Study of Beijiang catchment flash-flood forecasting model

Flash floods occur frequently and distribute widely in mountainous areas because of complex geographic and geomorphic conditions and various climate types. Effective flash flood forecasting with useful lead times remains a challenge due to its high burstiness and short response time. Recently, machine learning has led to substantial changes across many areas of study. In hydrology, the advent of novel machine learning methods has started to encourage novel applications or substantially improve old ones. This study aims to establish a discharge forecasting model based on Long Short-Term Memory (LSTM) networks for flash flood forecasting in mountainous catchments. The proposed LSTM flood forecasting (LSTM-FF) model is composed of T multivariate single-step LSTM networks and takes spatial and temporal dynamics information of observed and forecast rainfall and early discharge as inputs. The case study in Anhe revealed that the proposed models can effectively predict flash floods, especially the qualified rates (the ratio of the number of qualified events to the total number of flood events) of large flood events are above 94.7% at 1–5 h lead time and range from 84.2% to 89.5% at 6–10 h lead-time. For the large flood simulation, the small flood events can help the LSTM-FF model to explore a better rainfall-runoff relationship. The impact analysis of weights in the LSTM network structures shows that the discharge input plays a more obvious role in the 1-h LSTM network and the effect decreases with the lead-time. Meanwhile, in the adjacent lead-time, the LSTM networks explored a similar relationship between input and output. The study provides a new approach for flash flood forecasting and the highly accurate forecast contributes to prepare for and mitigate disasters.

Floods are one of the most devastating natural hazards around the world, and flash floods are extremely dangerous to property and human lives. This study reviews the current flash flood forecasting in China and proposes a new methodology to develop a flash flood potential index (FFPI) based on GIS to further improve China's flash flood guidance system and national weather forecasting services. We collected datasets required to calculate the FFPI and tested a series of schemes to determine the method to calculate the FFPI over China. The result indicates that the calculated China's FFPI distribution can represent the risk of flash flood well, especially in Central and Southeast China. It is also implied that the southwest and western complex terrains are more susceptible to rain storms. Then we chose Yunnan Province as a test-bed and also yielded good results which can offer reference to the current China's flash flood forecasting system. Further research can be placed on incorporation of soil moisture and rainfall frequency components into the FFPI calculation method for dynamic warning and forecasting of flash floods.

<p>Radar rainfall nowcasting promises to result in more accurate and timely rainfall forecasts up to several hours in advance, compared to numerical weather prediction model forecasts. As flood early warning systems can benefit from this, we assessed the potential of radar rainfall nowcasting for (flash) flood forecasting with a large sample of 659 individual rainfall events for 12 catchments in the Netherlands. In this assessment, we tested four open-source nowcasting algorithms: Rainymotion Sparse (RM-S), Rainymotion DenseRotation (RM-DR), Pysteps deterministic (PS-D) and Pysteps probabilistic (PS-P) with 20 ensemble members. Eulerian Persistence (EP) forecasts and zero precipitation input (ZP) were the benchmark. The discharge forecasts had a 12-h forecast horizon and were issued for every 5-min step in the available nowcasts for the set of events. We regarded the simulations using the observed radar rainfall as reference.</p><p>We found that rainfall and discharge forecast errors increase with both increasing rainfall intensity and spatial variability. For the discharge forecasts, this relationship also depends on the initial conditions, as the forecast error increases more quickly with rainfall intensity, when the initial conditions are wet and groundwater tables are shallow. Overall, discharge forecasts using RM-DR, PS-D and PS-P outperform the other nowcasting methods.</p><p>Furthermore, we tested the potential for forecasting threshold exceedances by setting the highest discharge, per event, as threshold. Compared to benchmark ZP, a threshold exceedance is, on average, forecast 223 (EP), 196 (RM-S), 213 (RM-DR), 119 (PS-D) and 143 min (PS-P) earlier. The relatively high time profits with EP are counterbalanced by both a high false alarm ratio (FAR) and inconsistent forecasts. Contrarily, PS-D and PS-P show both lower FAR and inconsistency index values, which can lead to more trust in the simulations.</p><p>Based on this large-sample analysis, we conclude that all nowcasting methods have shown a benefit for short-term discharge forecasting compared to issuing no rainfall forecasts at all, though all have shortcomings. As forecast rainfall volumes are a crucial factor in (flash) flood forecasting, we recommend a future focus on improving this aspect in nowcasting.</p>

This article focuses on the forecasting of flash floods using the Algorithm of Storm Prediction as a new tool to predict convective precipitation, severe phenomena and the risk of flash floods. The first part of the article contains information on methods for predicting dangerous severe phenomena. This algorithm uses mainly data from numerical weather prediction models (NWP models), database of historic weather events and relief characteristics describing the influence of orography on the initiation of atmospheric convection. The result section includes verification of predicted algorithm outputs, selected NWP models and warnings of CHMI and ESTOFEX on three events related to the floods that hit the Zlín Region between years of 2015 - 2017. The main result is a report with prediction outputs of the algorithm visualized in maps for the territory of municipalities with extended competence and their regions. The outputs of the algorithm will be used primarily to increase the effectiveness of preventive measures against flash floods not only by the Fire Rescue Service of Czech Republic but also by the flood and crisis management authorities.

PREVIEW is an European Commission FP6 Integrated Project which aims at developing, on an European level, new geo-information services for natural and industrial risk management. The work package WP4340 focuses on forecast of Mediterranean flash floods. Phase 1 was devoted to the assessment of the usefulness of kilometric scale atmospheric model forecast for hydrological applications and to the development of hydro-meteorological coupled systems based on high-resolution atmospheric models and hydrological models able to reproduce the hydrological behaviour of Mediterranean catchments. Four high-resolution models at 2–3 km resolution have been run on five flash-flood cases over the French Cevennes-Vivarais and Italian Piedmont regions; models are MM5 (by NOA), COSMO-2 (by MeteoSwiss), MESO-NH (by Meteo-France) and COSMO-LAMI (by Arpa Piemonte). To investigate the benefit of coupling atmospheric and hydrological models, the quantitative precipitation forecasts (QPF) have been verified against observations using both classical and categorical statistical scores, while the sensitivity of the QPF to the model initial conditions has been also examined. In addition, the various hourly precipitation forecasts were supplied as input to hydrological models to evaluate through the simulated discharges the value of high resolution forecasts for hydrological forecast purposes. Clearly the hydrological verification conclusions differ from the QPF verification ones and show the usefulness of developing such hydrological verification as the one performed here.

Physically based, distributed hydrological model(PBDHM) was proposed for long time, and was regarded to have the potential to improve the flood forecasting accuracy. But unfortunately, this is still in the dream due to some existing challenges, and the biggest one is model parameter determination. Initially, it was assumed that parameter of PBDHM should be derived from the terrain properties directly, such as the DEM, land use/cover(LUC) types and soil types, not calibrated like lumped conceptual model(LCM) by employing optimization algorithm. In fact, PBDHM’s parameter calibration is also infeasible considering its huge number of model parameters, that could be up to millions or even to billions. As there is no “optimal” references for deriving PBDHM’s parameters directly from terrain properties, PBDHM’s capability for real-time flood forecasting has been weakened, so limiting its use mainly in scientific studies. In this study, the author assumes that PBDHM also needs parameter “calibration”, and the theory and framework for PBDHM parameter optimization have been presented. Based on the Liuxihe model, which was proposed for watershed flood forecasting, an automatic parameter optimization algorithm has been proposed by employing Particle Swarm Optimization (PSO). With parameter optimization, flood simulation accuracy of Liuxihe model has been improved largely, and very importantly, its performance is very stable. Not like LCM, model performance fluctuates sharply, thus limiting its capability being used for real-time flood forecasting. From dozens case studies in China, it also has been found that hydrological data from only one flood event is enough for parameter optimization, not like LCM, which requires hydrological data from a series of flood events. This finding is significant particularly for data-poor watershed, which makes PBDHM’s parameter optimization feasible for most of the world watersheds. With this advances, Liuxihe model has been used in several Chinses watersheds for real-time flood forecasting, and successful forecasting have been achieved. These successful implementations have proven that PBDHM has entered a new era for the real world application.

Flash floods account for a large proportion of flood-based fatalities, and they are becoming more frequent due to climate change. A global flash flood warning system therefore has the potential to be life saving.Standard approaches to flash flood forecasting – such as the Flash Flood Guidance (FFG) system in the US National Weather Service (NWS), or the European Runoff Index based on Climatology (ERIC) in the European Flood Awareness System (EFAS) – are utilizing recent weather and soil conditions, physiographic characteristics of basins, and weather forecasts, in order to produce a forecast for possible flash floods. These forecasts are not disseminated directly to the public. Instead, they are firstly refined by hydrologists that have intimate knowledge of the relevant basins and of previous flood events. This introduces a difficulty to scaling these methods worldwide, as the training of professional hydrologists in every region is costly and time consuming.Recent applications of Machine Learning (ML) to hydrology show that a learning system has the potential to train on data-rich basins and generalize to data-poor basins, with a skill that is comparable to state of the art hydrological models. In this work we attempt to build a ML model to produce daily flash flood forecasts, based on globally available weather reanalysis and physiographic characteristics from HydroATLAS. We discuss the model architecture, and evaluate it against NWS Flash Flood Warnings (FFW). While such models may not surpass the skill of a professional hydrologist, they have the potential to provide reasonable warnings in regions that do not currently have any such system in place.

Catchment response to precipitation is a complex, highly nonlinear, threshold-driven process influenced by various factors, including rainfall characteristics, topography, soil type, and antecedent soil moisture conditions. Identifying specific rainfall and soil moisture values beyond which runoff is generated—potentially leading to hazardous river discharges—is crucial in hydrology and fundamental for flash flood forecasting. However, this is not always straightforward, especially in data-scarce regions, due to ground data’s limited availability and representativeness. Satellite-based soil moisture retrievals provide a viable alternative for defining these thresholds, particularly in data-scarce catchments, though their application still needs to be explored. This study examines the ability of Soil Moisture Active Passive (SMAP) level 4 data to define surface runoff (SR) generation thresholds in six small catchments (i.e., smaller than 250 km2) with varied lithologies and land uses in the hilly and mountainous regions of central Italy—catchments representative of many in the Mediterranean. We analyze responses to 435 rainfall–runoff events, focusing on the antecedent soil moisture within the top 30 cm, estimated by SMAP level 4 (L4), up to 3 h before rainfall onset. Our findings show that the threshold-driven behavior of runoff generation can be effectively captured only by considering SMAP L4 antecedent soil moisture and rainfall depth together. In contrast, the use of rainfall or soil moisture alone is insufficient. This study demonstrates that SMAP L4 is a viable alternative to in situ observations for SR threshold identification, with practical applications for issuing flash flood warnings and alerts.

Modeling a densely urbanized watershed with an artificial neural network, weather radar and telemetric data

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.

[Departement_IRSTEA]Eaux [TR1_IRSTEA]ARCEAU

A comparative study of Machine Learning and Deep Learning methods for flood forecasting in the Far-North region, Cameroon

A simple, but rather complete hydro‐meteorological system for ensemble flash flood forecasting is described. Special focus is put on the use of precipitation forecasts, essentially quantitative and overall probabilistic (PQPFs). A first issue addressed concerns the matching of the time‐step used by the meteorological forecasts providers (12 or 24 h) with the appropriate hydrological time‐step for quick responding catchments (1 h or less). A second issue concerns the forecasts management between two updating cycles of the PQPFs. Finally, ensemble discharge forecasts for a case‐study event observed in Vogüé (South‐eastern France) are illustrated as an application of our flood forecasting system. Copyright © 2008 Royal Meteorological Society

Quantifying uncertainty associated with flash flood warning or forecast systems is required to enable informed decision making by those responsible for operation and management of natural hazard protection systems. The current system used by the U.S. National Weather Service (NWS) to issue flash-flood warnings and watches over the Unites States is a purely deterministic system. The authors propose a simple approach to augment the Flash Flood Guidance System (FFGS) with uncertainty propagation components. The authors briefly discuss the main components of the system, propose changes to improve it, and allow accounting for several sources of uncertainty. They illustrate their discussion with examples of uncertainty quantification procedures for several small basins of the Illinois River basin in Oklahoma. As the current FFGS is tightly coupled with two technologies, that is, threshold-runoff mapping and the Sacramento Soil Moisture Accounting Hydrologic Model, the authors discuss both as sources of uncertainty. To quantify and propagate those sources of uncertainty throughout the system, they develop a simple version of the Sacramento model and use Monte Carlo simulation to study several uncertainty scenarios. The results point out the significance of the stream characteristics such as top width and the hydraulic depth on the overall uncertainty of the Flash Flood Guidance System. They also show that the overall flash flood guidance uncertainty is higher under drier initial soil moisture conditions. The results presented herein, although limited, are a necessary first step toward the development of probabilistic operational flash flood guidance forecast-response systems.

While concept of boosting ensemble data mining techniques has been recently attracted a lot of attention for flood forecasting, mainly on non-urbanised river basins or reservoirs [1,2], time-series boosting, i.e., contribution of last timestep prediction to the next forecasting model is a new era, especially for real-time operation of flood forecasting models in the shape of early warning systems.This study aims to provide time-series boosting for ensemble flood forecasting model through adding forecasted water level of one timestep before as an input of training base models to previous proposed rainfall feature, especially rainfall duration, intensity, evidence of past rainfall and season occurrence [3]. Several weak learner data mining techniques are developed for various forecast lead times and recorded in data cube structure that can be used for developing time-series boosted ensemble model. This novel model was tested for real case study of Hanwell urban drainage systems located in the west London, UK for a period of 20 years data with 15min intervals. Confusion matrix is employed for performance assessment and the model is compared by conventional benchmark gradient boosted models.Results shows the added feature can significantly increase the accuracy of overflow detection of all developed base models, especially for longer timesteps. More specifically, adding the new feature to the model can increase the accuracy rate from 84% for the best developed base model to 93% in 3hrs-ahead predictions. More importantly, the model can decrease underestimation miss rate from 45% to only 21% for the same forecast lead time. Furthermore, new time-series boosted ensemble model can noticeably increase overflow detection rate, where hit rate increase from 78% to 88% in 3hrs-ahead predictions. Overall, the concept of time-series boosted ensemble modelling can overcome the problem of missing and false alarm of real-time operation by adding the previous situation of catchment to the forecasting procedure.