|Wednesday, September 16th|
Sadegh Sadeghi Tabas, P.hD. student of Water Res. En
1:20 PM - 1:40 PM
Surface flow models provide a comprehensive assessment of rainfall-runoff processes to understand how a catchment system responds to different environmental conditions. Although, many modelling systems including conceptual and distributed physical hydrological approaches provide a concise expression of flow path heterogeneity and limited system response function at the catchment scale. Recently, deep leaning methods have been proposed for empirical rainfall-runoff modelling and to complement existing hydrological models (both distributed physical and conceptual approaches), particularly in a catchment where data to support process-based model is limited. This study implemented a new and innovative Deep Learning (DL) modelling system to simulate daily streamflow timeseries across a mix urban and rural coastal catchment in North Carolina, USA. We used a new generation of deep learning neural network (DLNN), i.e., Recurrent Neural Networks (RNNs), that seamlessly transformed data into intelligence and simulated sequential flow rates based on a set of collected flow factors. Furthermore, the architecture of RNNs provided a neural basis for efficient training procedure that had the ability to intelligently integrate any mathematical or logical algorithm into the simulation decision process. Analysis suggests that the effects of input data characteristics on model performance (sequential data), the uncertainty associated with forcing data, the amount of training data, and the correlation among different attributes of data series are important factors for a neural computational development underlying surface flow processes. RNN algorithm was able to learn the complexity of order or temporal dependence between observations, thereby it was capable of accurately model the complex multivariate sequences of a coastal rainfall-runoff process. Further investigation of RNN simulation revealed that while model architecture was important, training on a large amount of dataset was necessary to enforce spatio-temporal hierarchical relationships of data with the catchment attributes. Together our results provide an algorithmically informed simulation on the dynamics of daily streamflow simulation and maybe applicable to other complex catchment and climate settings.
Brianna Pagán, VITO, VITO
1:40 PM - 2:00 PM
The Internet of Water (IoW) is a large-scale permanent IoT sensor network with 2500 water quality sensors spread across Flanders, Belgium. This intelligent system will permanently monitor water quality and quantity in real-time. Such a dense network of sensors with high temporal resolution will provide unprecedented volumes of data for drought, flood and pollution management, prediction and decisions. Here we present several data mining and machine learning initiatives along with a database infrastructure which supports environmental modelling efforts and large scale monitoring networks like IoW. Examples include interpolating grab sample measurements to river stretches to monitor salinity intrusion. A shallow feed forward neural network is trained on historical grab samples using physical characteristics of the river stretches. Such a system allows for salinity monitoring without complex convection-diffusion modeling, and for estimating salinity in areas with less monitoring stations. Another highlighted project is the coupling of neural network and data assimilation schemes for water quality forecasting. A long short-term memory recurrent neural network is trained on historical water quality parameters and remotely sensed spatially distributed weather data. Using forecasted weather data, a model estimate of water quality parameters are obtained from the neural network. A Newtonian nudging data assimilation scheme further corrects the forecast leveraging previous day observations, which can aid in the correction for non-point or non-weather driven pollution influences. Calculations are supported by an optimized database system developed by the Hasselt University which further exploits data mining techniques to estimate water movement and timing through the Flanders river network system. As geospatial data increases exponentially in temporal and spatial resolutions, scientists and water managers must consider the tradeoff between computational resources and physical model accuracy. These type of hybrid approaches allows for near real-time analysis without computational limitations and will further support research to make communities more climate resilient
Jonathan Goodall, University of Virginia
2:00 PM - 2:20 PM
Technical approaches for enabling more open and reproducible computational models are gaining attention in the environmental modelling and software community. We see three main themes emerging from this research: (1) advancing data sharing platforms, (2) using containers and notebooks for encapsulating complete computational environments and analyses, and (3) developing higher-level Application Programming Interfaces (APIs) for simulation models to make them more scriptable and notebook-friendly. In this research, we explore an approach for leveraging these topics into a model agnostic framework able to support open and reproducible environmental modeling. The framework’s design consists of data sharing achieved through online repositories, notebook-based and containerized modeling analyses in the cloud, and model APIs allowing for the abstraction of lower-level details for model configurations, execution, and visualization. We present an example implementation of the approach using HydroShare as an online repository, CUAHSI and CyberGIS JupyterHubs as computational environments, and pySUMMA as an example model API for the Structure for Unifying Multiple Modeling Alternatives (SUMMA) hydrologic model. We demonstrate how the approach can be leveraged for a study by (1) creating and organizing HydroShare resources for the study’s data and model files and (2) using Jupyter notebooks and pySUMMA to reproduce figures from the past study. We will discuss within this context more nuanced views of reproducibility and remaining challenges to achieving computational reproducibility in environmental modelling not addressed through this research.
Vidya Samadi, Clemson University, United States
2:20 PM - 2:40 PM
Successive flood events have brought new challenges to human life, civil infrastructure system and the environment in the southeast United States. To address flooding impacts, this study developed artificial intelligence algorithms (image processing approaches) to detect floodwater extent on inundated roadways from image data captured by smartphones, traffic cameras, etc. A sample dataset collected in real-time from recent flooding in South Carolina, USA and location-matched reference images are used to compute flood depth and inundation extend. Different algorithms have been used including Cloud Vision API in the Google Cloud Console and Convolutional Neural Networks (CNNs). To create high-quality training datasets of annotated images, at least 100-200 images are required to train the algorithm and label the objects. For the sake of comparing the image quality of the compressed image with the original image, we calculated and compared the structural similarity (SSIM) index of the two images in addition to mean square error (MSE) or peak signal-to-noise ratio (PSNR). Analysis suggests that differences in image resolution and lighting, and environmental conditions have significant impact on annotating an image with a label and score. Further, crowdsourced images showed discrepancies in labeling dry/flooded image pairs specifically for differentiating “flood”, “floodplain”, “waterway”, and “road”. Specifically, all models showed high precision in detecting critical infrastructures (>70% precision) in the images such as road, bridges, and reservoirs while revealed significant challenges in detecting flood and water (50%-55% precision). More research is underway to integrate image processing algorithm with the watershed geometry for flood severity and inundation assessment and to apply the results for emergency response purposes in real time.
Benjamin Bowes, University of Virginia
2:40 PM - 3:00 PM
The traditional gravity-driven stormwater systems that coastal urban communities typically rely on to manage flooding are increasingly stressed by sea level rise and climate change. While recent research has shown that retrofitting these passive systems as smart cyber-physical systems controlled in real-time can improve their performance, methods for automating and optimizing that real-time control is an active area of study. This research explores deep reinforcement learning (RL) to create control policies for these systems. In RL, an agent learns control policies by interacting with an environment and receiving rewards or penalties based on the agent's actions. In this case, an RL agent controls valves in a simulated stormwater system, which uses observed rainfall and tide data as input, and is penalized if flooding occurs. The RL agent must learn to manage the depth of water in two retention ponds by opening and closing valves, which is complicated by the fact that the agent must learn to time releases of water from the ponds with changing tidal conditions at the stormwater system outlet. The RL agent’s performance is compared to (i) a passive, gravity-driven system and (ii) a model predictive control (MPC) strategy. Results show that the RL agent can learn to proactively manage water levels in the retention ponds based on current and forecast conditions, and can reduce flooding compared to the passive, gravity-driven system. In contrast to MPC, which in this case uses a heuristic approach to perform online optimization, the RL agent is better suited for real-time control because it can be trained offline and used with a lower real-time computational cost than MPC, allowing it to scale to larger systems. This research helps to inform control strategies for smart stormwater systems by allowing them to learn from and adapt to a wide range of conditions.
Khouloud Gader, University of Carthage, National Agronomic Institute of Tunisia (INAT), Tunisia
3:20 PM - 3:40 PM
Tackling the threats of climate change has become a global challenge. Rainfall in the Mediterranean region is characterized by a large spatiotemporal variability. In this research study, an Empirical Statistical Downscaling (ESD) approach was used to generate regression-based climate projections for the horizon 2100 at the level of six main climate stations in the Medjerda catchment in Tunisia. The MIROC5 model outputs for historical and future conditions were used, together with the reanalysis NCEP data and RCP emissions data (RCP4.5 and RCP 8.5). The downscaled values were based on correlations between large scale atmosphere, ocean variables and precipitation at the station level. Historical monthly rainfall data at six gauges during 30 years, from 1981 until 2010, archived given by the national climate data center NCDC were used. Data gaps were filled with data provided by the National Institute of Meteorology (INM). To choose the appropriate regression model, six different types of statistical models were tested: generalized linear models, selected GLM, generalized additive models, selected GAM, random forest models, and artificial neural networks based. It is concluded that the best prediction skill varies between stations. Further, it is shown that precipitation in the Medjerda catchment, for the horizon 2100, will undergo a decline of between 9% and 35% depending on the RCP 4.5 scenario. This decline will be more pronounced for the RCP 8.5 scenario.
Jacob Kravits, University of Colorado, University of Colorado Boulder, United States
3:40 PM - 4:00 PM
Machine learning methods often require tuning hyperparameters to optimize their performance. Often these hyperparameters are selected based on prior assumptions or a single-objective optimization. However, these techniques fail to capture tradeoffs between type I (i.e., false positive) and type II (i.e., false negative) misclassifications. This presentation advances the process of hyperparameter optimization by analysing tradeoffs among multiple classification objectives. Our approach starts with feature selection, coupled with a machine learning classification model. We employ the BORG multiobjective evolutionary algorithm to explore different values of hyperparameters and identify tradeoffs among objectives describing misclassifications, including area under the receiver operating characteristic curve, precision, and accuracy. Such an approach is broadly applicable to environmental applications where type I and type II errors have differing consequences, empowering analysts to make informed choices of hyperparameter values when applying machine learning algorithms to real-world situations. The approach is demonstrated on the novel classification problem of dams deemed to have a high or not-high hazard potential (HP). A machine learning algorithm “learns” to classify existing dam hazard classifications based on features such as dam height, length, reservoir size, and downstream population. This is a problem where type I and type II errors could have dire implications because a dam with a high HP means that failure or misoperation would cause probable loss of human life. In this research, we develop a data-driven dam HP classification model, demonstrating its feasibility with National Inventory of Dams entries in the northeastern United States.
Stefano Bagli, GECOsistema Srl, italy
4:00 PM - 4:20 PM
Integrating big open data and emerging information and communication technologies in flood risk management may well represent a next crucial stage in supporting robust, evidence-based decision-making. Flood risk management policies, such as the EU Floods Directive, are increasingly demanding and challenging. Authorities and stakeholders are clearly in need of innovative and cost-effective solutions to assess and map flood risk. SmartFLOOD is an innovative Artificial Intelligence Cloud-Web platform developed to drastically reduce cost and processing time in mapping Fluvial Flood Hazard at large-scales and at high resolution anywhere in the world. The web platform generates in real-time comprehensive flood hazard map for every place in Europe (at 25 m resolution) and for the entire planet (at 90 m resolution). SmartFLOOD’s intelligence is composed by three main components: 1) A high resolution Geomorphic Flood Index – GFI generated by processing available high resolution and hydrologically corrected Digital Elevation Models. 2) A set of available benchmark flood hazard maps usually available at lower resolution or limited spatial extension. 3) An efficient Machine Learning binary classification algorithm able to quickly provide the calibrated flood prone area. The Web Service is designed to provides end-users with the possibility to consistently and cost-effectively extend local detailed flood studies (i.e. those coming from complex flood models, field studies or remote sensing detection) in uncovered data-scarce regions, to extrapolate them to larger scales and downscale to finer resolutions. The service is fully operative at global scale and is available at https://gecosistema.com/smartflood.
Faria Zahura, University of Virginia, University of Virginia, United States
4:20 PM - 4:40 PM
Coastal communities have been threatened by increased frequency and severity of flooding in recent years, which is likely to worsen due to climate change. Flood prediction models are essential tools to help reduce the socio-economic disruptions caused by escalating flood events. However, traditional flood prediction approaches require powerful computing machines and have long run-times because they employ 1D/2D dual drainage models. Therefore, these models are unsuitable for real-time street flood forecasting. Since machine learning models can have smaller computational costs and fast run-times, they may have the potential to be used as an alternative approach. This study demonstrates the use of a machine learning algorithm, Random Forest, as a surrogate model to emulate flood depth on streets from a high-fidelity 1D pipe/2D overland flow physics-based model, TUFLOW. The Random Forest surrogate model was trained and evaluated using the top 20 storm events from 2016 to 2018 in the coastal city of Norfolk, VA, USA. The surrogate model uses topographic (e.g. elevation, topographic wetness index, depth to water) and environmental (e.g. rainfall, tide) features for 17,000 road segments in the city as input. Hourly water depth on streets simulated by the TUFLOW model is used as the surrogate model output. Results show that the surrogate model can match the duration and depth of flooding on streets from the TUFLOW model with high accuracy while reducing the computational time by a factor of 1,500. The surrogate model also exhibits the potential to identify problem spots in the physics-based model. In summary, this study shows how machine learning models trained on detailed physics-based models can provide a computationally efficient solution for real-time, street-scale flood warning within urban contexts.
Jonathan Goodall, University of Virginia
4:40 PM - 5:00 PM
Coastal communities are facing repetitive flooding caused by both tidal and rainfall-driven events. Climate change and sea level rise are resulting in more frequent and intense rainfall events, and sea-level rise is causing urban drainage infrastructure to be less effective during high tide periods when outfalls can be either partially or fully submerged. We are studying this problem and its impact on transportation and stormwater infrastructure systems within urban coastal communities. Our study region and partner in the research is the City of Norfolk, Virginia, USA. Norfolk is a historic town in coastal Virginia that is home to the largest Navy base in the world, the second busiest port on the United States East Coast, and is one of 100 Rockefeller Resilient Cities in the world. Our approach is to apply principles from the field of cyber-physical systems (CPS) to improve flood resiliency in coastal communities facing nuisance flooding. We are using real-time observational networks, crowdsourced data, physics-based and machine learning modeling approaches, model predictive control, and economic and social science methods to explore ways to better understand and mitigate the impacts of street-scale flooding. This presentation will provide a high-level overview of our research project including exploring (1) how real-time control of stormwater infrastructure systems can help to improve the resilience of these systems during nuisance flooding events by strategically holding back rainfall runoff and preventing tidally driven stormwater backups, (2) how both physics-based and machine-learning methods can be combined to real-time decision support, and (3) how reputation system approaches can be used to measure trust in crowdsourced rainfall datasets. These and related activities on the project are aimed at the common goal of leveraging real-time data from a variety of sources, innovative modelling techniques, and community-driven decision making to improve community resilience to nuisance flooding.