Copernicus Emergency Management Service Research Articles

Leveraging Sentinel-2 and Geographical Information Systems in Mapping Flooded Regions around the Sesia River, Piedmont, Italy

Sentinel-2 data are crucial in mapping flooded areas as they provide high spatial and spectral resolution but under cloud-free weather conditions. In the present study, we aimed to devise a method for mapping a flooded area using multispectral Sentinel-2 data from optical sensors and Geographical Information Systems (GISs). As a case study, we selected a site located in Northern Italy that was heavily affected by flooding events on 3 October 2020, when the Sesia River in the Piedmont region was hit by severe weather disturbance, heavy rainfall, and strong winds. The method developed for mapping the flooded area was a thresholding technique through spectral water indices. More specifically, the Normalized Difference Water Index (NDWI) and the Modified Normalized Difference Water Index (MNDWI) were chosen as they are among the most widely used methods with applications across various environments, including urban, agricultural, and natural landscapes. The corresponding flooded area product from the Copernicus Emergency Management Service (EMS) was used to evaluate the flooded area predicted by our method. The results showed that both indices captured the flooded area with a satisfactory level of detail. The NDWI demonstrated a slightly higher accuracy, where it also appeared to be more sensitive to the separation of water from soil and areas with vegetation cover. The study findings may be useful in disaster management linked to flooded-area mapping and area rehabilitation mapping following a flood event, and they can also valuably assist decision and policy making towards a more sustainable environment.

Global prediction of extreme floods in ungauged watersheds

Floods are one of the most common natural disasters, with a disproportionate impact in developing countries that often lack dense streamflow gauge networks1. Accurate and timely warnings are critical for mitigating flood risks2, but hydrological simulation models typically must be calibrated to long data records in each watershed. Here we show that artificial intelligence-based forecasting achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a five-day lead time that is similar to or better than the reliability of nowcasts (zero-day lead time) from a current state-of-the-art global modelling system (the Copernicus Emergency Management Service Global Flood Awareness System). In addition, we achieve accuracies over five-year return period events that are similar to or better than current accuracies over one-year return period events. This means that artificial intelligence can provide flood warnings earlier and over larger and more impactful events in ungauged basins. The model developed here was incorporated into an operational early warning system that produces publicly available (free and open) forecasts in real time in over 80 countries. This work highlights a need for increasing the availability of hydrological data to continue to improve global access to reliable flood warnings.

Mediterranean Wildfires’ Effect on Soil Quality and Properties: A Case from Northern Euboea, Greece

Mediterranean Wildfires’ Effect on Soil Quality and Properties: A Case from Northern Euboea, Greece

Global hydrological reanalyses: The value of river discharge information for world‐wide downstream applications – The example of the Global Flood Awareness System GloFAS

AbstractGlobal hydrological reanalyses are modelled datasets providing information on river discharge evolution everywhere in the world. With multi‐decadal daily timeseries, they provide long‐term context to identify extreme hydrological events such as floods and droughts. By covering the majority of the world's land masses, they can fill the many gaps in river discharge in‐situ observational data, especially in the global South. These gaps impede knowledge of both hydrological status and future evolution and hamper the development of reliable early warning systems for hydrological‐related disaster reduction. River discharge is a natural integrator of the water cycle over land. Global hydrological reanalysis datasets offer an understanding of its spatio‐temporal variability and are therefore critical for addressing the water–energy–food–environment nexus. This paper describes how global hydrological reanalyses can fill the lack of ground measurements by using earth system or hydrological models to provide river discharge time series. Following an inventory of alternative sources of river discharge datasets, reviewing their advantages and limitations, the paper introduces the Copernicus Emergency Management Service (CEMS) Global Flood Awareness System (GloFAS) modelling chain and its reanalysis dataset as an example of a global hydrological reanalysis dataset. It then reviews examples of downstream applications for global hydrological reanalyses, including monitoring of land water resources and ocean dynamics, understanding large‐scale hydrological extreme fluctuations, early warning systems, earth system model diagnostics and the calibration and training of models, with examples from three Copernicus Services (Emergency Management, Marine and Climate Change).

An Analysis of the Lagged Relationship between Anomalies of Precipitation and Soil Moisture and Its Potential Role in Agricultural Drought Early Warning

Abstract The standardized precipitation index (SPI) is the most commonly used index for detecting and characterizing meteorological droughts, and it is also extensively used as a proxy variable for soil moisture anomalies (SMA) for the purpose of monitoring agricultural drought in absence of long-term soil moisture observations. However, the potential capability of SPI to warn of the time-lagged soil water deficit—following the well-known “drought cascade” effect—is often overlooked in agricultural drought studies. In this research, a time-lagged correlation analysis is used to evaluate the relationship between the SMA dataset, generated as part of the Global Drought Observatory of the European Union’s Copernicus Emergency Management Service, and a set of SPIs derived from the ERA5 reanalysis produced by the European Centre for Medium-Range Weather Forecasts. The possibility to achieve an optimal agreement between SPI and SMA that also preserves the early warning skills of SPI is evaluated. The results suggest that if only the correlation between SPI and SMA is considered, the maximum agreement is usually obtained with a zero lead time (almost 80% of the cases), with SPI-3 representing the best option in about 40% of the grid cells at global scale. By also accounting for the benefits of a positive lead time, short accumulation periods tend to be favored, with SPI-1 being the optimal choice in about one-half of the cases, and 10–20 days of lead time in more than 90% of the grid cells is achieved without any significant reduction in either correlation or skill in drought extreme detection.

Global seasonal prediction of fire danger

The European Centre for Medium range weather forecast (ECMWF) on behalf of the Copernicus Emergency Management Service (CEMS) has recently widened the fire danger data offering in the Climate Data Store (CDS) to include a set of fire danger forecasts with lead times up to 7 months. The dataset incorporates fire danger indices for three different models developed in Canada, United States and Australia. The indices are calculated using ECMWF Seasonal Forecasting System 5 (SEAS5) and verified against the relevant reanalysis of fire danger based on the ECMWF Re-Analysis (ERA5). The data set is made openly available for the period 1981 to 2023 and will be updated regularly providing a resource to assess the predictability of fire weather at the seasonal time scale. The data set complements the availability of seasonal forecast provided by the Copernicus Emergency Management Service in real time.A preliminary analysis shows that globally anomalous conditions for fire weather can be predicted with confidence 1 month ahead. In some regions the prediction can extend to 2 months ahead. In most situations beyond this horizon, forecasts do not show more skill than climatology. However an extended predictability window, up to 6-7 months ahead is possible when anomalous fire weather is the results of large scale phenomena such as the El Niño Southern Oscillation and the Indian Ocean Dipole, often conducive of extensive fire burning in regions such as Indonesia and Australia.

Employing Copernicus Land Service and Sentinel-2 Satellite Mission Data to Assess the Spatial Dynamics and Distribution of the Extreme Forest Fires of 2023 in Greece

In 2023, Greece faced its worst wildfire season, with nine major fires causing unprecedented environmental damage of 1470.31 km2. This article uses Copernicus Land Monitoring Service and Sentinel-2 data, employing advanced remote sensing and GIS techniques to analyze spatial dynamics, map burn severity, assess fire extent, and highlight pre-fire tree density and land cover. The study focuses on the catastrophic fire in the Evros region and the damage to the National Forest Park of Dadia–Lefkimmi–Soufli. It also analyzes significant fires in Rhodes, Attica, Thessaly, Evia, Corfu, and Magnesia, emphasizing the compounded challenges posed by terrain, climate, and human factors in those areas. Additionally, the climate data for each affected area were compared with the weather conditions prevailing at the time of the fires. Copernicus Land Cover and Tree Density data are integrated to aid future management, assessment, and restoration. The analysis of maps and fire statistics underscores a notable pattern: areas with higher pre-fire tree density experienced correspondingly higher burn severity. This research underscores the crucial role of such data in assessing wildfire impact. In addition, compared with Copernicus Emergency Management Service, the burned area maps validate the accuracy and reliability of the utilized satellite data. The total burned area was assessed with a high accuracy rate of 96.28%.

Application of Gated Recurrent Unit Neural Network for Flood Extraction from Synthetic Aperture Radar Time Series

Floods are a sudden and influential natural disaster, and synthetic aperture radar (SAR) can image the Earth’s surface almost independently of time and weather conditions, making it particularly suitable for extracting flood ranges in time. Platforms such as Google Earth Engine (GEE) can provide a large amount of SAR data and preprocess it, providing powerful assistance for real-time flood monitoring and time series analysis. However, the application of long-term series data combined with recurrent neural networks (RNNs) to monitor floods has been lacking in current research, and the accuracy of flood extraction in open water surfaces remains unsatisfactory. In this study, we proposed a new method of near real-time flood monitoring with a higher accuracy. The method utilizes SAR image time series to establish a gated recurrent unit (GRU) neural network model. This model was used to predict normal flood-free surface conditions. Flood extraction was achieved by comparing and analyzing the actual flood surface conditions with the predicted conditions, using a parameter called Scores. Our method demonstrated significant improvements in accuracy compared to existing algorithms like the OTSU algorithm, Sentinel-1 Dual Polarized Water Index (SDWI) algorithm, and Z-score algorithm. The overall accuracy of our method was 99.20%, which outperformed the Copernicus Emergency Management Service (EMS) map. Importantly, our method exhibited high stability as it allowed for fluctuation within the normal range, enabling the extraction of the complete flood range, especially in open water surfaces. The stability of our method makes it suitable for the flood monitoring of future open-access SAR data, including data from future Sentinel-1 missions.

Sentinel-1-based analysis of the severe flood over Pakistan 2022

Abstract. In August and September 2022, Pakistan was hit by a severe flood, and millions of people were impacted. The Sentinel-1-based flood mapping algorithm developed by Technische Universität Wien (TU Wien) for the Copernicus Emergency Management Service (CEMS) global flood monitoring (GFM) component was used to document the propagation of the flood from 10 August to 23 September 2022. The results were evaluated using the flood maps from the CEMS rapid mapping component. Overall, the algorithm performs reasonably well with a critical success index of up to 80 %, while the detected differences can be primarily attributed to the time difference of the algorithm's results and the corresponding reference. Over the 6-week time span, an area of 30 492 km2 was observed to be flooded at least once, and the maximum extent was found to be present on 30 August. The study demonstrates the ability of the TU Wien flood mapping algorithm to fully automatically produce large-scale results and how key data of an event can be derived from these results.

Exploring MapSwipe as a Crowdsourcing Tool for (Rapid) Damage Assessment: The Case of the 2021 Haiti Earthquake

Abstract. Fast and reliable geographic information is vital in disaster management. In the late 2000s, crowdsourcing emerged as a powerful method to provide this information. Base mapping through crowdsourcing is already well-established in relief workflows. However, crowdsourced post-disaster damage assessment is researched but not yet institutionalized. Based on MapSwipe, an established mobile application for crowdsourced base mapping, a damage assessment approach was developed and tested for a case study after the 2021 Haiti earthquake. First, MapSwipe’s damage mapping results are assessed for quality by using a reference dataset in regard to different aggregation methods. Then, the MapSwipe data was compared to an already established rapid damage assessment method by the Copernicus Emergency Management Service (CEMS). Crowdsourced building damage mapping achieved a maximum F1-score of 0.63 in comparison to the reference data set. MapSwipe and CEMS data showed only slight agreement with Cohen’s Kappa values reaching a maximum of 0.16. The results highlight the potential of crowdsourcing damage assessment as well as the importance for a scientific evaluation of the quality of CEMS data. Next steps for further integrating the presented workflow into MapSwipe are discussed.

Object- Versus Pixel-Based Unsupervised Fire Burn Scar Mapping under Different Biogeographical Conditions in Europe

Wildfire detection and mapping is crucial for managing natural resources and preventing further environmental damage. In this study, we compared two methods of mapping burn scars using Sentinel-2 satellite imagery, a pixel-based approach and an object-based approach, at test sites located in various climatic zones with diverse land cover synthesis. The study aimed to determine the advantages and limitations of each method in terms of accuracy and precision in detecting burn scars. The results showed that both methods could detect burn scars with high accuracy, but with some limitations. The F1 score was in the range of 0.64–0.89 for the object-based approach, and 0.58–0.90 for the pixel-based approach. The pixel-based method produced a more precise delineation of the burnt area, but it was only suitable for detecting burn scars in a limited area of interest. The object-based method, on the other hand, was able to detect burn scars over a larger area accurately but with some commission errors. The results of both methods were also compared to the Copernicus Emergency Management Service (CEMS) rapid mapping product.

Integrating Copernicus Satellite Products and Ground-Truthing for Documenting and Monitoring the Impact of the 2022 Extreme Floods in Pakistan on Cultural Heritage

The catastrophic floods that hit Pakistan in summer 2022 represent the latest example of climate change-induced extreme events occurring in South Asia. In addition to the dramatic impact on population and infrastructures, this event threatened UNESCO World Heritage Sites (WHS) and properties of national interest. However, while a wealth of national and international mapping initiatives were conducted based on satellite imagery to assess damage to people, buildings and economic goods, the impact caused to archaeological sites and monuments has not been fully unveiled yet. To bridge this gap and provide an integrated approach that can be used by local end-users to assess damage and, in turn, collect evidence to inform and improve risk management plans, the present paper integrates Sentinel-1 and 2 imagery and mapping products derived from them (e.g., Copernicus Emergency Management Service Global Flood Mapping) with ground-truthing and geospatial datasets. Through a multidisciplinary collaboration between geologists, archaeologists, remote sensing and satellite image analysts, the integration methodology was tested on UNESCO and other heritage sites of national relevance located in the two mostly affected regions of Khyber Pakhtunkhwa and Sindh that were intentionally selected to represent different typologies of cultural heritage and governance in Pakistan. Finally, the information extracted from the present analysis was assessed in relation to the current national and international legislations, the official state of conservation reports and the activities conducted at each site to protect them against flood events. Given the accessibility to ready-to-use Copernicus products and that the present analysis can be replicated over time and other sites, the proposed methodology provides a feasible means to exploit satellite data in post-disaster mapping situations and contribute to the decision-making process for risk management.

Evaluating the potential of burn severity mapping and transferability of Copernicus EMS data using Sentinel-2 imagery and machine learning approaches

ABSTRACT The abiotic and biotic conditions in forest ecosystems can be significantly influenced by forest fires. However, difficulties in policy decisions for restoration inevitably occur in the absence of information on the damaged forests, such as location, area, and burn severity. In this study, eight spectral indices calculated from Sentinel 2 MSI imagery and machine learning algorithms (Random Forest (RF) and Support Vector Machine (SVM)) were used for mapping burned areas and severity. Two study sites with similar meteorological environment (dry season) and species (coniferous vegetation) were tested, and dataset (EMSR448) from Copernicus Emergency Management Service (CEMS) was used as the reference truth. RF showed better performance for classifying pixels from classes with similar properties than SVM. Normalized Burn Ratio (NBR) and Green Normalized Difference Vegetation Index (GNDVI) showed high importance in assessing fire severity suggesting that it may be effective for identifying senescent plants. The results also confirmed that the CEMS dataset has transferability as a reference truth for fire damage classification in other regions. Implementation of this method enables fast and accurate mapping of the area and severity of destructive damage by forest fires, and also has applicability for other disasters.

Integrating image segmentation in the delineation of burned areas on Sentinel-2 and Landsat 8 data

Precise and regularly updated maps of surface burned area extent are essential for fire management. It is challenging to obtain this information through ground surveys due to high cost. Satellite remote sensing has become a fundamental and cost-effective tool for acquiring such information in a temporal context. This paper features the integration of image segmentation in a methodological framework to delineate the extent of burned areas on Sentinel-2 and Landsat 8 optical data. A rule-based and a supervised classification approach are employed to support this aim. In order to evaluate the performance of the proposed approaches, the results were compared to the relevant Copernicus Emergency Management Service (EMS) maps. The results obtained using the rule-based approach depicted adequate agreement with the EMS product (86.9% using Landsat 8 data and 85.4% using Sentinel-2 data), whereas those obtained using a supervised classification approach depicted noticeable agreement with the EMS product (88.6% using Landsat 8 data and 90.7% using Sentinel-2 data). Finally, the widely used error matrix was calculated for each classification approach. We obtained 96% overall accuracy value using Landsat 8 data and 98% overall accuracy value using Sentinel-2 data. The results depict that the proposed methodology can be very useful for the delineation of burned areas.

Multi-Temporal Sentinel-1 SAR and Sentinel-2 MSI Data for Flood Mapping and Damage Assessment in Mozambique

Floods are one of the most frequent natural disasters worldwide. Although the vulnerability varies from region to region, all countries are susceptible to flooding. Mozambique was hit by several cyclones in the last few decades, and in 2019, after cyclones Idai and Kenneth, the country became the first one in southern Africa to be hit by two cyclones in the same raining season. Aiming to provide the local authorities with tools to yield better responses before and after any disaster event, and to mitigate the impact and support in decision making for sustainable development, it is fundamental to continue investigating reliable methods for disaster management. In this paper, we propose a fully automated method for flood mapping in near real-time utilizing multi-temporal Sentinel-1 Synthetic Aperture Radar (SAR) data acquired in the Beira municipality and Macomia district. The procedure exploits the processing capability of the Google Earth Engine (GEE) platform. We map flooded areas by finding the differences of images acquired before and after the flooding and then use Otsu’s thresholding method to automatically extract the flooded area from the difference image. To validate and compute the accuracy of the proposed technique, we compare our results with the Copernicus Emergency Management Service (Copernicus EMS) data available in the study areas. Furthermore, we investigated the use of a Sentinel-2 multi-spectral instrument (MSI) to produce a land cover (LC) map of the study area and estimate the percentage of flooded areas in each LC class. The results show that the combination of Sentinel-1 SAR and Sentinel-2 MSI data is reliable for near real-time flood mapping and damage assessment. We automatically mapped flooded areas with an overall accuracy of about 87–88% and kappa of 0.73–0.75 by directly comparing our prediction and Copernicus EMS maps. The LC classification is validated by randomly collecting over 600 points for each LC, and the overall accuracy is 90–95% with a kappa of 0.80–0.94.

An Unsupervised Burned Area Mapping Approach Using Sentinel-2 Images

The frequency and severity of large, destructive fires have increased in the recent past, with extended impacts on the landscape, the human population, and ecosystems. Earth observations provide a means for the frequent, wide coverage and accurate monitoring of fire impacts. This study describes an unsupervised approach for the mapping of burned areas from Sentinel-2 satellite imagery, which is based on multispectral thresholding, and introduces an adaptive thresholding method. It takes into account the localized variability of the spectral responses in a two-phase approach. The first phase detects areas that are burned with a high probability, while the second phase adaptively adjusts this preliminary mapping by expanding and refining its boundaries. The resulting classification contains two main classes of interest: burned and unburned. The latter is further classified into four (4) fire impact severity classes, according to the Copernicus Emergency Management Service (CEMS) and the NASA United States Geological Survey (USGS)’s widely acknowledged nomenclature examples. Three distinct wildfire events are assessed, which occurred during the summers of 2020 and 2021 in Greece and Portugal. The classification accuracy is calculated by juxtaposing the classification outputs to burned area validation maps created through the photointerpretation of very high-resolution (VHR) satellite imagery. The corresponding CEMS On-Demand Mapping products are also juxtaposed against the validation maps for comparison purposes. The accuracy assessment showcases that the unsupervised approach closely follows the capacity provided by the CEMS maps (e.g., the kappa coefficient—k—of the proposed unsupervised approach is 0.91, 0.83 and 0.83 for the events processed, while the CEMS products achieve a k of 0.94, 0.93 and 0.8, respectively). The proposed approach considers the variability of the affected areas’ spectral response; thus, it generalizes well to different areas, e.g., areas characterized by different land cover types. It seems to offer an effective means of mapping the wildfire-induced changes, which can be further incorporated and used by forest fire management services and further decision support systems complementary to the CEMS maps.

Regional Adaptability of Global and Regional Hydrological Forecast System

Our paper aims to improve flood forecasting by establishing whether a global hydrological forecast system could be used as an alternative to a regional system, or whether it could provide additional information. This paper was based on the operational Global Flood Awareness System (GloFAS) of the European Commission Copernicus Emergency Management Service, as well as on a regional hydrological forecast system named RHFS, which was created with observations recorded in the Wangjiaba river basin in China. We compared the discharge simulations of the two systems, and tested the influence of input. Then the discharge ensemble forecasts were evaluated for lead times of 1–7 d, and the impact on the forecasts of errors in initialization and modelling were considered. We also used quantile mapping (QM) to post-process the discharge simulations and forecasts. The results showed: (1) GloFAS (KGE of 0.54) had a worse discharge simulation than RHFS (KGE of 0.88), mainly because of the poor quality of the input; (2) the average forecast skill of GloFAS (CRPSS about 0.2) was inferior to that of RHFS (CRPSS about 0.6), because of the errors in the initialization and the model, however, GloFAS had a higher forecast quality than RHFS at high flow with longer lead times; (3) QM performed well at eliminating errors in input, the model, and the initialization.

BADI: A NOVEL BURNED AREA DETECTION INDEX FOR SENTINEL-2 IMAGERY USING GOOGLE EARTH ENGINE PLATFORM

Abstract. Forest fires are natural events that occur in numerous ecosystems worldwide and cause significant damage to human, ecological and socio-economic factors. It is also crucial to obtain useful information on the distribution and density of burned areas on large scale. An efficient way to map large regions is through remote sensing (RS). Nevertheless, the complex scenario and similar spectral signature of features in multispectral bands can lead to many false positives, making it difficult to extract the burned areas accurately. Multispectral data from Sentinel-2 satellite images allow the development of novel burned area indices, as more spectral data is recorded in the Red-Edge region. This research aims to develop a new burned area detection index (BADI) at 20 m spatial resolution in the google earth engine platform to detect the wildfire-affected areas in southwest of Iran using Sentinel-2 satellite imagery. The BADI spectral index has been specially designed to take benefit of the Sentinel-2 spectral bands and use a spectral combination of bands that are reasonable for post-fire burned regions detection. The final results indicated that the proposed index by applying a post-processing stage works well in the case of the study area to identify the burned areas. At the same time, it can satisfactorily suppress the complicated and irrelevant changes in the scene. Furthermore, the BADI index is rapid and can provide the burned areas map in near real-time. According to the Copernicus Emergency Management Service (EMS) reference data, maps of the burned areas were produced with a kappa coefficient of 0.92 and an overall accuracy of 92.15%, which demonstrated a good result in comparison to similar spectral indices.

Intercomparison of Automated Near-Real-Time Flood Mapping Algorithms Using Satellite Data and DEM-Based Methods: A Case Study of 2022 Madagascar Flood

Numerous algorithms have been developed to automate the process of delineating water surface maps for flood monitoring and mitigation purposes by using multiple sources such as satellite sensors and digital elevation model (DEM) data. To better understand the causes of inaccurate mapping information, we aim to demonstrate the advantages and limitations of these algorithms through a case study of the 2022 Madagascar flooding event. The HYDRAFloods toolbox was used to perform preprocessing, image correction, and automated flood water detection based on the state-of-the-art Edge Otsu, Bmax Otsu, and Fuzzy Otsu algorithms for the satellite images; the FwDET tool was deployed upon the cloud computing platform (Google Earth Engine) for rapid estimation of flood area/depth using the digital elevation model (DEM) data. Generated surface water maps from the respective techniques were evaluated qualitatively against each other and compared with a reference map produced by the European Union Copernicus Emergency Management Service (CEMS). The DEM-based maps show generally overestimated flood extents. The satellite-based maps show that Edge Otsu and Bmax Otsu methods are more likely to generate consistent results than those from Fuzzy Otsu. While the synthetic-aperture radar (SAR) data are typically favorable over the optical image under undesired weather conditions, maps generated based on SAR data tend to underestimate the flood extent as compared with reference maps. This study also suggests the newly launched Landsat-9 serves as an essential supplement to the rapid delineation of flood extents.

Using a long short-term memory (LSTM) neural network to boost river streamflow forecasts over the western United States

Abstract. Accurate river streamflow forecasts are a vital tool in the fields of water security, flood preparation and agriculture, as well as in industry more generally. Traditional physics-based models used to produce streamflow forecasts have become increasingly sophisticated, with forecasts improving accordingly. However, the development of such models is often bound by two soft limits: empiricism – many physical relationships are represented empirical formulae; and data sparsity – long time series of observational data are often required for the calibration of these models. Artificial neural networks have previously been shown to be highly effective at simulating non-linear systems where knowledge of the underlying physical relationships is incomplete. However, they also suffer from issues related to data sparsity. Recently, hybrid forecasting systems, which combine the traditional physics-based approach with statistical forecasting techniques, have been investigated for use in hydrological applications. In this study, we test the efficacy of a type of neural network, the long short-term memory (LSTM), at predicting streamflow at 10 river gauge stations across various climatic regions of the western United States. The LSTM is trained on the catchment-mean meteorological and hydrological variables from the ERA5 and Global Flood Awareness System (GloFAS)–ERA5 reanalyses as well as historical streamflow observations. The performance of these hybrid forecasts is evaluated and compared with the performance of both raw and bias-corrected output from the Copernicus Emergency Management Service (CEMS) physics-based GloFAS. Two periods are considered, a testing phase (June 2019 to June 2020), during which the models were fed with ERA5 data to investigate how well they simulated streamflow at the 10 stations, and an operational phase (September 2020 to October 2021), during which the models were fed forecast variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS), to investigate how well they could predict streamflow at lead times of up to 10 d. Implications and potential improvements to this work are discussed. In summary, this is the first time an LSTM has been used in a hybrid system to create a medium-range streamflow forecast, and in beating established physics-based models, shows promise for the future of neural networks in hydrological forecasting.