from monthly and daily scales using the SWAT model, LSTM model, and coupled model of both. To improve the computational speed of SWAT and reduce the computing time of the model, the test statistic t and the significance index p were used to screen out the 10 indexes that have a high impact on the sensitivity of the model, and the larger the absolute value of t, the closer p is to 0, the more sensitive the parameter is (Bhattacharya et al. 2020). For LSTM, the model effect is more stable when the number of iterations is 300, the time step increases with the number of iterations, and other relevant parameters are set. The parameters required in the two models are shown in Tables 3 and 4. Table 5 shows the fitting effects for each period from 1991 to 2019. The findings indicate that the fitting coefficients (R²) of both the SWAT model and LSTM model are above 0.89 on the monthly scale, but the fitting coefficients of the coupled model are above 0.91. The outcomes of the NSE coefficients show that the NSE coefficients of the SWAT model and LSTM model are above 0.79 for all three periods, and the NSE coefficients of the coupled model are above 0.85. The simulation effect is better. On the daily scale, the fitting coefficients (R²) of the SWAT model and the LSTM model are above 0.79, but the fitting coefficients of the coupled model are above 0.84. The results of the NSE coefficients show that the NSE coefficients of the SWAT model and the LSTM model for all three periods are above 0.78, and the NSE coefficients of the coupled model are above 0.81.

In the GWUP, the lowest EFAD of EWL I was 63.19% (1963) before the mutation and 62.09% (1998) after the mutation, which is not a big change, but after the mutation, it reached 100.00% in several years, accounting for 65.52% of all years. In addition, the EFAD of other EWLs decreased more significantly, with the EFAD of the EWL V reaching a maximum of 12.64% (1965) before the mutation and less than 2.00% in all years after the mutation. For the FSFP, the above phenomenon is more significant, but the EFAD of the EWL IV before the mutation is up to 40.44% (1963), and after the mutation is 20.22% (2018), which still exerts a coercive pressure on the river ecosystem. The change of river EFAD from EWL V to EWL I is the result of reservoir scheduling, and the gradient operation of reservoirs on rivers can effectively alleviate the problem of uneven intra-annual distribution of river flow, which leads to the change of EWL to ensure the stability of river ecosystems (Tu et al. 2015).

The attribution analysis's findings revealed that the EFAD of EWL IV for GWUP was most affected by climate change (165.20%). The EFAD of EWL V for FSFP was most impacted by climate change (100.00%) and the EFAD of EWL II was second most affected by climate change (91.02%), but the impact of climate change on the EFAD of EWL II for the GWUP was opposite to that of FSFP (−62.19%). In addition, human activities had the greatest influence on the EFAD at EWL II for the GWUP (162.19%), and at EWL III for the FSFP (76.95%), and the impact of human activities on the EFAD at EWL V for both the GWUP and all year was the same. The results of the study also demonstrated how human activity and climate change affect EFAD alterations. They were the same for all periods except for EWLs II, IV, and V during the GWUP and EWLs II and V during the whole year.

Ecological flow can maintain suitable flow conditions in a water body to meet the minimum flow requirements for ecosystem needs, so determining the ecological flow of a river is of great significance for the preservation and protection of the natural environment. First, the ecological flow can help to maintain the stability of environmental conditions such as humidity, water temperature, and water quality in aquatic habitats, as well as provide the oxygen, food, and habitat needed by organisms, and help to maintain the balance of various ecosystems including wetlands, rivers, lakes, and coastal zones (Liu et al. 2023). Furthermore, appropriate ecological flows can create diverse habitats, protect and promote species diversity in various ecosystems, and thus facilitate the reproduction and migration of a wide range of aquatic plants and animals (Keller et al. 2019). Qian et al. (2012) found that, among the components of river flow, baseflow remains relatively stable over a long period and is a key factor in keeping river ecosystems stable, which is one of the substrates to maintain the river ecosystem. However, few previous studies have investigated the ecological flow that maintains the river from the perspective of baseflow. In this study, the baseflow process during the year was separated by the baseflow separation method, and the ecological flow thresholds for different periods of the year were determined by the Tennant method. For river management practice, the ecological flow thresholds during the FSFP and GWUP periods can be used as one of the guidelines for the joint scheduling of the gradient reservoirs. At the same time, these thresholds can be used as an important reference basis for river management and protection, which can help the managers develop a more scientific and reasonable river management strategy: by ensuring that the ecological flow in the river is not lower than these thresholds, the river aquatic ecosystems can be effectively protected, and the ecological function and biodiversity of the river can be maintained. This is of great significance for ensuring the survival of aquatic organisms, realizing the sustainable use and management of rivers, and guaranteeing the long-term healthy development of river ecosystems. However, due to the complexity of the formation and evolution mechanism of river baseflow, the results of baseflow segmentation by C–M digital filtering in this study can only partially represent baseflow. Furthermore, the existence of multiple baseflow separation methods, each with its specific assumptions and limitations that lead to inconsistent results, adds to the uncertainty of applying baseflow separation techniques. Gan et al. (2022a) analyzed the effectiveness of seven basic flow separation methods, including digital filtering, and found that the simulated basic flow process line by digital filtering can reflect the hysteresis effect of runoff, which is consistent with the relevant results of basic flow separation reflected in this study. In addition, baseflow separation techniques typically rely on long-term, continuous runoff data. In areas where data are scarce or unavailable, the difficulty and accuracy of baseflow separation can be compromised. An in-depth study of the formation mechanism of baseflow, the development of new and more accurate baseflow separation methods, and a comparison of the performance of different methods are also an important direction for future research. The ecological flow threshold of the river can be delineated from the point of view of a more thorough analysis of the formation mechanism of the river's baseflow in the future. Moreover, this study only investigated the ecological flow of the river from the hydrological process and did not consider the ecological water demand of the river ecosystem. Subsequent studies can further incorporate the water requirements of fish, water quality, and other ecological conditions in the watershed to improve the science and accuracy of baseflow-based ecological flow threshold calculation (Fornaroli et al. 2016).

Under the ecological flow thresholds, this study classified the ecological flow of the river into five EWLs, and an ecological flow early warning system was constructed to discriminate the EWL and to support the most efficient distribution of local water resources. The SWAT–LSTM coupled model is used to predict future flow and create an ecological flow warning system to direct cascade reservoir operations. The ecological flow warning system has profound potential and practical value in protecting the living environment of aquatic organisms, maintaining the health of ecosystems, optimizing water resource allocation, improving water resource utilization efficiency, flood control, and disaster reduction, promoting sustainable economic and social development, and advancing ecological civilization construction by monitoring and warning the flow changes of rivers, lakes, and other water bodies in real-time. It is an important tool for achieving the dual protection of rational utilization of water resources and the ecological environment. The findings indicate that the accuracy of the coupled simulation has been enhanced both on the daily and monthly scales, and good results have been achieved when relying on the flow predicted by the coupled model to discriminate the EWL of the river. However, the SWAT model has a complex structure and requires a relatively large database, which limits the application of the model when dealing with large-scale datasets or making real-time predictions. Moreover, LSTM, as a black-box model, has prediction results that are difficult to interpret intuitively, which may not satisfy the demand for model transparency and interpretability in areas such as water resource management, and these factors limit the development of SWAT–LSTM models. Follow-up research could develop the application of machine learning interpretable models in hydrology to serve directions such as regional water resource management. Research on how to optimize the structure and algorithms of SWAT–LSTM models to improve their performance and computational efficiency under different data conditions is also an important future research direction. In addition, while this study examined the evolutionary patterns and driving causes of river ecological flows, it did not examine the impact of numerous influences, such as plant cover and land use, on the accuracy of flow safety level predictions (Hammond & Fleming 2021). In the course of subsequent research, the model can be further optimized to create a model with better predictive performance, which can provide early warning of the ecological flow safety classification of rivers and serve the ecological scheduling of reservoirs.

In order to explore the impact of changes in the hydrological regime on whether the actual flow of rivers can meet the needs of river ecosystems, this study analyzed the EFAD of the river in different periods before and after the abrupt change in the hydrological regime. The outcomes demonstrated that, as a result of the river terrace reservoir building, the safety level of river flow was transformed from other levels to EWL I after the sudden hydrological change, which provided a suitable flow for river organisms, guaranteed a suitable habitat environment for river organisms, and improved the safety level of river ecosystems. According to related research, drought events with more significant impacts occurred in 1963, 2003, 2007, 2011, and 2018 in the Xiangjiang River Basin, and the drought events would affect the river flow, which had a direct impact on the river EFAD (Wang et al. 2020). However, after the hydrological mutation, this effect was mitigated better during the GWUP and weaker during the FSFP. Therefore, the regulation of river ecological flow during the FSFP can be strengthened in the gradient scheduling of reservoirs to cope with the risk of drought. Ma et al. (2019) found that the guaranteed ecological flow rate of the Yangtze River Basin in China was decreasing year by year by assessing the Yangtze River Basin's many transects' ecological flow conditions, and the results of the study were different from the present one, which was because the early warning of the ecological flow was delineated by the present study range, which is the minimum flow to avoid serious damage to the river ecosystem, so the EFAD will be increased. In addition, the transition of the safety level of river flow from other levels to EWL I indicates a trend toward homogenized river flow evolution, but activation of fish reproductive behavior and triggering of key life cycle events such as incubation, as well as migration and habitat connectivity, need to be supported by flow pulses, and therefore, there is a need to explore ecological flow thresholds for maintaining normal pulsation state rivers (Espínola et al. 2017; van der Sleen & Rams 2023).

In addition, the results of the RVA show that the EFAD before and after the mutation has changed significantly, and when analyzing the contribution rate of climate change and human activities to the change of the EFAD, it may be observed that there is a clear difference in the roles that human activities and climate change play at various periods. From the EWLs shift, it can be seen that the increase in the EFAD of EWL Ⅰ is mainly transformed by the EFAD of EWLs Ⅱ, Ⅲ, and Ⅳ. 2163. In the GWUP, the transformation of EWL Ⅱ is mainly impacted by human activities, and the transformation of EWLs Ⅲ, and Ⅳ is mainly impacted by climate change; in the FSFP, the transformation of EWLs Ⅱ and Ⅳ is mainly impacted by climate change, and the transformation of EWL Ⅲ is mainly affected by human activities; in the whole year's point of view, the transformation of EWL Ⅱ is mainly affected by human activities, and the transformation of EWLs Ⅲ and Ⅳ is mainly impacted by climate change. This suggests that the FSFP after the mutation year can play a dominant role in flow regulation due to sufficient precipitation, but during the GWUP due to the decrease in precipitation, the combined operation of the terrace reservoirs is needed to compensate for the decrease in river flow (Cheng et al. 2020; Huang et al. 2020; Xie et al. 2020). This is consistent with the research results of Guo et al. (2022) in the Xiangjiang River, which showed that human activities were the dominant factor in runoff changes during the dry season, while climate change was the dominant factor during the wet season. From the perspective of the whole year, human activities play a greater role in EWL II, but a more feeble role in EWLs III and IV, which also indicates that the current operation mechanism of the Xiangjiang River terrace reservoirs has some deficiencies, and the reservoir operation managers and decision-makers should strengthen the capture and flow supplementation of low-flow events to fulfill the flow requirements for the stability of the river ecosystem (Olabiwonnu et al. 2022).

River basin water resource management and the health of river ecosystems depend heavily on the reasonable assessment of ecological flow thresholds and the forecast of safe levels of ecological flow in rivers. In this research, the ecological flow threshold of Xiangjiang River was determined according to the base flow separation method combined with the Tennant method, and the EWLs of the river were determined based on the BFDC. The early warning system of river ecological flow was also established through the SWAT–LSTM coupled model, and the effects of climate change and human activities on the change of EFAD were quantitatively separated by combining the separated framework. The proposed methods for establishing an ecological flow warning system have filled the gap in the relevant research field. The results of the study are as follows: (1) The base flow separation method in conjunction with the Tennant method was used to estimate the river's ecological flow thresholds during the GWUP and the FSFP, and the base flow relative time curves were used to classify the EWLs; the results of the study can be used to serve the joint scheduling strategy of terrace reservoirs. (2) The SWAT–LSTM coupled model was used to build the early warning system for river ecological flow, and the early warning accuracy rate exceeded 89.58% on the monthly scale and 84.07% on the daily scale. The research results can be used to predict the safety level of river ecosystems under future changing environments. (3) The impact of human activities and climate change on the changes in EFAD varies significantly in different periods. However, on a year-round scale, the role of human activities is greater for EWL II but weaker for EWLs III and IV. This indicates that the current operation of the terrace reservoirs in Xiangjiang River is inadequate in the capture of low-flow events and flow replenishment; decision-makers need to further adjust the operational strategies of the tiered reservoirs.

It was found that the ecological flow early warning system we constructed can be good for ecological flow early warning. Still, this study only explored the ecological flow of the river from the hydrological process and did not consider the ecological water demand of the river ecosystem. In fact, the ecological flow of a river needs to meet the water demand of aquatic organisms and the water environment so the water demand of fish, water quality, and other ecological conditions in the watershed can be further combined in the subsequent study so as to improve the science and accuracy of the calculation of the ecological flow threshold based on base flow. In addition, the activation of key life cycle events such as fish reproductive behavior and triggering of hatcheries, as well as migration and habitat connectivity, requires the support of flow pulses, so it is necessary to explore the ecological flow thresholds of the river under the conditions of maintaining the normal pulses of the river, which points to the direction of the subsequent research. The results of this study can also provide a paradigm for ecological flow warning in other watersheds and provide a basis for reservoir ecological flow scheduling, regional water resource allocation, and aquatic habitat health protection.

This study was funded by the Basic Research Project of Key Scientific Research Projects of Colleges and Universities of Henan Province (23ZX012), the National Natural Science Foundation of China (Grant no. 51779094), 2025 Henan Provincial Natural Science Foundation (252300421334)，2025 Henan Provincial Outstanding Youth Science Foundation (252300421049) and the Henan Provincial Science and Technology Research Project (232102320033).

W. G. rendered support in acquisition of funds, management of projects, resources, research, and supervised the work. G. W. the stages involved in this process include conceptualization, data curation, formal analysis, investigation, methodology, original draft writing, and editing. F. H. and L. H. validated the work, developed the methodology, rendered support in formal analysis, and investigated the project. X. B., B. W., Y. L., C. S., and Z. Y. visualized the study, support in research, and rendered support ion formal evaluation. H. W. rendered support in acquisition of funds and project management.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.