Deep learning has opened new opportunities for streamflow modeling, enabling the extraction of complex hydrological patterns from large-sample observations. However, large-scale streamflow prediction remains challenging, requiring robust generalization across diverse hydroclimatic conditions and reliable transfer to ungauged regions. To address these challenges, we propose a MAML-EA-LSTM framework that integrates Model-Agnostic Meta-Learning (MAML) with an entity-aware Long Short-Term Memory (EA-LSTM) architecture. The framework is evaluated using 468 basins from the CAMELS dataset. Results show that the proposed MAML-EA-LSTM achieves a mean Nash-Sutcliffe Efficiency (NSE) of 0.75, outperforming LSTM (0.70), EA-LSTM (0.69), MAML-LSTM (0.71), and the conceptual Sacramento Soil Moisture Accounting (SAC-SMA) model (0.57), with improvements observed in most basins. Reductions in both high-flow and low-flow bias further indicate enhanced predictive stability across contrasting hydrological conditions. Dominant attribute identification analysis shows that MAML-EA-LSTM also improves hydrological behavior learning, yielding an overall hit rate of 71.2%, higher than LSTM (43.6%), EA-LSTM (60.7%) and MAML-LSTM (51.5%). Transfer experiments under strictly spatially disjoint conditions further demonstrate stronger out-of-sample transferability in data-scarce basins. The proposed framework achieves an NSE of 0.553 with only one year of target-basin observations for fine-tuning. These findings indicate that the integration of structural inductive bias and meta-learning provides a promising pathway toward more robust and transferable large-sample hydrological modeling.

Accurate early warning of flash floods is critical for prompt decision-making in mitigating disaster impact. However, most current applications of flash-flood warning are based on deterministic approaches, and the inherent uncertainty that exists has not been fully considered. This study proposed a probabilistic flash-flood warning approach by incorporating hydrological modelling uncertainty. The Monte Carlo (MC)-based parameter selection method, together with probability density analysis, was used in assessing the probability of warning criteria being exceeded. Moreover, an optimal decision rule was introduced to enhance the reliability of the flash-flood warning. The results show that the proposed approach provides more informative results by generating the probability distribution estimation and probabilistic thresholds, enabling the user to choose their own decision rule. The probabilistic approach with the optimal threshold has a better performance (CSI = 0.58) than the deterministic approach (CSI = 0.41), especially in the reduction of the number of false alarms (from 37 to 19 events), which shows better reliability and confidence. The results highlight the improvement of the proposed approach by incorporating the uncertainty in hydrological modelling, which can effectively quantify the potential impact risk and aid decision-making to issue warnings. Specifically, a range of possible outcomes are transformed into actionable decisions for issuing reasonable flash-flood warnings with a lead time of 1-3 h. This study provides new insights into the application of the probabilistic approach in flash-flood warning and is expected to enhance practical applications.

Accurate prediction of root zone soil moisture (RZSM) is critical for advancing hydrological modeling and water cycle characterization. To improve RZSM estimation in ungauged regions and elucidate the role of catchment attributes in RZSM dynamics in time and space, this study proposed a novel regionalization framework that integrates catchment attribute classification with surface soil moisture (SSM) similarity metrics. We investigate the viability of extrapolating RZSM data from gauged to ungauged catchments, with emphasis on the adaptability of the Soil Moisture Analytical Relationship (SMAR) model and the influence of catchment attributes on prediction performance. The results show that the calibrated SMAR model effectively simulates RZSM patterns, achieving a mean root mean square error (RMSE) of 0.040 cm³/cm³ for the validation periods. Additionally, the results reveal significant disparities between SSM and RZSM dynamics across the catchment, underscoring the pronounced influence of catchment attributes on SSM-RZSM coupling. Notably, parameter regionalization strategies combining catchment attribute-based site grouping, including topographic wetness index (TWI), soil depth, and leaf area index (LAI), produced more accurate RZSM predictions (mean RMSE = 0.081 cm³/cm³) than results from relying solely on SSM similarity (mean RMSE = 0.145 cm³/cm³). The superior performance of TWI-based groupings highlights topography’s essential role in modulating nonlinear SSM-RZSM relationships. These insights underscore the interdependence between soil moisture dynamics and catchment attributes in headwater catchments, illustrating the value of catchment physiographic features in constraining predictive uncertainty for RZSM in ungauged regions.

Obtaining accurate information regarding root zone soil moisture (RZSM) is a critical element of effective hydrological and agricultural management practices. Previous studies have relied on surface soil moisture (SSM) values, which are more easily measured, to estimate RZSM using the Soil Moisture Analytical Relationship (SMAR) model or regression method. However, the performance of these two types of methods in areas with complex topography still needs more exploration. Here, we assess the accuracy of these two types of methods in a forested mountainous catchment, using daily SSM measurements from 32 monitoring sites. The results show that both methods are capable of accurately estimating RZSM with a high NSE (>0.950) during the validation period. Additionally, they exhibit excellent model transferability at ungauged sites. Spatially, both methods perform better in drier areas than in wetter areas. Temporally, both methods are better in the wet–cold season than in the dry–warm season. Overall, both methods demonstrate comparable performance in the catchment, with NSE values of 0.986 and 0.951 during the validation period, respectively. The regression method is more suited to complex hydropedological environments characterized by long-term soil moisture monitoring and nonlinear hydropedological behaviors. Conversely, the SMAR model is better suited for flat areas and less spatial variability in microtopography. Moreover, the estimation of RZSM by both methods is influenced not only by soil moisture conditions but also by local factors including terrain topography, soil depth, and the degree of subsurface hydrological connectivity. This study adds to our understanding of RZSM estimation from SSM in complex terrain and will act as a reference for selecting appropriate methods of RZSM estimation. The results of this study underscore a discernible relationship between surface and deep soil moisture across varying spatial and temporal scales.

Accurate flash flood warning is crucial for hazard risk management. Rainfall threshold is one of the most effective approaches for flash flood early warning. However, the capability of different rainfall-threshold methods has been insufficiently evaluated and validated, and the influence of data limitations on the model performance still needs to be explored. In this work, we comprehensively evaluate and compare three rainfall-threshold methods (empirical, hydrological, and probabilistic) based on soil moisture conditions for flash flood warnings in two small mountainous catchments with different data features. Results show that all three methods obtain good accuracy (Hit Rate (HR) > 0.80, False Alarm Rate (FAR) < 0.20) in both catchments, indicating that rainfall threshold is an efficient warning approach for flash flood management. Especially, the hydrological method shows the best predictive performance, with the highest Critical Success Index (CSI) (0.61 and 0.41, respectively in the two study catchments) and the lowest FAR (0.07 and 0.12), suggesting it is the most appropriate method for catchments where the hydrological model can be properly calibrated. Empirical and probabilistic methods are not far behind, showing comparable performance (CSI = 0.53 and 0.53, respectively for the Qingxi River (QXH) catchment). Catchment data availability, which reflects the representativeness of the relevant flash flood process conditions, greatly impacts the forecasting and model application. The results of this work improve our understanding of the applicability and reliability of these methods under various conditions, and aid in selecting or establishing a robust rainfall-threshold method for flash flood warning under complex environments.

Advanced warning of flash flooding plays a vital role in mitigating risk. Established warning models focus on deterministic indicators and do not consider the probability of flash flood occurrences. This study proposed a probabilistic framework based on CNN-LSTM-MultiHead-Attention (CLMA) by considering the threshold optimization for the flash flooding warnings, and compared it with a deterministic method based on hydrological modeling. In addition to the pivotal factor of rainfall, the proposed model also incorporated antecedent precipitation index (API) and rainfall pattern (RP) as input factors to explore the hidden patterns of flash flood occurrences. The results show that the proposed method gave more informative results than the deterministic method by additionally generating hourly probability values. This greatly reduced false and missing alarms and enabled warnings to be issued average 1-3 h in advance. Meanwhile, incorporating API and RP data into the CLMA model enhanced its ability to forecast flash flooding, with API having a greater effect than RP. Moreover, probabilistic threshold optimization using the Best_DIFF criterion improved the CLMA model, decreasing the Euclidean Distance (ED) from 0.12 to 0.05 and increasing the Critical Successful Index (CSI) from 0.54 to 0.83. This study demonstrates that deep learning has the potential to be applied within a probabilistic framework for the flash flood forecasting, which can effectively assess risk information and support early decision-making to issue warnings and flood control actions.

Accurate streamflow predictions in mountainous regions are crucial for water resource management and flood mitigation. Deep learning (DL) models, which have been widely used for streamflow predicting recently, can simulate the nonlinear hydrological relationships but may not capture the underlying laws of physics. This study proposed a Causality-Guided Deep Learning (CGDL) model to enhance the streamflow predicting for mountainous regions by incorporating physics-based causal inference and improved multivariate Transfer Entropy (IMTE) algorithm. We assessed the CGDL model through a case study in a mountainous catchment using in situ hydrometeorological variables (precipitation, temperature, and humidity, etc.). The results demonstrated that CGDL outperformed DL and process-based (PB) models, achieving a higher NSE (Nash-Sutcliffe Efficiency) of 0.805, compared to 0.701 for DL and 0.716 for PB during the testing period. Furthermore, CGDL significantly reduced the EHF (Error of High Flow) to-9.7%, versus-14.9% for DL and -22.0% for the PB model in the testing period, highlighting its efficiency in high flow predictions. The CGDL also showed superior robustness and generalization when extending forecast lead time and simulating beyond the bounds of the training data. Additionally, the SHAP analysis indicated that CGDL provided greater interpretability than the DL model. This study demonstrates that integrating causality knowledge into deep learning models has the potential to enhance streamflow predicting in mountainous regions. It is helpful for improving our understanding of hydrological processes and decision-making to issue flood warnings.

Understanding the spatiotemporal variability of soil moisture is essential for advancing hydrological, ecological, and agricultural management practices. This study investigates soil moisture dynamics across three distinct hydroclimatic periods (wet, dry, and wet-dry transitions) in a headwater catchment in central Pennsylvania, USA, characterized by heterogeneous soils and variable topography. We aim to elucidate key influencing factors and optimize catchment-scale monitoring strategies. Using the Index of Temporal Stability (ITS), we assessed the temporal stability of soil moisture and identified representative locations. Redundancy Analysis (RDA) was applied to explore the impacts of influencing factors on soil moisture variability. Results revealed that dominant controls varied substantially across periods and soil horizons. The temporal stability of soil moisture was the highest during wet periods, primarily governed by topographic attributes, but decreased during dry and transitional periods. During dry periods, soil moisture dynamics were influenced by a combination of soil composition and slope characteristics, while transitional periods exhibited strong sensitivity to Preferential Flow Frequency (PFF) and slope attributes. Among time-variant factors, air temperature and evapotranspiration emerged as critical regulators of soil moisture dynamics. A single representative location, characterized by Topographic Wetness Index (TWI) values close to the catchment's mean and relatively low PFF, provided robust estimates of areal mean soil moisture across the study period (R-2 > 0.86). However, additional monitoring locations were needed to maintain accuracy during dry and transitional periods. These findings improve the understanding of soil moisture variability in forested catchments and offer insights for designing cost-effective monitoring networks to support hydrological modeling and management.

Spatiotemporal dynamics of soil moisture and its hysteresis behavior are crucial for the integration of soil, vegetation, and hydrological processes in headwater catchments. While extensive research has been conducted on soil moisture patterns within distinct dry or wet seasons, the transitions between these seasons have received relatively little attention. Therefore, exploring the variability in soil moisture and its hysteresis characteristics during these transitional periods is imperative. This study delves into the relationship between spatial mean 8 and standard deviation sigma theta (sigma theta(8)) of soil moisture over both time and space, utilizing three years of daily monitoring data from 33 sites within a 0.08 km2 headwater catchment in Pennsylvania, USA, with a focus on the seasonal transition periods. Through empirical orthogonal function (EOF) analysis, key factors that influence the occurrence of the hysteresis cycles in the sigma theta(8) relationship were identified. The findings of this study reveal a linear trend in the sigma theta(8) relationship over time during transitional periods, contrasting with behavior observed within the dry and wet seasons. Notably, hysteresis cycles in the sigma theta(8) relationship are present in each season, predominantly during wetter periods. These cycles are categorized into two types: those without reorganization, featuring wetting-up and drying-down phases, and those with reorganization, which also include a reorganization phase. The first two EOFs explain 75.3 % of the variation in hysteresis cycles and exhibit significant correlations with soil characteristics, topography, and vegetation. More specifically, topography and soil properties primarily influence soil moisture patterns during the transition from wet to dry seasons, whereas vegetation and topography are more influential during the transition from dry to wet. The observed spatiotemporal variability and the occurrence of hysteresis cycles in soil moisture supplement the behaviors and mechanisms of soil moisture in seasonal transitions in forested headwater catchments.

Soil moisture data assimilation (SM-DA) is a valuable approach for enhancing streamflow prediction in rainfall-runoff models. However, most studies have focused on incorporating remotely sensed SM, and their results strongly depend on the quality of satellite products. Compared with remote sensing products, in situ observed SM data provide greater accuracy and more effectively capture temporal fluctuations in soil moisture levels. Therefore, the effectiveness of SM-DA in improving streamflow prediction remains site-specific and requires further validation. Here, we employed the Ensemble Kalman filter (EnKF) to integrate daily SM into lumped and distributed approaches of the Xinanjiang (XAJ) hydrological model to assess the importance of SM-DA in streamflow prediction. We observed a general improvement in streamflow prediction after conducting SM-DA. Specifically, the Nash-Sutcliffe efficiency increased from 0.61 to 0.65 for the lumped and from 0.62 to 0.70 for the distributed approaches. Moreover, the efficiency of SM-DA exhibits seasonal variation, with in situ SM proving particularly valuable for streamflow prediction during the wet-cold season compared to the dry-warm season. Notably, daily SM data from deep layers exhibit a stronger capability to improve streamflow prediction compared to surface SM. This indicates the significance of deep SM information for streamflow prediction in mountain areas. Overall, this study effectively demonstrates the efficacy of assimilating SM data to improve hydrological models in streamflow prediction. These findings contribute to our understanding of the connection between SM, streamflow, and hydrological connectivity in headwater catchments.

Satellite-based precipitation products (SBPPs) are essential for rainfall quantification in areas where ground-based observation is scarce. However, the accuracy of SBPPs is greatly influenced by complex topography. This study evaluates the performance of Integrated Multi-satellite Retrievals for GPM (IMERG) and Global Satellite Mapping of Precipitation (GSMaP) in characterizing rainfall in a mountainous catchment of southwestern China, with an emphasis on the effect of three topographic variables (elevation, slope, aspect). The SBPPs are evaluated by comparing rain gauge observations at eight ground stations from May to October in 2014–2018. Results show that IMERG and GSMaP have good rainfall detection capability for the entire region, with POD = 0.75 and 0.93, respectively. In addition, IMERG overestimates rainfall (BIAS = −48.8%), while GSMaP is consistent with gauge rainfall (BIAS = −0.4%). Comprehensive analysis shows that IMERG and GSMaP are more impacted by elevation, and then slope, whereas aspect has little impact. The independent evaluations suggest that variability of elevation and slope negatively correlate with the accuracy of SBPPs. The accuracy of GSMaP presents weaker dependence on topography than that of IMERG in the study area. Our findings demonstrate the applicability of IMERG and GSMaP in mountainous catchments of Southwest China. We confirm that complex topography impacts the performance of SBPPs, especially for complex topography in mountainous areas. It is suggested that taking topographical factors into account is needed for hydrometeorological applications such as flood forecasting, and SBPP evaluations and retrieval technology require further improvement in the future for better applications.

黄土高原地区由于其短历时高强度的暴雨特性引起洪水陡涨陡落，不仅影响防洪安全，也使得预报难度加大。为避免单一模型预报结果的不确定性，以岔巴沟流域作为典型代表，采用多个模型（水箱模型、NAM模型、陕北模型）对其进行洪水预报研究，并对比分析了不同模型的应用效果。结果表明，三个模型在该地区均有一定的适应性；无论对洪峰或洪水过程线，水箱模型模拟效果整体均为最优，且具有更好的稳定性；对部分大洪水，NAM模型对于峰现时间把控更好，模拟精度更高；陕北模型总体对于涨洪阶段模拟效果更好。研究结果对岔巴沟流域防洪减灾具有重要的意义，亦可为黄土丘陵沟壑区洪水预报及水土保持等工作提供参考依据。

鉴于山洪突发性强、历时短、陡涨陡落等致使在模拟预报过程中具有较大难度和不确定性问题,构建了基于深度学习的LSTM网络模型进行山洪确定性预报和概率预报,从精度和可靠度两方面研究其在西南山区的适用性.并以西南山洪易发区寿溪河流域为例进行模拟,结果显示LSTM网络模型更易发现暴雨洪水之间的深层规律,验证期平均纳什效率系数达0.954,与BP模型相比,显著提升了洪水预报精度,尤其是大洪水;概率预报有效降低了山洪预报的不确定性,洪峰附近的流量数据基本落入预报区间内,有效提高了预报可靠度.

Suspended sediment transport is one of the essential processes in the geochemical cycle. This study investigated the role of rainfall thresholds in suspended sediment modeling in semiarid catchments. The results showed that rainfall-sediment in the study catchment (HMTC) could be grouped into two patterns on the basis of rainfall threshold 10 mm. The sediment modeling based on LSTM model with the rainfall threshold (C-LSTM scheme) and without threshold (LSTM scheme) were evaluated and compared. The results showed that the C-LSTM scheme had much better performances than LSTM scheme, especially for the low sediment conditions. It was observed that in the study catchment, the mean NSE was marginally improved from 0.925 to 0.934 for calibration and 0.911 to 0.924 for validation for medium and high sediment (Pattern 1); while for low sediment (Pattern 2), the mean NSE was significantly improved from -0.375 to 0.738 for calibration and 0.171 to 0.797 for validation. Results of this study indicated rainfall thresholds were very effective in improving suspended sediment simulation. It was suggested that the incorporation of more information such as rainfall intensity, land use, and land cover may lead to further improvement of sediment prediction in the future.

水沙模型是计算和评估水土流失和泥沙侵蚀的重要工具.黄土高原是我国土壤侵蚀最严重的地区之一,针对黄土高原地区水沙特点,耦合降雨径流模块、土壤侵蚀模块和泥沙输移模块,构建了基于物理机制的水沙模型,其中降雨径流模块采用水箱模型,土壤侵蚀模块包括降雨溅蚀、梁峁坡侵蚀、沟谷坡侵蚀、沟道侵蚀,泥沙输移模块采用泥沙平衡方程与蓄泄方程联合方法.以黄河支流无定河的西南部黄土丘陵沟壑区岔巴沟流域作为典型代表进行水沙模拟,并将模拟结果与MUSLE(Modified Universal Soil Loss Equation)模型进行对比,探究不同的水沙模型对于模拟结果的影响.模拟结果表明:①构建的水沙模型详细考虑黄土高原侵蚀过程特点,物理机制更明确,侵蚀计算更加全面.②构建的水沙模型具有更高的精度和可靠性,优于基于经验公式MUSLE模型,其模型平均纳什效率系数NSE为0.722,平均沙量相对误差绝对值为21.3％;基于MUSLE的模型平均NSE为0.576,平均沙量相对误差绝对值为32.8％.该研究成果为黄土高原地区泥沙侵蚀过程研究和水沙防治提供一定的科学依据.

Satellite remote sensing precipitation is useful for many hydrological and meteorological applications such as rainfall-runoff forecasting. However, most studies have focused on the use of satellite precipitation on daily, monthly, or larger time scales. This study focused on flash flood simulation using satellite precipitation products (IMERG) on an hourly scale in a poorly gauged mountainous catchment in southwestern China. Deep learning (long short-term memory, LSTM) was used, merging satellite precipitation and gauge observations, and the merged precipitation data were used as inputs for flood simulation based on the HEC-HMS model, compared with the gauged precipitation data and original IMERG data. The results showed that the application of original IMERG data used directly in the HEC-HMS hydrological model had much lower accuracy than that of gauged data and merged data. The simulation using the merged precipitation in HEC-HMS exhibited much better performances than gauged data. The mean NSE improved from 0.84 to 0.87 for calibration and 0.80 to 0.84 for verification, while the lower NSE improved from 0.81 to 0.84 for calibration and 0.73 to 0.86 for verification, which showed that accuracy and robustness were both significantly improved. Results of this study indicate the advances of remote sensing precipitation with deep learning for flash flood forecasting in mountainous regions. It is likely that more significant improvements can be made in flash flood forecasting by employing multi-source remote sensing products and deep learning merging methods considering the impact of complex terrain.

针对目前中小河流洪水预报难度较大、预报精度偏低的问题,以西南山区荥经河流域为例,基于HEC-HMS构建了荥经河流域分布式水文模型,以2009～2018年23场洪水进行参数率定,4场洪水进行模型验证.通过参数敏感性分析和有效率定,模拟的洪水过程与实测洪水过程吻合较好,其中模拟洪峰流量平均误差为6.81％,平均纳什效率系数为0.84.可见构建的HEC-HMS洪水预报模型对荥经河流域有很好的模拟效果,可为西南山区其他中小河流的洪水预报提供借鉴和参考.

山洪灾害对山区人民生命财产、经济建设发展危害极大,有效的山洪预报对降低山洪灾害具有重要作用.以西南山洪易发区——寿溪河流域为研究区,采用HEC-HMS模型构建山洪预报模型.模拟结果显示,洪峰流量相对误差均在20％以内,峰现时差均在3 h以内,相关系数均值达0.9以上,Nash效率系数为0.855,合格率为100％,表明该模型在寿溪河流域适用性较好.由于山区小流域下垫面情况复杂,地形起伏大,因此进一步探究流域DEM空间精度(12.5、30、90 m)对HEC-HMS洪水模拟结果的影响.研究结果表明,DEM分辨率对流域模型参数有一定程度影响,从而导致最终洪水模拟结果的差异;整体而言,DEM分辨率精度越高,寿溪河洪水模拟结果的稳定性及准确性更好,尤其对历时短和降雨过程复杂的场次洪水模拟精度产生的影响更大.

中小河流洪水预报是当前国际水文科学中亟待研究的重大科学问题之一.由于中小河流存在分布广、降水及下垫面空间异致性强、产汇流时间短、突发性强、水文资料欠缺等特点,因此预报精度较低,预见期短,预报难度大.目前国内外学者对中小河流洪水预报开展了一定的研究,主要集中在中小河流洪水形成机理、洪水预报模型、缺资料中小河流水文模型参数确定方法、洪水预报耦合降水预报等方面.对中小河流洪水预报相关研究进展和存在的问题进行了总结,并指出未来在多源信息高效融合、精细化洪水模拟、高精度降水预报等方面还应进行更深入的研究.