The Xinanjiang (XAJ) model is widely used and a de facto standard hydrological model in China. However, its implementation remains ambiguous due to the 5 mm inflow-based subdivision, originally introduced to control numerical errors in time discretization. Ambiguity persists regarding its necessity, scope, and threshold, leading to inconsistent implementations and hindering model reproducibility. To address this issue, we introduce a differential-form XAJ model and employ high-order numerical solvers to obtain a high-accuracy benchmark. Idealized experiment and large-sample applications are conducted to evaluate the subdivision method. Results show that subdivision is necessary for error control and should be applied across all modules. The 5 mm threshold is generally adequate, but smaller thresholds are required in a subset of watersheds. This study provides the first quantitative assessment of the subdivision method, resolves a long-standing ambiguity in the XAJ model, and highlights the importance of explicitly accounting for numerical implementation in hydrological modeling.

River network routing is a key component linking spatially distributed hillslope runoff to streamflow response at any point in the river network. While recent advances have reformulated conceptual hydrological models into differential form to reduce numerical errors, routing modules have largely remained in difference form, resulting in numerical inconsistency between modules. This study proposes a matrix-vector differential-form Muskingum (MVDM) method for river network routing within an ordinary differential equation framework. Starting from the classical difference-form Muskingum equations, storage and flux variables are identified to derive a differential formulation, which is then reorganized into a compact matrix-vector system for segmented reaches and generalized to river networks using connectivity matrix and block-matrix structures. Two numerical experiments demonstrate that the difference-form Muskingum method introduces a noticeable numerical discrepancy relative to the analytical solution, and further generates a coupling-induced error that accounts for about 8.7% of the peak flow when integrated with a hydrological model. The propsed MVDM effectively controls these numerical errors. Application to the Lianjiang River watershed further confirms the effectiveness and applicability of the MVDM method, achieving improved model skill performance and reduced low-flow bias relative to the classical Muskingum method. This study revisits the classical river-network routing method from the perspective of differential equations, and enables fully differential coupling with conceptual hydrological models, allowing seamless integration with modern numerical solvers for improved model performance.

Scientific flood season segmentation serves as the foundation for determining the flood-limited operating water levels across different periods, providing crucial support for reservoir flood control safety operations and optimal water resource utilization. Under the background of climate change, the traditional static flood-limited water level management model based on fixed dates struggles to adapt to variations in flood season patterns. This study aims to establish a scientifically sound flood season segmentation scheme, providing a basis for dynamic control of flood-limited water levels across different periods, thereby improving water resource utilization efficiency while ensuring flood control safety. This study focuses on the Wuluwati Reservoir and employs the circular distribution method and the Fisher optimal partition method to conduct its flood season segmentation calculations. First, the circular distribution method is used to analyse the concentration and periodic characteristics of flood occurrences in the basin. Subsequently, the Fisher optimal partition method is applied to perform statistical segmentation of the historical hydrological series. Based on this analysis, the flood season of the Wuluwati Reservoir is comprehensively determined as: the pre-flood season from 1 June to 2 July, the main flood season from 3 July to 27 August, and the post-flood season from 28 August to 30 September. To objectively evaluate the rationality of the segmentation results, the improved Cunderlik method was employed to examine the rationality of 15 segmentation schemes based on relative superiority degree. The results show that the scheme with the main flood season from 3 July to 23 August achieves the highest relative superiority degree (0.930). The comprehensively determined segmentation of this study (3 July-27 August) encompasses this optimal interval, demonstrating that the flood season segmentation for the Wuluwati Reservoir is reasonable and effective.

Accurate soil moisture (SM) estimation is essential for applications such as flood forecasting, irrigation management, and drought monitoring. Data assimilation (DA) provides a powerful means to integrate multi-source SM observations, yet conventional approaches-such as the ensemble Kalman filter (EnKF) and its variants-are constrained by linear update formulations and Gaussian assumptions. To overcome these limitations, we introduce the ensemble deep learning filter (EnDLF), a novel DA framework that replaces the Kalman-based formula with a deep learning (DL)-based nonlinear mapping capable of capturing complex relationships and non-Gaussian patterns. To reduce the computational burden of retraining the DL model at each assimilation step, EnDLF employs a transfer learning strategy that fine-tunes model parameters across assimilation cycles, substantially shortening training time. Applied to an SM assimilation case with pronounced non-Gaussianity, EnDLF outperforms EnKF, reducing root-mean-square error (RMSE) by 10.53 % and 9.09 % and improving Kling-Gupta efficiency (KGE) by 5.13 % and 3.90 % for surface (0-10 cm) and subsurface (10-50 cm) SM, respectively. Enhanced SM estimates further translate into improved streamflow predictions, increasing Nash-Sutcliffe efficiency (NSE) from 0.848 with EnKF to 0.864 with EnDLF. These results demonstrate the effectiveness of integrating DL into DA to capture nonlinear dependencies and highlight EnDLF's potential for assimilating other hydrological variables with non-Gaussian characteristics, such as precipitation and streamflow.

Based on three generalized beach profiles, this study utilizes a numerical modeling approach to investigate the combined effect of tides and irregular waves on groundwater dynamics and marine-sourced salt transport during beach recovery. Compared to tidal-only conditions, the inclusion of non-overtopping waves amplifies differences between berm profile and storm profile in upper saline plume (USP) features (area and salt content), salt-freshwater mixing zones. With overtopping waves, these differences are further magnified due to high-salinity zones formed by infiltration of ponding water in backshore depressions behind berms. Particle tracking shows that wave action introduces additional deceleration phases in particle movement, leading to more complex travel speed distributions. Furthermore, overtopping waves intensify water and salt exchange across the aquifer, elevating intertidal saline infiltration (ISI) and submarine groundwater discharge (SGD) by up to 1,248% and 182%, respectively. This research underscores the critical role of wave overtopping in enhancing subsurface mixing and solute transport during beach recovery, with important implications for managing seawater intrusion in vulnerable shorelines.

Climate change is redefining meteorological triggers of agricultural drought in the upper-middle Hanjiang River Basin (UMHRB), threatening water and food security. We coupled a calibrated VIC model with bias-corrected NEX-GDDP-CMIP6 forcings to reconstruct (1980-2020) and project (2021-2100) root zone soil moisture. Copula-Bayesian conditional probability yields agricultural drought trigger thresholds (ADTTs) from mild to extreme events, while the Lindeman-Merenda-Gold (LMG) metric, proportional marginal variance decomposition (PMVD), and relative weight analysis (RWA) apportion the influence of temperature, precipitation, and evaporation. Under SSP2-4.5 and SSP5-8.5, annual and seasonal ADTTs rise basin-wide; moderate thresholds increase most - so by 2100, a one-class weaker meteorological drought (SPI-3 approximate to -1.0) can trigger agricultural drought. Converging thresholds across severity classes foreshadow more frequent, short, extreme soil-moisture deficits. Temperature dominates annual ADTT inflation, whereas precipitation and evaporation govern most seasonal responses. These dynamic, driver-specific thresholds provide actionable metrics for early warning, crop-insurance triggers, and adaptive water allocation, and the framework is transferable to other monsoon-influenced basins. The results inform management of two national inter-basin water transfer schemes traversing the UMHRB.

Detached breakwaters, a widely adopted coastal protection measure, effectively mitigate coastal erosion. Previous studies have primarily focused on the hydrodynamic and morphodynamic characteristics of detached breakwaters; however, the impact of beach morphology changes induced by detached breakwaters on groundwater flow and solute transport in coastal aquifers remains poorly understood. This study investigated the impacts of beach morphology changes behind detached breakwaters on groundwater flow and salinity distribution in coastal unconfined aquifers by constructing an idealized numerical model featuring a stable salient landform. The results indicate that the salient shifts the tide-driven upper saline plume seaward and expands its distribution range, while shortening the seawater intrusion distance. Although beach morphology changes significantly reduce the salinized volume and total salt mass, the volume of saltwater-freshwater mixing zone has greatly increased. Beach morphology changes also alter the average particle transit time in the freshwater discharge zone: transit time increases within the deposition area of the beach, and transit time decreases slightly within the scouring area of the beach. Compared to the flat beach with a fixed slope, the salient enhances aquifer-ocean exchange fluxes, increases tide-driven recirculation percentages, and decreases density-driven recirculation percentages. Overall, compared to areas outside the sheltered area, submarine groundwater discharge and seawater intrusion within the sheltered area of detached breakwaters are more sensitive to inland freshwater flux, tidal amplitude, and hydraulic conductivity. This study highlights the complexity of coastal unconfined aquifer systems under the influence of detached breakwaters. It provides novel theoretical insights into how detached breakwater-induced beach morphology changes regulate seawater intrusion processes, advancing our mechanistic understanding of land/sea sourced solute transport dynamics in subterranean estuaries.

In recent years, the irregular wave characteristics of ocean dynamics have often been overlooked in the study of the driving mechanism of groundwater movement in coastal aquifers. To clarify the propagation mechanisms of groundwater fluctuations driven by irregular waves in beach aquifers, we employed spectral analysis based on numerical simulations to examine the energy migration processes and evolution characteristics of wave signals at different frequencies. It elucidates the response mechanism of groundwater movement characteristics (head, velocity) to irregular waves in the sea. The energy density in the low-frequency region is enhanced compared to the incident wave and continuously increases in the direction away from the sea within the aquifer. The wavelet power corresponding to the 1/2 spectral peak frequency is significantly enhanced. The energy density in the high-frequency region is generally weaker than that of the incident waves, and the wavelet power corresponding to double spectral peak frequency is enhanced. The correlation between incident waves and groundwater fluctuations is highest near the spectral peak period. This study addresses some problems in modeling surface water–groundwater interactions under irregular wave conditions and provides a theoretical reference for investigating the impacts of extreme climate events (such as typhoon waves and low-frequency offshore oscillations generated by storm surges) on seawater intrusion into coastal groundwater systems.

The distributed hydrologic models (DHMs) evolved from lumped hydrologic models, inheriting their modeling philosophy along with persistent numerical-error issues. Historically, these models tend to use established one-dimensional (1D) methods for slope concentration, which often struggle to effectively represent complex terrains. In this study, we formulated a purely differential form of mathematical equations for the distributed Xinanjiang model and developed a fully coupled numerical solution framework. We also introduced two-dimensional (2D) diffusion wave equations for surface slope concentration and derived 2D linear reservoir equations for subsurface slope concentration to replace their 1D counterparts. This culminated in the development of a two-dimensional differential form of the distributed Xinanjiang (TDD-XAJ) model. Two numerical experiments and the application of the TDD-XAJ model in a humid watershed were conducted to demonstrate the model. Our results suggested that (a) numerical errors in the existing distributed Xinanjiang model are significant and may be exacerbated by a potential terrain amplification effect, which could be effectively controlled by the fully coupled numerical framework within the TDD-XAJ model; (b) the 2D slope concentration methods showed enhanced terrain capture ability and eliminated the reliance on the flow direction algorithms used in 1D methods; and (c) the TDD-XAJ model exhibited improved simulation capabilities compared to the existing model when applied in the Tunxi watershed, particularly for flood volume. This study emphasizes the need to revisit DHMs which stem from lumped hydrological models, focusing on model equations and numerical implementations, which could enhance model performance and benefit the hydrological modeling community.

The evolution of hydrological processes was highly nonlinear due to climate change, human activities, and changes in the watershed subsurface. Typical deep learning (DL) schemes still had limitations in terms of accuracy, stability, and computational complexity. In this study, a novel design based on the “encoding-decoding” architecture was adopted, drawing upon recent advances in the field of artificial intelligence. Incorporating a wavelet feature extractor, the Transformer encoder module was retained to encode the features of the input sequences, while a new module, TimesBlock, was introduced to realize runoff prediction. A new model called WaveTransTimesNet (WTTN) was proposed. The experimental results demonstrated that WTTN exhibited superior prediction performance, excellent generalization ability, and strong robustness, with the Kling-Gupta Efficiency coefficient (KGE) reaching 0.94, 0.95, and 0.96, respectively. Compared to other benchmark models, the proposed model not only identified and utilized periodic trend features in the runoff series more effectively but also predicted runoff more accurately. Additionally, it performed exceptionally well in peak prediction.

This study evaluates the Nielsen set-up empirical equation's applicability for groundwater dynamics under irregular waves through numerical simulation, comparing phase-resolving and phase-averaged methods under three peak enhancement factor (gamma) conditions. Key findings demonstrate that both numerical methods generate stronger horizontal pressure gradients than the empirical equation, yielding larger upper saline plume (USP) features (area, salt content) and greater saltwater penetration depths. The phase-averaged method produces more persistent pressure gradients, resulting in enhanced USP characteristics compared to the phase-resolving method, which promotes broader salt-freshwater mixing through wave-induced water level fluctuations. Reduced gamma values correlate with more uniform wave energy distribution, amplifying swash extent and pore pressures, thereby intensifying USP quantities and method-dependent discrepancies. Interface flux analysis reveals phase-averaged conditions yield higher peak infiltration (30 % greater) and marginally stronger submarine groundwater discharge (SGD) than phase-resolving results, with gamma reduction further accentuating these differences. The set-up empirical equation consistently underestimates flow and salt dynamics by 21 similar to 53 % across main metrics (USP area: similar to 53 %; mixing area: similar to 49 %; SGD: similar to 24 %; total influx: similar to 21 %), demonstrating its unreliability for wave-groundwater coupling systems. These findings strongly support the need for physics-based numerical approaches in coastal aquifer management, particularly for regions experiencing intensified wave variability. The study establishes gamma as a key modulator of USP dynamics and method selection criteria for subsurface transport modeling.

Understanding the link between landscape pattern and water conservation service (WCS) is crucial for effective water resource conservation and landscape planning. However, the impact of landscape configuration on WCS remains unclear. This study explored the spatiotemporal relationships between WCS and landscape pattern in the Zoige Plateau (ZP), a vital waterhead of the yellow river, from 1990 to 2020. The results showed that woodland and high-coverage grassland exhibited the highest WCS capacity and contributed the greatest to total WCS, suggesting the importance of protecting coverage and area of woodland and grassland. From 1990 to 2020, the areas of high- and medium-coverage grasslands significantly decreased, implying adverse effects of landscape composition changes on ZP's WCS during the past few decades. The effects of landscape configuration metrics on WCS exhibited significant variability across years, climate scenarios and regions, highlighting the significant role of climatic and underlying surface. Under realistic climatic scenarios, for greater WCS, in most parts of the western and northern ZP, which negatively correlated with aggregation index, patch density, and mean fractal dimension index but positively correlated with shannon's diversity index, it is better to increase the proportion of large patches, reduce landscape fragmentation, and maintain diverse patch types. Conversely, in southeastern part of the central ZP, where WCS showed an opposite correlation with the aforementioned metrics, maintaining simpler patch type diversity and higher landscape connectivity may be useful. These findings provided important insights for identifying sensitive areas where landscape pattern changes affect WCS and optimizing landscapes to sustain ZP's WCS.

Flooding in riverine basins remains a recurring disaster, often leading to extensive property destruction and, in extreme scenarios, loss of lives. In recent years, the Shouchang River Basin in Zhejiang Province, China, has experienced increasing flood risks, driven by a combination of extreme weather events, urban expansion, and alterations in natural land use. Managing these events is becoming increasingly crucial to minimize the impact on vulnerable communities and critical infrastructure. This study develops an integrated framework for flood forecasting and hydrodynamic floodplain mapping using HEC-HMS and HEC-RAS 6.5 over a 10 km stretch of the Shouchang River upstream of Shouchang Town. The hydrological model (HEC-HMS) simulates rainfall-runoff processes across five sub-basins, using observed rainfall and streamflow data from four gauging stations, to capture key flow dynamics. Based on local plans for Shouchang Town, a total of 28 villages are situated within exposure areas of sub-basin 5. Out of villages only 22 rescue centers are found to be unaffected and thus effective for sheltering flood victims. Four rescue centers, Yongjiaqiao, Henanli, Ximen, and Datanbian would need relocation to higher grounds, including adding new resettlement sites and modifying transfer plans and routes. Simulations show that, while flood defenses protect most regions under upstream flows of 1,200 m3/s, the levees along Shili Shouchngjiang Ecology Leisure Greenway breach once this threshold is surpassed. The study highlights the need to review the existing flood evacuation and transfer analysis system, given that some evacuation centers could be exposed to flood associated risks.

Fractured coastal aquifers are particularly susceptible to saltwater intrusion (SWI), and the effects of fractures on SWI affect the fate of land-sourced and sea-sourced nutrients in coastal aquifers, but the reactivity of biogeochemical process in such aquifers remains largely unknown. A numerical variable-density groundwater flow and reactive transport model was used to evaluate the effects of fracture characteristics (vertical position of horizontal fracture, horizontal fracture length, vertical fracture length, and horizontal position of vertical fracture) on the reactivity of mixing-dependent and mixing-independent reactions (represented as denitrification and sulfate reduction, respectively) in fractured coastal aquifers. Simulation results show that denitrification removes up to 64%-92% of land-sourced nitrate, and sulfate reduction transform only about 5 parts per thousand of sea-sourced sulfate prior to discharge. Under the model settings and conditions tested, the simulation results indicate that the fracture characteristics of the aquifer are a major factor in regulating the behavior of both mixing-dependent and mixing-independent reactions. Horizontal fractures promote or attenuate denitrification and sulfate reduction primarily by affecting the area of the upper saline plume, and reactant solutes mixing. Vertical fractures primarily affect reactant solutes mixing to promote or attenuate denitrification and sulfate reduction. Mixingdependent reactivity increases with the size of the mixing zone and solute supply, while mixing-independent reactivity is mainly controlled by solute supply and area of the upper saline plume. These results have important implications for understanding biogeochemical processes in fractured coastal aquifers and for the management of coastal ecosystems.

Remote sensing data primarily provide insights into surface soil moisture (SM). By integrating auxiliary data such as topography and meteorology, deep learning (DL) can enable the estimation of SM beneath the soil surface. However, DL models often overlook underlying physical mechanisms and lack interpretability, which may compromise their reliability in predicting SM. In this study, we introduce a physics-guided DL (PGDL) approach for estimating multi-layer SM. Our method utilizes convolutional neural networks (CNN) to analyze spatial features and long short-term memory networks (LSTM) to capture temporal dynamics, effectively modeling the spatiotemporal characteristics of SM. Crucially, we integrate the Richardson-Richards equation, which describes the spatiotemporal dynamics of SM, as a constraint within the CNN-LSTM model's loss function. Additionally, we constrain the LSTM activation function to keep estimated values within the residual and saturated SM range. This integration guides the model to learn physical characteristics, thereby enhancing accuracy and interpretability. Our method (named PGDL-CNN-LSTM) outperforms other machine learning methods (CNN-LSTM, extreme gradient boosting, and random forest) across multi-layer (10 cm, 20 cm, 30 cm, 40 cm, and 50 cm). Specifically, for 10 cm SM, our method achieves a 9.6 % reduction in root mean square error and a 92.0 % reduction in physical inconsistency compared to CNN-LSTM alone. Furthermore, we extend surface SM to estimate root-zone SM at 10 km resolution. Our method accurately captures dynamic changes in SM and remains effective even with a 50 % reduction in data. In conclusion, integrating physical constraints significantly enhances the depth and interpretability of SM estimation from remote sensing data. These findings highlight the potential of our method for various hydrological applications.

Basin-scale runoff forecasting requires controlling absolute relative errors within 0.2 for accurate predictions. This study examines the sensitivity of simulated flood deviations in frequently flooded humid regions to discrepancies in precipitation inputs driving the Weather Research and Forecasting (WRF)-Hydro model. Key parameters of various WRF-Hydro modules are calibrated from 23 flood events between 2003 and 2017. Two precipitation experiments were designed with minimized uncertainty in both model and parameterization to explore the sensitivity of streamflow changes to precipitation misestimations, both in total and graded precipitation. The sensitivity of runoff relative errors is more pronounced than that of precipitation relative errors, influenced by the magnitude, variability, and duration of rainfall. During processes involving heavy rainstorm, precipitation absolute relative errors increased by 50%, while runoff absolute relative errors increased by more than 60%. This sensitivity trend is linked to variations in components generated by misestimated precipitation. Runoff relative errors are higher when precipitation is misestimated at high rainfall levels compared to the same misestimation at low rainfall levels. Consequently, for effective flood simulation and prediction, it is crucial to emphasize the accuracy of precipitation data, especially during high rainfall events. Future research will incorporate more realistic precipitation forcing mechanisms and additional metrics for evaluating precipitation forecasts.

This study examined the influence of preferential flow on pore water flows and marine nitrogen transport reaction in variable saturation and variable density coastal aquifers. The 2-D unconfined aquifer model established was based on the software COMSOL by coupling the dynamic and chemical processes together. The results showed that preferential flow affects groundwater flow and salinity distribution, leading to a more complicated mixing process. The preferential flow resulted in an increase in mixing zone area and the upper saline plume area of 10.33 and 2.62 m2, respectively, a decrease in saltwater wedge area of 7.22 m2, and an increase in nitrate (NO3-) removal efficiency from 7.9% to 8.97%. The NO3- removal efficiency increases progressively with the depth (h) and quantity (n) of preferential flows; however, it decreases after a certain quantity. Further quantitative analysis revealed an increase in the intensity of nitrification and dissolved oxygen inflow flux with preferential flow depth and quantity increase. This phenomenon usually occurs on coasts where biological caves are abundant. The results also offer significant implications for designing engineering measures to mitigate saltwater intrusion and are significant to prevent groundwater quality deterioration in coastal zones.

Flood forecasting in data-scarce catchments is challenging for hydrologists. To address this issue, a regional long short-term memory model (R-LSTM) is proposed. Given the diverse physical characteristics of sub-catchments, this model scalarises the runoff data based on catchment attributes including area, confluence path length, slope and minimum and maximum runoff values, thereby eliminating the local influence and producing a geomorphological-runoff factor as the model input. To assess the effectiveness of R-LSTMs for flood forecasting in data-scarce basins, the Jiaodong Peninsula in China was selected as the study area. The proposed R-LSTMs are compared with local LSTMs, regional LSTMs that do not use catchment attributes, or regional LSTMs that incorporate catchment attributes in different ways. The results show that R-LSTMs outperform the benchmarking LSTM models, especially in the simulation of flood peaks. The study indicates the potential of regionalization and the benefit of building the scalarised inputs of runoff data for regional LSTM that consider catchment attributes meticulously. The research findings can provide a reference for flood forecasting in data-scarce regions. Hydroclimatic consistent area contains multiple basins (nested and independent). A generalised scalarisation method is established to eliminate the influence of local hydrologic and geomorphologic factors; and the scalarised factors are integrated based on hydrological insights, which allows data aggregation across catchments. The unified dataset serves as the input for regional LSTM model, which can be adopted in data-sparse basins within the consistent area. image

Subsurface dams have been recognized as one of the most effective measures for preventing saltwater intrusion. However, it may result in large amounts of residual saltwater being trapped upstream of the dam and take years to decades to remove, which may limit the utilization of fresh groundwater in coastal areas. In this study, field-scale numerical simulations were used to investigate the mechanisms of residual saltwater removal from a typical stratified aquifer, where an intermediate low-permeability layer (LPL) exists between two high-permeability layers, under the effect of seasonal sea level fluctuations. The study quantifies and compares the time of residual saltwater removal (Tre) for constant sea level (CSL) and seasonally varying sea level (FSL) scenarios. The modelling results indicate that, in most cases, seasonal fluctuations in sea level facilitate the dilution of residual saltwater and thus accelerate residual saltwater removal compared to a static sea level scenario. However, accounting for seasonal sea level variations may increase the required critical dam height (the minimum dam height required to achieve complete residual saltwater removal). Sensitivity analyses show that Tre decreases with increasing height of subsurface dam (Hd) under CSL or weaker sea level fluctuation scenarios; however, when the magnitude of sea level fluctuation is large, Tre changes non-monotonically with Hd. Tre decreases with increasing distance between subsurface dam and ocean for both CSL and FSL scenarios. We also found that stratification model had a significant effect on Tre. The increase in LPL thickness for both CSL and FSL scenarios leads to a decrease in Tre and critical dam height. Tre generally shows a non-monotonically decreasing trend as LPL elevation increases. These quantitative analyses provide valuable insights into the design of subsurface dams in complex situations.

The development of artificial intelligence has introduced new perspectives to the field of hydrological forecasting. However, there is still a lack of research on efficiently identifying the physical characteristics of runoff sequences and developing prediction models that consider global and local sequence features. This study proposes a parallel computing prediction model called IMCAEN (Integrated Multi-Feature Causal Dilated Convolutional Attention Encoder Network) to address these issues. Unlike existing models, this model can monitor fluctuations and anomalies in time series. Incorporating the CDC-AA (Causal Dilated Convolutional Network with Aggregation Attention) and encoder structure captures both local sequence variations and global abrupt anomalies, allowing for comprehensive attention to sequence features. When predicting runoff data from three different hydrological conditions, the IMCAEN model achieved NSEC (Nash-Sutcliffe Efficiency Coefficient) values of 0.98, 0.97, and 0.88, respectively, and outperformed benchmark models in other evaluation indicators as well. Given the opacity of the feature distribution process in AI models, SHAP (SHapleyAdditive exPlanations) analysis and spatial expression of feature distribution are used to assess the contribution of each feature variable to the long-term trend of runoff and to verify the distribution of features trained in each module. The proposed IMCAEN model efficiently captures local and global information in the runoff evolution process through parallel computing and shared features, enabling accurate runoff forecasting and providing critical references for timely warnings and predictions.