Integrating machine learning ensembles and flood classification for enhanced flood forecasting with dynamic parameter weighting

[Departement_IRSTEA]Eaux [TR1_IRSTEA]ARCEAU

Flood forecasting is becoming increasingly important across the world. The exposure of people and property to flooding is increasing and society is demanding improved management of flood risk. At the same time, technological and data advances are enabling improvements in forecasting capabilities. One area where flood forecasting is seeing technical developments is in the use of probabilistic forecasts—these provide a range of possible forecast outcomes that indicate the probability or chance of a flood occurring. While probabilistic forecasts have some distinct benefits, they pose an additional decision-making challenge to those that use them: with a range of forecasts to pick from, which one is right? (or rather, which one(s) can enable me to make the correct decision?). This paper describes an innovative and transferable approach for aiding decision-making with probabilistic forecasts. The proposed risk-based decision-support framework has been tested in a range of flood risk environments: from coastal surge to fluvial catchments to urban storm water scales. The outputs have been designed to be practical and proportionate to the level of flood risk at any location and to be easy to apply in an operational flood forecasting and warning context. The benefits of employing a benefit-cost inspired decision-support framework are that flood forecasting decision-making can be undertaken objectively, with confidence and an understanding of uncertainty, and can save unnecessary effort on flood incident actions. The method described is flexible such that it can be used for a wide range of flood environments with multiple flood incident management actions. It uses a risk-based approach taking into account both the probability and the level of impact of a flood event. A key feature of the framework is that it is based on a full assessment of the flood-related risk, taking into account both the probability and the level of impact of a flood event. A recommendation for action may be triggered by either a higher probability of a lower impact flood or a low probability of a very severe flood. Hence, it is highly innovative as it is the first application of such a risk-based method for flood forecasting and warning purposes. A final benefit is that it is considered to be transferrable to other countries.

An approach for improving the capability of a coupled meteorological and hydrological model for rainfall and flood forecasts

Ensemble predictions and perceptions of risk, uncertainty, and error in flood forecasting

Flood forecasting has proved to be an effective and helpful non-structure method which can effectively reduce the impact of floods by providing early warnings ahead of time. Among many flood forecasting models, Rainfall-Runoff Correlation Diagram (RRCD) gives more accurate prediction by making use of large amount of history data and acting as experience expect, and thus is widely applied in operational flood forecasting. This paper proposed a new method to draw the RRCD through an equation instead of subjectivity. The method is then applied to Zhang River Reservoir with 126 flood events from historical meteorological and hydrological data from 1971 to 2000. Results show that this method gives performance as good as experience in more objective aspect and can be easily applied to consider the influence of rainfall hours cast on RRCD.

The paper proposes a methodological approach and software for managing information processes in systems for modeling natural objects on the example of a system for operational forecasting of fl oods. The proposed approach is based on application of cyber-physical systems, the industrial Internet and fog computing concepts. The results of testing the technologies for managing information processes demonstrate increasing both reliability of modeling and effi ciency of using the fl ood forecasting system computing resources.

ABSTRACTThe aim of this study is to assess the influence of sensor locations and varying observation accuracy on the assimilation of distributed streamflow observations, also taking into account different structures of semi-distributed hydrological models. An ensemble Kalman filter is used to update a semi-distributed hydrological model as a response to measured streamflow. Various scenarios of sensor locations and observation accuracy are introduced. The methodology is tested on the Brue basin during five flood events. The results of this work demonstrate that the assimilation of streamflow observations at interior points of the basin can improve the hydrological models according to the particular location of the sensors and hydrological model structure. It is also found that appropriate definition of the observation accuracy can affect model performance and consequent flood forecasting. These findings can be used as criteria to develop methods for streamflow monitoring network design.

Stress is an almost basic human instinct, especially in the internet age. People’s health is being endangered by psychological stress. It is critical to diagnose stress early on to provide proactive therapy. In the age of the Internet, social media has an enormous impact on human thinking. It can cause mental health issues such as stress but can also play an important part in detecting it. Advancements in machine learning and natural language processing have enabled information extraction from a huge amount of raw textual data. In this study machine learning approach is used to detect stress in a social media text. Data used in this study is from Reddit. Classical and Ensemble machine learning approaches were used for stress detection. Classical Techniques include Decision Tree, Logistic Regression, support vector machine, Random Forest, and Na ̈ıve Bayes. Ensemble approaches used are boosting, bagging, and voting. The ensemble approach was able to provide better results than all machine learning baselines on the dataset. Also, it was able to outperform all non-transformed-based neural network architectures discussed in the baseline. The best-performing model in this study is Logistic regression with a 76.6% F1 Score.

Problems and prospects in long and medium range weather forecasting

Optimized machine learning approaches for identifying vertical temperature gradient on ballastless track in natural environments

Great advancement has been achieved in the last decade in 2D shallow water solvers for flood modelling. However, their application to physically-based flood forecasting continues to be experimental and not widespread. One of the central challenges towards operational flood forecasting with 2D solvers is their computational cost, which needs to be reconciled with the required lead times for forecasts to be of use. Nonetheless, these solvers have great potential to improve flood forecasting predictions, especially when it comes to flash floods, for which the established 1D and conceptual models may be significantly less applicable.The shallow water solver SERGHEI-SWE leverages on robust and efficient numerical techniques and is implemented for High Performance Computing (HPC), allowing its use in supercomputers and opening new opportunities in 2D flood forecasting. In this contribution, we present proof-of-concept simulations of several flood events in different catchment and river systems. We show that, with SERGHEI-SWE, it is possible to run very high resolution flood simulations for large hydrological systems with runtimes significantly lower than the event duration. This property is essential to enable operational forecasting with useful lead times.We run simulations on three river reaches, in the Italian river Po (125 km reach between Boretto and Pontelagoscuro) and in one of its tributaries, the river Secchia (20 km reach), and a meandering reach of the Ebro river through the city of Zaragoza. We also perform flash flood simulations on a 5 km2 district of Nice (France), and in a 50 km2 agricultural catchment in Jaén (Spain). The focus of the exercise is on the computational performance aspect and not on the model performance. The results show that high resolution simulations can be done with runtimes in the order of 100 times faster than real time, potentially allowing a very good forecast lead time. We also explore different combinations of computational resources, model resolution and ensemble size to explore the flexibility of the modelling approach under different computational systems, which may be available for flood forecasting.

Classification and prediction tasks are common in health research. With the increasing availability of vast health data repositories (e.g. electronic medical record databases) and advances in computing power, traditional statistical approaches are being augmented or replaced with machine learning (ML) approaches to classify and predict health outcomes. ML describes the automated process of identifying ("learning") patterns in data to perform tasks. Developing an ML model includes selecting between many ML models (e.g. decision trees, support vector machines, neural networks); model specifications such as hyperparameter tuning; and evaluation of model performance. This process is conducted repeatedly to find the model and corresponding specifications that optimize some measure of model performance. ML models can make more accurate classifications and predictions than their statistical counterparts and confer greater flexibility when modelling unstructured data or interactions between covariates; however, many ML models require larger sample sizes to achieve good classification or predictive performance and have been criticized as "black box" for their poor transparency and interpretability. ML holds potential in family medicine for risk profiling of patients' disease risk and clinical decision support to present additional information at times of uncertainty or high demand. In the future, ML approaches are positioned to become commonplace in family medicine. As such, it is important to understand the objectives that can be addressed using ML approaches and the associated techniques and limitations. This article provides a brief introduction into the use of ML approaches for classification and prediction tasks in family medicine.

Performance optimization of wireless networks is typically complicated because of high computational complexity and dynamic channel conditions. Considering a specific case, the recent introduction of intelligent reflecting surface (IRS) can reshape the wireless channels by controlling the scattering elements' phase shifts, namely, passive beamforming. However, due to the large size of scattering elements, the IRS's beamforming optimization becomes intractable. In this article, we focus on machine learning (ML) approaches for complex optimization problems in wireless networks. ML approaches can provide flexibility and robustness against uncertain and dynamic systems. However, practical challenges still remain due to slow convergence in offline training or online learning. This motivated us to design a novel optimization-driven ML framework that exploits the efficiency of model-based optimization and the robustness of model-free ML approaches. Splitting the control variables into two parts allows one part to be updated by the outer loop ML approach while the other part is solved by the inner loop optimization. The case study in IRS-assisted wireless networks confirms that the optimization-driven ML framework can improve learning efficiency and the reward performance significantly compared to conventional model-free ML approaches.

The Red River of the North basin (RRB) is vulnerable to spring snowmelt flooding because of its flat terrain, low permeability soils, and the presence of river ice jams resulting from the river's northward flow direction. The onset and magnitude of major flood events in the RRB have been very difficult to forecast, in part due to limited field observations of snow water equivalent (SWE). Coarse-resolution (25-km) passive microwave observations from satellite instruments are well suited for the monitoring of SWE. Despite routine use in the Earth sciences community to document the impacts of climate change, the use of passive microwave observations in operational flood forecasting is rare. This paper compares daily satellite passive microwave SWE observations from special sensor microwave/imager (SSM/I) and special sensor microwave imager/sounder (SSMIS), advanced microwave scanning radiometer for earth observing system (AMSR-E), and advanced microwave scanning radiometer 2 (AMSR2) from 2003 to 2016 to modeled output from the SNOw Data Assimilation System (SNODAS) and Global Snow Monitoring for Climate Research -2 (GlobSnow-2) in the RRB to determine the differences between the remotely sensed SWE estimates and the model products currently used in flood forecasting. Results show statistically significant differences between the satellite observations and SNODAS in the northern region of the basin that were not evident in the southern region. Satellite estimates of peak SWE in the forecast subbasins in the northern region were 40–125% higher than the model results which points to the lack of ground observations used to constrain the model simulations. This paper recommends that satellite SWE observations should be considered for improving operational snowmelt forecasting in the RRB.

TIGGE (THORPEX International Grand Global Ensemble) was a major part of the THORPEX (Observing System Research and Predictability Experiment). It integrates ensemble precipitation products from all the major forecast centers in the world and provides systematic evaluation on the multimodel ensemble prediction system. Development of meteorologic-hydrologic coupled flood forecasting model and early warning model based on the TIGGE precipitation ensemble forecast can provide flood probability forecast, extend the lead time of the flood forecast, and gain more time for decision-makers to make the right decision. In this study, precipitation ensemble forecast products from ECMWF, NCEP, and CMA are used to drive distributed hydrologic model TOPX. We focus on Yi River catchment and aim to build a flood forecast and early warning system. The results show that the meteorologic-hydrologic coupled model can satisfactorily predict the flow-process of four flood events. The predicted occurrence time of peak discharges is close to the observations. However, the magnitude of the peak discharges is significantly different due to various performances of the ensemble prediction systems. The coupled forecasting model can accurately predict occurrence of the peak time and the corresponding risk probability of peak discharge based on the probability distribution of peak time and flood warning, which can provide users a strong theoretical foundation and valuable information as a promising new approach.