Tropical cyclones in southeast Africa pose significant flood risks, driven by high winds that can cause storm surge flooding, and extreme rainfall often leading to devastating inland impacts. Understanding current and future flood risk from extreme rainfall in the region is hindered by the limited observational network. Although reanalysis and satellite rainfall data improve spatial coverage, their coarse resolution restricts the accurate representation of intense convective rainfall associated with tropical cyclones. To address this limitation, fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis (ERA5) reanalysis data were dynamically downscaled to a convection-permitting resolution using the Conformal Cubic Atmospheric Model (CCAM) from 2014 to 2023. This study assesses the ability of kilometer-scale CCAM simulations to improve the characterization of extreme rainfall from six impactful tropical cyclones affecting Malawi, Madagascar, and Mozambique. Comparisons are made against reanalysis data, satellite-derived rainfall, and rain gauge observations. Results show that CCAM captures higher rainfall totals and improves the underestimation bias of extreme hourly rainfall present in the other datasets. Additionally, CCAM reveals detailed rainband structures, including evidence of double rainbands within the eyewall, and inner and outer spiral rainbands. This work has demonstrated that the current reanalysis is not sufficient to properly characterize flood risk from tropical cyclones. The kilometer-scale model provides a more realistic simulation of flood-relevant rainfall. These advancements are essential for understanding and managing flood risk under current and future climate scenarios, supporting adaptive actions to mitigate tropical cyclone impacts.

Abstract. The representation of river routing and floodplain dynamics remains a key limitation in many global Earth system and numerical weather prediction models, despite their importance for hydrological extremes, water resources, and land–atmosphere interactions. This paper presents the implementation of a global hydrodynamic river routing capability within the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS) through the integration of the Catchment-based Macro-scale Floodplain (CaMa-Flood) model into the ecLand land surface scheme. The system enables the routine simulation of river discharge, water level, and inundation extent within a physically consistent Earth system modelling framework. We describe the scientific design and technical developments required to transition from research to operational implementation, including coupling strategies, state initialisation, parallelisation using a basin-based hybrid MPI/OpenMP approach, and integration within ECMWF's Research-to-Operations workflow. The implementation maintains consistency across multiple configurations, from offline experiments to fully coupled forecasts, while preserving operational constraints such as computational efficiency and forecast stability. The added value of the integrated system is demonstrated through a range of applications, including global water resources monitoring contributing to the World Meteorological Organisation State of Global Water Resources Report, real-time ensemble flood forecasting, high-resolution regional simulations, and land-surface diagnostics. Results illustrate that river discharge provides a powerful integrative constraint on the land water cycle and supports the identification of model deficiencies in hydrological processes. The system also enables the generation of continuous, global, high-frequency hydrological datasets suitable for both operational applications and emerging data-driven modelling approaches. This work establishes a new capability for global hydrological prediction within an operational Earth system model and provides a framework for future developments toward fully coupled atmosphere–land–river interactions.

Rainfall-Runoff (R-R) modeling is crucial for hydrological forecasting and water resource management, yet traditional deep learning approaches, such as Long Short-Term Memory (LSTM) networks, often overlook explicit runoff routing, leading to inaccuracies in complex river basins. This study introduces a novel LSTM-Graph Neural Network (GNN) framework that integrates LSTM for local runoff generation with GNN for spatial flow routing, leveraging river network topology as a directed graph. Applied to the Upper Danube River Basin using the LamaH-CE dataset (1987-2017), the model partitions the basin into 530 subbasins and evaluates four GNN architectures: Graph Convolutional Network (GCN), Graph Attention Network (GAT), Graph SAmple and aggreGatE (GraphSAGE), and Chebyshev Spectral Graph Convolutional Network (ChebNet). Results demonstrate that all LSTM-GNN architectures outperform the baseline LSTM, with LSTM-GAT achieving the highest performance (mean NSE = 0.61, KGE = 0.65, Correlation Coefficient = 0.84, RMSE reduction of similar to 35 %). Improvements are most evident in downstream stations with high connectivity and large contributing areas, where adaptive attention in GAT effectively captures heterogeneous upstream influences. These findings underscore the potential of GNN-based approaches for large-scale, spatially aware hydrological modelling and provide a foundation for future applications in flood forecasting and climate adaptation.

More than one billion people are exposed to flood risk globally, with this number projected to double by 2050. Global flood models underpin risk assessment and adaptation planning, yet typically assume that river bankfull capacity corresponds to a fixed two-year return period, neglecting spatial and temporal variability in channel characteristics. Here, we evaluate how inundated areas and population exposures respond when forced with empirically-derived bankfull capacities in the Mississippi basin using the Fathom Global Flood Model. We find that present-day bankfull flows generally correspond to return periods of less than one year, leading to systematic underestimation of flood extent (9-152%) and exposure (15-472%) across 5-, 20- and 100-year flood events, with the largest discrepancies for more frequent floods. We further show that historical changes in channel morphology can influence flood impacts at magnitudes comparable to projected climate change over multi-decadal timescales, depending on emission scenarios. Our work highlights a key structural limitation in current global flood modelling frameworks with implications for risk assessments.

Abstract The new film ‘Pressure’ brings renewed attention to one of the most dramatic weather forecasts ever. In June 1944, James Stagg faced the unenviable task of advising whether conditions would allow the D‐Day landings to proceed, with the lives of 160 000 troops and the fate of the war in the balance. Using AIFS‐ENS, the ECMWF machine‐learning ensemble forecast system, we recreate the forecasts in the days before 6th June. Even at short range, AIFS‐ENS moves the storm too rapidly into the Norwegian Sea giving undue confidence for favourable conditions for the landing in Normandy.

Abstract. Reservoirs fundamentally alter downstream river flow regimes, decoupling discharge from natural meteorological forcing and challenging standard hydrological prediction. While data-driven models, such as Long Short-Term Memory (LSTM) networks, show promise in regulated catchments, it remains unclear how training data composition across natural and regulated rivers influences model generalisability and behaviour. In this study, we investigate how the presence or absence of reservoir-influenced catchments in training data impacts model performance across different flow regimes and alters the physical drivers the models learn to rely on. Using carefully matched subsets of the CAMELS-GB dataset, we trained separate specialist LSTMs (reservoir and non-reservoir), a pooled Full LSTM, and a multi-headed Hydra-LSTM to investigate whether explicit architectural specialisation offers any advantage over pooled training alone. Models were evaluated on held-out test gauges using standard performance metrics and gradient importance analysis to interpret feature reliance. Our results demonstrate that exposure to reservoir-influenced catchments during training is essential. Models trained exclusively on natural catchments consistently overestimate the mean and variance of regulated flows. Conversely, training exclusively on reservoir-influenced data degrades performance on non reservoir-influenced rivers (KGE reduction of ≥ 0.1) giving importance primarily to anthropogenic static features, such as abstraction rates, at the expense of precipitation drivers. A single Full LSTM trained on combined data matched the performance of both specialist models in their respective domains, implicitly switching its feature reliance between regimes. The Hydra-LSTM performed comparably to the Full LSTM throughout, indicating that the shared body may act as a regulariser limiting over-specialisation, but that explicit architectural specialisation provides no further benefit under these conditions. We conclude that pooling training data across regimes is a highly effective strategy for general-purpose modelling. However, case studies highlight a fundamental limitation: purely meteorological inputs remain insufficient for predicting flows in heavily managed single-purpose reservoirs, where unobserved human operational decisions dominate the hydrograph.

Abstract Accurately estimating rainfall distributions, from-small-to-extreme totals, is crucial for addressing various environmental challenges (e.g., flood forecasting, water resource management, disaster preparedness). Global Numerical Weather Prediction (NWP) models can provide useful rainfall estimates; yet, they often misrepresent point-scale observations from rain gauges, underestimating the frequency of small rainfall totals and extreme values. In general, finer resolutions yield more accurate representation of gauge-based climatologies. Hence, this study provides a systematic, global verification of four NWP-modelled rainfall datasets of differing resolutions (with “resolution” meaning “horizontal grid spacing”) - ERA5’s Ensemble Data Assimilation (62 km, probabilistic), ERA5’s short-range forecasts (31 km, deterministic), short-range 46r1 ECMWF reforecasts (18 km, control run), and ERA5-ecPoint (point-scale, probabilistic)—against 20 years of global rain gauge observations, assessing each dataset’s ability to represent the entire rainfall distribution. Although in very mountainous areas (e.g., the Andes) ERA5-ecPoint underestimates zero-rainfall frequency and overestimates wet tail length, it dramatically improves upon raw NWP performance in many other regions by capturing more accurately the frequency of zeros, the “growth rates” of rainfall totals, and the wet tails. Moreover, due to its probabilistic nature, ERA5-ecPoint can estimate long return periods (e.g., 1000 years) without using distribution fitting, thereby offering valuable insights into extremely rare or unprecedented events at specific locations. Such findings underscore the importance of using post-processing to enhance the local-scale validity of global NWP models. Moreover, as climate change intensifies extreme rainfall events, such post-processing becomes crucial for estimating accurate long-period rainfall climatologies, as needed for effective mitigation and resilience building, particularly in areas lacking comprehensive and reliable rain gauge records.

Machine learning models have been used with success to produce accurate river discharge forecasts at multiple lead times. However, almost no research has been done to show if they are physically consistent across lead times. In the deterministic problem setting, where models output a single forecast with multiple leadtimes, these models mean-seeking by construction, predicting the most likely river flow for each day, regardless of how likely the resulting trajectory is to occur. This is important for forecasters who need to look at the multi-day properties of a forecast, such as the accumulated flow or number of days over threshold. When each leadtime is described as an independent distribution, the model provides no insight into how to connect the uncertainties at each lead time, as an ensemble forecast would. In this paper, we show that temporal consistency in machine learning forecasts cannot be assumed, and develop two methods for enforcing temporal consistency, the Conditional-LSTM and Seeded-LSTM. Through this, we create ensemble forecasts that successfully predict temporal properties of the 10-day hydrographs. We find that by explicitly training the model to treat the prediction of previous lead times as truth, our model better predicts temporal properties of 10-day hydrographs than by not doing so. Our approach allows users to efficiently generate as many ensemble members as desired, and we use our results to highlight the important of developing temporally consistent ensembles.

The White Nile from Lake Victoria through Lakes Kyoga and Albert to the Sudd wetlands forms a complex connected lake-river-wetland system where flood propagation, storage, and attenuation remain poorly quantified. Following unprecedented and persistent flooding across South Sudan in 2022, this study quantified system-scale flood-wave transit time and examined how long it takes a flood wave to travel from Lake Victoria to the Sudd and how upstream storage and connectivity shape multi-year flood behaviour. Using daily lake levels, discharge, CHIRPS rainfall, and MODIS-derived inundation for 2002–2024, we tracked sequential flood peaks through the Victoria–Kyoga–Albert–Sudd cascade and mapped monthly wetland dynamics across five South Sudan sub-catchments. Flood-wave tracking showed a mean system transit time of approximately 17 months (16.84±1.95 months; range 13.0–20.9 months), substantially longer than the commonly inferred four-to-five-month timescale based on seasonal peak alignment. Segmental analysis revealed rapid transmission from Victoria to Kyoga (mean 4.2 months) but strong attenuation through the Albert–Sudd reach (mean 9.3 months), consistent with extensive floodplain storage and backwater control. Correlations between Lake Victoria peaks and downstream wetland extents strengthened markedly after 2019, with r2 exceeding 0.8 at 9–13-month lags, confirming strong hydraulic coupling and long system memory. These statistical lags complement, but do not represent, physical flood-wave transit times. The 2019–2024 high-water regime was not a series of isolated rainfall events but a multi-year propagation of excess storage initiated by the 2019 positive Indian Ocean Dipole anomaly and consecutive rainfall seasons. Lake Victoria reached exceptional peak levels of 1136.48 m a.s.l. in 2020, 1136.50 m a.s.l. in 2021, and 1136.66 m a.s.l. in 2024, each exceeding the historical 1964 maximum (1136.42 m a.s.l.). Over the same period, the Sudd Wetland exceeded its previous MODIS-era maximum extent (81 496 km2 in 2016) in every year from 2019 to 2024, reaching annual maxima of 120 680, 111 684, 111 480, 163 475, 122 292, and 116 359 km2, respectively. Flood-persistence mapping shows a shift from rainfall-driven activation in the eastern Sudd (Baro-Akobbo-Sobat–White Nile) in 2019–2020 to sustained, inflow-driven inundation across the central and western Sudd (Bahr el Jebel–Bahr el Ghazal–Bahr el Arab) during 2020–2022, consistent with water pathway activation and backwater expansion under high antecedent storage. When compared with historical episodes in the 1870s and 1960s, the persistence and spatial reach of the 2019–2024 floods rank among the most extensive in the modern record. These results redefine the White Nile as a long-memory system in which upstream storage governs downstream flood risk over multi-season timescales and challenge interpretations of flood propagation based on seasonal timing alone, offering a new empirical basis for flood forecasting, wetland management, and anticipatory action in South Sudan and across the wider basin.

Abstract Bankfull discharge, the maximum flow a river can convey before spilling over its banks, is central to modelling flood risk and understanding river-channel evolution. Global flood inundation models assume a 2-year return period for bankfull conditions, but this assumption remains untested globally. Here we use observations and machine learning to estimate bankfull discharge at ~1-km resolution along a recently developed global river network. We show that the widely used 2-year discharge reasonably approximates bankfull-flow magnitude, but bankfull return periods vary substantially within and across climate zones. Bankfull conditions occur more frequently in tropical and temperate regions (median return periods of 1.5 and 1.8 years; interquartile ranges of 2.5 and 3.2 years, respectively) and less frequently in cold and arid regions (2.8 and 4.3 years; interquartile ranges of 4.8 and 6.0 years). This variation across climate zones provides a basis for improving the representation of channel capacity in global flood models.

Over a billion people globally are already exposed to the risk of flooding, but by 2050 this number is expected to double due to human-induced climate change, population growth, and encroachment into at-risk areas. Global Flood Models (GFMs) are vital tools for producing flood hazard maps supporting impact estimates and policy interventions. These GFMs represent river channels by typically assuming that the bankfull flow-carrying capacity equates to a river flow with a specified return period (RP) that is spatially and temporally invariant. However, bankfull capacity is determined by channel size, shape and roughness and so varies in response to erosion and sedimentation. To quantify the extent to which channel variability biases GFM predictions here we employ a typical GFM, the Fathom model, to a 135,000 km2 region of the Mississippi floodplain in a sensitivity analysis that evaluates how inundated areas and associated population exposures respond when forced with empirically-derived bankfull capacities. Our results show that since the typical RPs (< 1 year) of these present-day bankfull flows differ from the 2-year value normally assumed in GFMs, substantial underestimates of flood extent (9 to 59%, depending on flood magnitude) and populations exposed (15 to 118%) result. We also show that, over multi-decadal timescales, changes in past channel morphology are, depending on emissions scenario, of equal or greater importance in driving changes in simulated flood hazard and risk than changes in future climate. The evolution of bankfull capacity through space and time is therefore a first order control on flood hazard and risk, meaning it is vital that river channel variability and change is represented accurately in GFMs.

Over 70% of flood events recorded in the past two decades in the Global Flood Database and WorldFloods dataset have occurred in locations where complex channel systems occur. Here we define complex channel systems as parts of the river network that diverge, such as bifurcations, multi-threaded channels, canals and deltas. Yet, large scale flood models have, until now, used only single-threaded networks due to the lack of a river network that reflects complex channel systems . Therefore, these large-scale models fundamentally misrepresent the physical processes in these often highly populated areas, leading to sub-optimal estimates of flood risk.Using the new Global River Topology (GRIT) dataset, a global bifurcation and multi-directional river network (Wortmann et al. 2023), we extend the river channel bathymetry estimation routine of Neal et al. (2021) to model multi-channels with LISFLOOD-FP. We compare the multi-thread model results to observations and to previous versions of LISFLOOD-FP using a single-threaded river network in the Indus, Mekong and Niger rivers at 1 arc second (~30m). By using GRIT, we find marked improvements in model results, observing better connectivity to areas of the floodplain that are far from the main channel and more channel floodplain interactions in wetlands. This work paves the way to further our understanding of global flood risk and to finally consider the diverse, evolving nature of geomorphologically active river networks. As this work progresses, we will continue to model a typology of bifurcations and multi-directional rivers to help further our understanding of the significance of complex river systems.Neal, J., Hawker, L., Savage, J., Durand, M., Bates, P., & Sampson, C. (2021). Estimating river channel bathymetry in large scale flood inundation models. Water Resources Research, 57(5), e2020WR028301.Wortmann, M., Slater, L., Hawker, L., Liu, Y., & Neal, J. (2023). Global River Topology (GRIT) (0.4) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7629908

We evaluate the skill and jumpiness of the ECMWF medium-range ensemble (ENS) in predicting tropical cyclone genesis in the Atlantic basin. Focusing on the probabilistic performance of the ENS, we assess how far in advance the ENS can predict genesis, quantify the consistency (jumpiness) from run to run, and investigate what factors influence the skill and consistency. We find that first indications of genesis are picked up at least 7 days ahead in 50% of the observed cases, although strong signals often only appear less than 3 days before genesis. There are significant regional differences, with observed genesis events predicted 2-3 days earlier in the eastern Atlantic than in other areas. The genesis probabilities can be jumpy from run to run, and the jumpiest cases are in the more skillful regions (central and eastern Atlantic) and for situations where the initial signal for genesis appears at longer lead time. In the eastern Atlantic, there is a tendency for the ENS tracks to reach tropical storm strength earlier and further east than observed; this model bias can affect both skill and jumpiness of the genesis forecasts. Our results provide guidance to forecasters on how to use and interpret the ENS predictions. Areas for future work include the link between early intensification in the eastern Atlantic and African easterly wave activity, the relationship between skill and the TC development pathways, and the impact of systematic analysis differences between 0000 UTC and 1200 UTC on forecast intensity. SIGNIFICANCE STATEMENT: Forecasting where and when tropical cyclones will appear increases the lead time at which decision-makers can begin to take preparatory mitigating action. Numerical weather prediction models can provide important guidance but sometimes are not consistent from one run to the next. We evaluate the skill and consistency of a state-of-the-art global model in predicting the formation of tropical cyclones up to 10 days ahead and provide guidance to forecasters on how to use and interpret the model predictions. We show that the formation of tropical cyclones can be predicted 2-3 days earlier in the eastern Atlantic than in the western Atlantic and identify some of the factors influencing both skill and consistency.

Meteotsunami (meteorological tsunami) are globally occurring progressive shallow water waves with a period of between 2 and 120 min which result from sudden pressure changes and wind stress due to moving atmospheric systems. These waves are known to cause destruction to and loss of assets. Currently, there is no research into the impact of meteotsunami on coastal ecosystems such as saltmarshes, despite the significant role saltmarsh play in providing vital habitats for resident and migrating birds, natural flood defences and climate mitigation. As such the restoration of saltmarshes has emerged as a pivotal focus within the UK Government's environmental policy framework. This paper examines the impact of two meteotsunami events (2016 and 2021) on saltmarsh vegetation in the southwestern UK. An assessment of the vegetation pre and post event was undertaken using high resolution satellite imagery and the Normalised Difference Vegetation Index (NDVI). Results revealed that the 2016 meteotsunami exacted minimal vegetation change with a decrease in NDVI from 0.26 to 0.23 and a temporary reduction in coverage of 40 %, suggesting a potential resilience to single episodic disturbances. In contrast, the 2021 event, compounded by multiple significant storms and additional meteotsunami, led to a decline in NDVI values from 0.44 to 0.22 and a temporary reduction in vegetation coverage of 66 %. Both events indicated a short-term disruption with a relatively rapid rebound (within one to three months). However, the longer-term effects of such a disruption on the saltmarsh ecosystem need to be investigated further. This comparative analysis underscores the complex interactions between meteotsunami, climatic phenomena, and coastal vegetation dynamics, highlighting the necessity for ongoing monitoring and research to understand the resilience mechanisms of such ecosystems in the face of increasing climatic variability and extreme weather events.

Existing global river networks underpin a wide range of hydrological applications but do not represent channels with divergent river flows (bifurcations, multi‐threaded channels, canals), as these features defy the convergent flow assumption that elevation‐derived networks (e.g., HydroSHEDS, MERIT Hydro) are based on. Yet, bifurcations are important features of the global river drainage system, especially on large floodplains and river deltas, and are also often found in densely populated regions. Here we developed the first raster and vector‐based Global RIver Topology that not only represents the tributaries of the global drainage network but also the distributaries, including multi‐threaded rivers, canals and deltas. We achieve this by merging a 30 m Landsat‐based river mask with elevation‐generated streams to ensure a homogeneous drainage density outside of the river mask for rivers narrower than approximately 30 m. Crucially, we employ the new 30 m digital terrain model, FABDEM, based on TanDEM‐X, which shows greater accuracy over the traditionally used SRTM derivatives. After vectorization and pruning, directionality is assigned by a series of elevation, flow angle and continuity approaches. The new global network and its attributes are validated using gauging stations, comparison with existing networks, and randomized manual checks. The new network represents 19.6 million km of streams and rivers with drainage areas greater than 50 km 2 and includes 67,495 bifurcations. With the advent of hyper‐resolution modeling and artificial intelligence, GRIT is expected to greatly improve the accuracy of many river‐based applications such as flood forecasting, water availability and quality simulations, or riverine habitat mapping.

Climate variability is a significant driver of flood events. However, geomorphological changes in river channels, including variations in local and upstream sediment supply, play a crucial role in determining flood conveyance capacity and flood stage variations. The interplay between hydrology and geomorphology, and their relative impact on flood conveyance, can vary in different river systems depending on both the degree of internal channel dynamics and the nature and magnitude of external forcings. For example, rates of bank erosion, vegetation establishment on bar surfaces, and overbank sedimentation control the time required for floodplain reworking, the adjustment of channel morphology and the associated evolution of river flow conveyance capacity and stage-discharge relations.To investigate the relative significance of hydrological and geomorphological controls on flood-stage variability, we employ a new computationally-efficient model of river and floodplain morphodynamics. This model simulates the evolution of river morphology and flow conveyance capacity by representing the interaction between processes of bank erosion, floodplain construction and river bed-level change over multiple centuries. The simple nature of the model enables its application at large spatial scales – e.g., to explore global variations in the controls on flood conveyance and its sensitivity to future environmental change. Simulated changes in conveyance capacity for a range of environmental settings were evaluated against trends in observed river gauging datasets. Convergent cross-mapping analysis was then applied to investigate the cause-and-effect relationships between controlling factors, including: (i) hydrologic regime; (ii) river sediment load; (iii) floodplain composition (e.g., fine versus coarse sediment); and (iv) lateral river dynamics (e.g., rates of erosion and accretion). Our analysis quantifies the causality between these factors and the resulting variability in river morphology (width and bed level), flood stage and channel conveyance capacity. Results indicate that in dynamic river systems, while the importance of climate-driven hydrological changes in driving conveyance capacity changes are acknowledged, geomorphological changes – specifically, variations in sediment supply and lateral sediment sources – may dominate over climate-driven trends.

The accurate estimation of bankfull discharge (QBF) plays a central role in multiple disciplines including geomorphology, hydrology, and ecology. For example, bankfull discharge is an essential input in many large-scale flood models which are widely used in understanding flood risk across large scales. However, in the context of extremely limited bankfull discharge observations, these Global Flood Models (GFMs) typically assume that bankfull discharge has a spatially uniform recurrence interval, with a value of 1-2 years widely adopted. In reality, many studies have found that the recurrence of bankfull discharge is highly variable. Therefore, more reliable estimates of bankfull discharge that account for river variability across different regions and climate zones are vital. Here, we train a random forest model to estimate bankfull discharge from global datasets encompassing river catchment characteristics, river geometry, topography, reservoir capacity, hydrological and climate indicators, alongside a newly compiled bankfull discharge database with over two thousand observations. The trained machine learning model is then used to develop the first estimate of bankfull discharge for 22 million km of rivers globally, using a newly developed, high-resolution, multi-threaded river network, Global River Topology (GRIT, Wortmann et al., 2023). Independent testing against observed values of QBF shows that the random forest model has good performance (R2=0.79), and the estimated QBF has better accuracy compared to the use of uniform recurrence-interval flows. This is the first study to estimate bankfull discharge for rivers at the global scale. Our dataset aims to improve bankfull representation in large-scale flood modelling, and to support river and water resources research more generally.Wortmann, M., Slater, L., Hawker, L., Liu, Y., & Neal, J. (2023). Global River Topology (GRIT) (0.4) [Data set]. Zenodo. 10.5281/zenodo.7629907