Hydrological models include errors when reproducing real-world observations, due to uncertainties in their components that inevitably propagate to the simulated variable. A large body of research in streamflow prediction blends statistical learning into the hydrological sciences, modelling river discharge using meteorological variables and catchment attributes as predictors of observed streamflow.

We developed a novel hybrid framework that integrates information from the process-based global hydrological model PCR-GLOBWB to reduce prediction errors in streamflow simulations. Our statistical methodology employs simulated streamflow and state variables from PCR-GLOBWB as additional predictors of observed river discharge. These model outputs provide supplemental information that is effectively used in a random forest, trained on a global database of streamflow measurements, to improve estimates of simulated river discharge across the globe. PCR-GLOBWB was run for the years 1979-2019 at 30arcmin and daily resolution, and the simulated state variables were then aggregated to monthly time steps. A single random forest model was trained with these state variables, meteorological data and catchment attributes, as predictors of observed streamflow from 2286 stations worldwide.

Results based on cross-validation show that the model is capable of discerning between a variety of hydro-climatic conditions and river flow dynamics, improving KGE of PCR-GLOBWB simulations at more than 80% of testing locations and increasing median KGE from -0.02 in uncalibrated runs to 0.52 after post-processing. Performance boosts are usually independent of availability of streamflow data at a particular station, thus making our method a potential candidate in addressing prediction in poorly gauged and ungauged basins.

Further research is still needed to test the potential influence of additional predictors describing catchment and time-series behaviour. Cluster analysis is required to understand why the post-processing framework still performs poorly at some stations. For prediction purposes, future efforts should also be directed at testing the model at higher spatial resolutions globally, and at finer temporal resolutions.

How to cite: Magni, M., Sutanudjaja, E. H., Shen, Y., and Karssenberg, D.: Global streamflow modelling using process-informed machine learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1160, https://doi.org/10.5194/egusphere-egu23-1160, 2023.

Stable isotopic composition modelling of water is an important part of resource management studies. We present a tool that estimates water stable isotope compositions using discontinuous inputs in time and space through machine learning algorithms. This tool has a multi-stage coupled algorithm that firstly calculates the parameters defined by the user that potentially affect the isotopic composition such as meteorological parameters, then, integrates the results of different parameters and generates the isotopic composition models for each time window. Isocompy time windows can be defined flexibly based on the amount of spatial-temporal properties of the available data. A variety of decision-making algorithms are implemented in this tool as an optional support to the user in different stages: from dataset preprocessing, outlier detection, statistical analysis, feature selection, model validation and calibration to postprocessing. Reports, figures, datasheets and maps could be generated in each step to clarify the underlying processes.

All in all, this tool aims (1) to offer an integrated, open-source Python library that is dedicated to the water isotopic composition statistical-regression modelling (2) to potentially improve our understanding of the precipitation stable isotopes by implementing novel machine-learning tools; and (3) to ensure reproducible research in environmental studies.

How to cite: Hassanzadeh, A., Valdivielso, S., Vázquez-Suñé, E., Criollo, R., and Corbella, M.: An open source library for environmental isotopic modelling using machine learning techniques, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1672, https://doi.org/10.5194/egusphere-egu23-1672, 2023.

An intensified hydrological cycle due to climate change is expected to increase precipitation extremes, but how river flood magnitudes will respond to this change remains disputed. Historically, there is only limited observational evidence that increasing precipitation extremes directly translate into systematically increased flood magnitudes. The incongruence between extreme precipitation and flooding is likely related to the compounding nature of various flooding drivers such as snowmelt and antecedent soil moisture. This complex interplay between flooding drivers makes it challenging to predict flood risks under warming. In order to better understand how precipitation extremes affect river floods in a warming climate, it is essential to disentangle the impacts of different drivers and conditions in flood generation. In this study, we employ an interpretable machine learning approach together with a large-sample hydrological dataset to identify the impact of various drivers in flood generation across a myriad of globally distributed catchments. We analyze how these impacts change with warmer temperatures and how – in response – their relationships with flood occurrence and magnitude change. The results indicate that increases in precipitation extremes have indeed contributed increasingly to flood generation in many regions over the historical period. The fact that flood magnitudes did not necessarily increase is likely a result of decreasing contributions of other drivers. We further investigate how future floods may change given the continuously rising trend of precipitation extremes. Overall, the study emphasizes the value of interpretable machine learning in helping understand how flood risks are likely to change in a warming climate.

How to cite: Jiang, S. and Zscheischler, J.: Revealing how precipitation extremes impact river floods in a warming climate with interpretable machine learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2343, https://doi.org/10.5194/egusphere-egu23-2343, 2023.

Groundwater potential mapping (GWPM) in the coastal areas is central for the decion making and development of the society and the environment. The current study endeavours the delineation of the groundwater potential zones of Korba coastal aquifer in the Cap-Bon Peninsula in the North East of Tunisia, using three different machine Learning techniques : random forest (RF), boosted regression tree (BRT), and the ensemble of RF and support vector machine (SVM). In order to achieve the objective, 17 groundwater influencing factors including elevation, slope, aspect, slope length (LS), profile curvature, plan curvature, topographical wetness index (TWI), distance from streams, distance from lineaments, lithology, geomorphology, soil, land use, normalized difference vegetation index (NDVI),Stream Power Index,Drainage Density and rainfall were considered for inter-thematic correlations and overlaid with wells locations and Transmissivity data in a spatial database. A total of 225 wells locations were identified, which had been divided into two classes: training and validation, at the ratio of 70:30, respectively. The RF, BRT, and RF-SVM ensemble models have been applied to delineate the groundwater potential zones. These models were validated with area under the receiver operating characteristics (AUROC) curve. The accuracy of RF (92%) and hybrid model (89.2%) was more efficient than BRT (85.6%) model. The results of the study will help the decision-makers, government agencies, and private sectors for sustainable planification of  groundwater resources in the study area.

How to cite: Khammessi, I. and omar, H.: Application of machine learning techniques in groundwater potential mapping in the Korba coastal aquifer Cap Bon Peninsula, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14581, https://doi.org/10.5194/egusphere-egu23-14581, 2023.

Efficient estimation and forecast of reference evapotranspiration (ET_O) is crucial for water resources management and for developing an efficient irrigation practice that will help better utilization of scanty water resources. This is a more challenging task in data scarce regions. This study aimed at multi-step ahead prediction of ET_O across different cropping seasons and agro-climatic regions in India with six Machine learning (ML) based techniques using globally available fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) gridded reanalysis products (ERA5). For real-time, one-day, two-day, and seven-day ahead prediction of ET_O, this study evaluates and compares the capability and prediction accuracy of deep learning algorithms, i.e., Support Vector Regression (SVR), Multivariate Adaptive Regression Splines (MARS), Random Forest (RF), Multi-Layer Perceptron (MLP), one-dimensional Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). ML based models were developed using meteorological observations along with ERA5 inputs at three meteorological stations: Nagpur, Hyderabad and Bhubaneswar. The results indicate that MLP, SVR and CNN outperform other ML algorithms. High performing models, one each (MLP and SVR) from neural network-based models and kernel-based models respectively, are further utilized to scale up the analysis for gridwise ET_O forecasting across the whole of India. The Global Land Evaporation Amsterdam Model (GLEAM) dataset has been used as reference to evaluate gridwise ET_O forecasts. ET_O predicted by MLP model shows better agreement with GLEAM ET_O values during the Rabi cropping season (October-March) (MAE = 0.103 mm/day and NRMSE = 3.9 %) than during the Kharif season (June-September) (MAE = 0.151 mm/day and NRMSE = 4.5 %). As expected, the accuracy of the models drops with increase in the prediction horizon from real-time to seven-day; for instance, with MLP, MAE = 0.146 mm/day, R² = 0.955 for real-time and MAE = 0.173 mm/day, R² = 0.939 for seven-day ahead prediction over arid agro-climatic zone during Rabi season. However, even the minimum forecast performance observed in the semi-arid tropics region during Rabi season is reasonably good (MAE = 0.396 mm/day, R² = 0.704 for real-time evaluation and MAE = 0.445 mm/day, R² = 0.56 for seven-day ahead). This strengthens the potential of the proposed models for multi-step ahead ET_O forecasting across varied cropping seasons and agro-climatic regions without depending on meteorological station data.

How to cite: Mandal, N. and Chanda, K.: Comparison of Neural network based and Kernel based Machine Learning approaches for daily forecasting of reference evapotranspiration in data scarce regions, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11950, https://doi.org/10.5194/egusphere-egu23-11950, 2023.

Intermittent streams, where water ceases to flow during some time, are a unique habitat for freshwater biota that are adapted to these conditions and provide many ecosystem services. Shifts in intermittency patterns, for example due to climate change, are problematic. To quantify streamflow intermittency in all of Europe at a spatial resolution of 15 arc-sec (approx.. 500 m), we developed a machine learning approach that combines daily streamflow observations, the output of a global hydrological model as well as other physiogeographic data to estimate monthly time series of the number of no-flow days.

Daily streamflow observations at initially a selection of initially close to 2000 stations gauging stations across Europe from the SMIRES, GRDC and GSIM databases were used as target for training the ML model. We selected those stations with at least 18 complete (no day is missing) monthly records in the period 1980-2019. Predictors include monthly time series of simulated hydrological indicators at two spatial resolutions, 15 arc-sec (high resolution HR) and 0.5 arc-deg (approx. 50 km, low resolution LR) as well as static HR environmental indicators (e.g. drainage area).  The hydrological indicators were derived from the global hydrological model WaterGAP 2.2e. Its native LR output including surface runoff, and groundwater discharge was used for computing HR time series of monthly streamflow across all of Europe. A comparison of streamflow observations shows a reasonable fit to observations. HR hydrological indicators include specific streamflow in current and previous months.  Examples for LR hydrological predictors include the groundwater recharge to total runoff ratio and daily streamflow variability with each month.

We considered a sequential statistical modeling approach (in the first stage: binary classification, and in the second stage: multiclass classification) owing to the zero-inflated and imbalanced data issues. In the first stage, a Random Forest (RF) model is built up to classify a binary classification of each month as either intermittent (with at least one no-flow day) or perennial. Then, by taking into account only those stations that were in the first step either predicted or observed to be intermittent, we developed another model to predict four classes of intermittency (e.g. with 1-2, 3-15, 16-27, 28-31 of no-flow days per month). A random oversampling of non-perennial gauging stations was implemented for both stages in order to address the biases in the RF model caused by the class imbalance in the training data. Three cross-validation techniques were applied for estimating the model performance, hyperparameter tuning, and model selection, including non-spatial, spatial, and spatial-temporal cross-validations. Balanced class accuracy, sensitivity, specificity, and precision supported the model selected. The most important predictors for streamflow intermittency will be presented as well as the spatial distribution of the four intermittency classes in Europe (without Russia).

How to cite: Abbasi, M., Trautmann, T., and Döll, P.: Developing a random forest model to quantify streamflow intermittency in Pan-Europe at a spatial resolution of 15 arc-sec, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-4440, https://doi.org/10.5194/egusphere-egu23-4440, 2023.

While the observations from earth-observing satellites and in-situ weather meteorological and monitoring stations continue to expand, researchers are to deal with an abundance of data and, consequently, with a long modeling procedure. Machine learning algorithms, as universal nonlinear function approximation tools, are effective in analyzing and modeling spatio-temporal environmental data, more efficiently, either in time or in terms of the availability of various variables, than physically-based models.

Soil moisture is an essential climatic parameter, especially for understanding and forecasting variations in surface temperature, precipitation, drought, flood, and the effects of climate change. As a parameter with high spatial and temporal variability, it is a strong necessity for predictive models that embed spatially irregular measurements, which stand for spatially distributed weather meteorological and monitoring stations. To date, most approaches, that have been documented in the literature, model environmental data only at the discrete locations of the monitoring stations.

This research aims to employ a recently proposed methodology, for spatio-temporal prediction of environmental data (Amato et al., 2020), and to propose a new methodology for spatio-temporal prediction of soil moisture in lowland areas, making use of the basic hydrologic premise that precipitation and temperature strongly hinge on topography. The features of the machine learning models used to predict soil moisture within the research area are the meteorological parameters of several agro-meteorological weather stations that have been installed at the site.

The study area is the plain of Arta, located in the Epirus region (NW Greece), and includes the lower part of the watersheds of rivers Aracthos and Louros. The final receptor for upstream surface water and groundwater is a sensitive and complex system of wetlands, the marine ecosystem of Amvrakikos. The research area is receiving a lot of attention because of its agricultural characteristics and water infrastructures (extensive irrigation and drainage network, pumping stations, and hydroelectric dam).

Acknowledgments: This research has been co-financed by the European Union and Greek national funds through the Operational Program Competitiveness, Entrepreneurship and Innovation, under the call SUPPORT OF REGIONAL EXCELLENCE (project code MIS: 5047059).

Amato, F., Guignard, F., Robert, S., & Kanevski, M. (2020). A novel framework for spatio temporal prediction of environmental data using deep learning. Scientific Reports, 10(1), Article 1. https://doi.org/10.1038/s41598-020-79148-7

How to cite: Chrysanthopoulos, E., Pouliaris, C., Tsirogiannis, I., and Kallioras, A.: Forecasting soil moisture on a spatial and temporal scale using machine learning algorithms, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11925, https://doi.org/10.5194/egusphere-egu23-11925, 2023.

Water sciences have greatly contributed to the proliferation of machine learning in the twenty-first century, especially through engineering hydrology. This process has consequently necessitated transfer of core theory and knowledge of machine learning to the domain of hydrological sciences. In this regard, it is noteworthy that published academic literature played a substantial role in supporting development and learning of hydrologists. Specifically, research articles (and book sections) that review machine learning concepts and algorithms along with their applications in hydrology bolster progress of science by presenting encapsulated information (e.g, in the form of literature synthesis). Despite the rapid increase in the number and scope of such research articles, a systematic understanding of how this line of research publications has evolved with respect to their scientific context, objectives, and methods is still lacking. Hereby, we present an analysis of review papers in hydrology and machine learning based on a  systematic search strategy. The overview includes analysis of bibliographic information, review types (objective, focus theme, etc.), review methodologies (narrative, systematic, etc.) as well as thematic context (hydrology subjects and machine learning topics). We believe that our analysis can provide important insights into topics and discussions in hydrology and machine learning that need further exploration by hydrologists. Furthermore, the public online library on Zotero (https://www.zotero.org/groups/4828386/machine_learning_hydrology_review_papers/library) might encourage more participation towards sustainable literature search and active reading on this subject at the intersection of two fundamental disciplines, machine learning and hydrology.

How to cite: Dogulu, N., Olusola, A. O., and Papacharalampous, G.: Machine learning and hydrological sciences: A systematic overview  of review papers, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10575, https://doi.org/10.5194/egusphere-egu23-10575, 2023.

Real-time reservoir operations are highly dependent on decisions made by reservoir operators, which are difficult to simulate accurately with process-based hydrological models. Data-driven models, particularly those based on machine learning (ML), have been shown to be able to overcome the limitations typically encountered in process-based models. Despite a large number of ML studies in reservoir operation modelling, only few studies have focused on ML model performance in real-time reservoir operation and outflow forecasting. This study aims to investigate the capabilities of the Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) and Gated Recurring Unit (GRU) in simulating and reforecasting real-time (daily) reservoir operation and outflow, considering uncertainties in input data, training-testing periods and different algorithms. A major, multi-purpose reservoir, namely the Sirikit reservoir, in the upper Chao Phraya River basin Thailand was used as the case study. The main inputs for the ML operation models included the daily reservoir outflow, inflow, storage and the month of the year. We applied the distributed wflow_sbm model for inflow simulation (using MSWEP precipitation data) and inflow reforecasting (using ECMWF precipitation data). Daily reservoir storage was obtained from observations and real-time recalculation based on the reservoir water balance. The ML operation models were trained and tested with 10-fold cross-validation. Results show that RNN, LSTM and GRU can reconstruct real-time reservoir operation and provide accurate outflow when training data cover both regular and extreme conditions. For multi-day reforecasting, the model performances are appropriate for the current day up to 2-day lead times for low outflows and up to 6-7 days for high outflows. GRU is potentially the most accurate, robust and convenient model to be used in practice. We conclude that with some further improvements, the ML operation models can be effective and applicable tools to support decision-making for real-time operational water management.

How to cite: Wannasin, C., Brauer, C., Uijlenhoet, R., and Weerts, A.: Modelling and reforecasting real-time reservoir operation and outflow with neural networks: case study of the multi-purpose Sirikit reservoir in Thailand, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14421, https://doi.org/10.5194/egusphere-egu23-14421, 2023.

Changes in the level of the world freshwater storage, the Terrestrial Water Storage Anomalies (TWSA), may be induced by natural variability, climate change, and human activities. Since 2002 the Gravity Recovery and Climate Experiment (GRACE) has been measuring the Earth’s gravity field providing estimates of the TWSA at the global scale.

Here, we aim to develop a machine learning model that can reproduce the GRACE monthly time series covering the period between 2002 and 2017 from climate data, identifying to what extent the TWS fluctuations have been caused by climate variability. We used a Long Short-Term Memory (LSTM) neural network trained with meteorological variables (precipitation, air temperature, solar net radiation, snow cover, relative humidity, and leaf area index) and soil properties data (soil porosity, soil texture, and clay, sand, and silt fractions). Our results show that the model is able to consistently reconstruct the observed freshwater anomalies, especially in the humid regions. Furthermore, we observed that as climate change trends are removed from input data, the bias between model predictions and observed data becomes larger, proving the influence of climate change on TWSA.

How to cite: Palazzoli, I., Ceola, S., and Gentine, P.: Predicting Terrestrial Water Storage Anomalies at the Global Scale with a Machine-Learning Model, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-16641, https://doi.org/10.5194/egusphere-egu23-16641, 2023.

Snow plays a significant role in the hydrology of numerous regions across the globe. A major portion of precipitation above 45^O N latitude falls as snow. The accumulated snow melts slowly and contributes to infiltration and runoff processes. Therefore, studying the quantity and fate of water from snowmelt is essential. Reduced snow storage would lead to snow droughts, which can have an enormous impact on the water resources of snow-dominated catchments, such as those in the western United States. For that reason, it is essential to identify the time and severity of snow droughts efficiently. This study proposes SnoDRI, a new index that could identify and measure snow drought events. SnoDRI is a machine learning-based index estimated from several snow-related variables utilizing novel machine learning algorithms. The model uses a combination of mutual information and a self-supervised learning algorithm of an autoencoder. We use random forests for feature extraction for SnoDRI and to assess the importance of each variable. We use NLDAS-2 reanalysis data from 1981 to 2021 for the Western United States to study the efficacy of the new snow drought index. The results are validated by verifying the coincidence of actual snow drought events and the interpretation of our new index. We will discuss how well the new drought index performs and help in better identification of snow droughts.

How to cite: Rasiya Koya, S., Kanti Kar, K., Srivastava, S., and Roy, T.: SnoDRI: A Machine Learning Based Index to Measure Snow Droughts, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8968, https://doi.org/10.5194/egusphere-egu23-8968, 2023.

Freshwater availability in coastal areas depends on the withdrawal from tidal rivers and is severely threatened by saltwater intrusion, especially in the dry season. Freshwater availability is associated with natural factors and human activities. Although analyses of freshwater availability under saltwater intrusion is problematic, it has received limited attention in the literature. We propose a new framework, i.e. regulation by avoiding saltwater withdrawal (RASW), where the relationships among saltwater intrusion, upstream streamflow, and water supply are established, using hybrid data-driven method coupling wavelet transform and random forest, and considering data on streamflow, tide, wind, salinity of withdrawal stations, capacities of withdrawal projects and reservoirs, and water demand. RASW contains three phases, i.e. estuary salinity-exceedance simulation, upstream streamflow distribution design, and local water supply security analysis. The method is tested on the water supply for Zhuhai-Macao of the Guangdong-Hong Kong-Macao Great Bay Area, South China. Results demonstrate that the salinity-exceedance simulation model using a hybrid data-driven method is quite accurate. The meta-Gaussian copula efficiently simulates the six-dimensional distribution of upstream monthly streamflow and is appropriate for streamflow distribution scenario design. Water supply security benefits greatly from the joint river-reservoir regulation mode. But for a given exceedance frequency of average streamflow, the modes and security situations are diverse, due to various streamflow distributions, i.e. extremely low streamflow and its occurrence time. The proposed framework facilitates integrated decision-making for water supply security in coastal areas. Moreover, the capacities of facilities should be carefully considered according to local conditions, and streamflow distribution design can be utilized as a management tool to regulate water supply system. 

How to cite: Wu, H., Tu, X., Chen, X., Singh, V. P., Alfonso, L., Lin, K., Liu, Z., and Lai, R.: A Framework for Water Supply Regulation in Coastal Areas by Avoiding Saltwater Withdrawal Considering Upstream Streamflow Distribution, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-3299, https://doi.org/10.5194/egusphere-egu23-3299, 2023.

In recent years data-driven techniques, specifically LSTMs, have outperformed conceptual hydrological models for rainfall-runoff prediction. However, even though great progress has been made to explain the internal functioning of the model ((Kratzert, et al., 2019); (Lees, et al., 2022)), their interpretation is still not as straightforward as conceptual models. Additionally, latent variables, different from the target quantity, need postprocessing methods to be extracted. One way to combine the flexibility of data-driven techniques with the interpretability of conceptual models is the use of hybrid models. In our contribution,  we will present results from applying a similar technique as (Kraft, Jung, Korner, & Reichstein, 2020) and (Feng, Liu, Lawson, & Shen, 2022), in which an artificial neural network dynamically calculates the parameters of the conceptual model. This approach increases the model flexibility, allows the inclusion of multiple information sources, and compensates for model uncertainty, while maintaining the straightforward interpretability of the conceptual part. In this contribution, we will look at the performance of the hybrid model, analyze the parameter variation over time, and present a technique to avoid parameter cross-compensation.

References

Feng, D., Liu, J., Lawson, K., & Shen, C. (2022). Differentiable, learnable, regionalized process-based models with multiphysical outputs can approach state-of-the-art hydrologic prediction accuracy. Water Resources Research. doi:https://doi.org/10.1029/2022WR032404

Kraft, B., Jung, M., Korner, M., & Reichstein, M. (2020). HYBRID MODELING: FUSION OF A DEEP LEARNING APPROACH AND A PHYSICS-BASED MODEL FOR GLOBAL HYDROLOGICAL MODELING. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 1537--1544. doi:https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-1537-2020

Kratzert, F., Herrnegger, M., Klotz, D., Hochreiter, S., & Klambauer, G. (2019). NeuralHydrology--Interpreting LSTMs in Hydrology. In W. Samek, G. Montavon, A. Vedaldi, L. Hansen, & K.-R. Müller, Explainable AI: Interpreting, Explaining and Visualizing Deep Learning (pp. 347--362). Springer.

Lees, T., Reece, S., Kratzert, F., Klotz, D., Gauch, M., De Bruijn, J., . . . Dadson, S. (2022). Hydrological concept formation inside long short-term memory (LSTM) networks. Hydrology and Earth System Sciences, 3079-3101. doi:https://doi.org/10.5194/hess-26-3079-2022

How to cite: Acuna, E., Ehret, U., Bäuerle, N., and Loritz, R.: Hybrid modelling in hydrology by Neural Network-based prediction of conceptual model parameters, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-3575, https://doi.org/10.5194/egusphere-egu23-3575, 2023.

How to cite: M. Ndiwago, D., Nijzink, R., Ley, C., J. Schymanski, S., and S. Hale, J.: Thermodynamic integration via Replica Exchange Hamiltonian Monte Carlo for faster sampling and model comparison, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2910, https://doi.org/10.5194/egusphere-egu23-2910, 2023.

Over the past decades, urban non-point source (NPS) pollution has been the most severe threat to the urban water environment. The sharp increase of impervious surface and the high level of particulate matter from massive human activities exacerbated the water quality of surface runoff, leading to the significant urban NPS pollution globally whereby it is of importance to have a deep knowledge on the accumulation and transport of pollutants. A series of traditional physical models have been developed to simulate the runoff generating as well as the NPS pollution. However, a disadvantage of process-based modelling is its great demand for a large amount of field data which may normally be inaccessible, as well as the demand for the expertise in applying appropriate modelling method on specific study area. Empirical models do not characterize complex physical processes of NPS pollution and thus require fewer data and modelling skills. Nevertheless, the limitation is that these modelling approaches are region-sensitive and spatially untransferable. It is challenging to fill the gap between the requirement of urban water environment management and existing modelling performance on NPS pollution, in the absence of a more effective model with high accuracy, easy employment, and spatial transferability.

Machine learning approach has been utilized in environmental studies for decades, and was originally believed to be a black box that can barely provide any physical insight into environmental processes. However, an approach named physics-informed neural networks (PINN) was proposed lately and then applicated in dynamical system. This approach embeds differential equations of priori knowledge into neural network to make modelling interpretable and generalizable. In this study, a physical process embedded LSTM network was proposed to formulate the cumulation and transport of urban NPS pollution in rainfall runoff, based on the coupling of LSTM and differential equations of classic exponential build-up/wash-off processes. Water quality data of urban runoff from sampling and continuous real-time monitoring campaigns distributed in China, USA and New Zealand were collected and fed into proposed network to model the primary NPS pollutant TSS. The results revealed that the hybrid PINN model excels the vanilla LSTM approach and auto-calibrated SWMM approach in accuracy and convenience. The interpretable model also enhanced the cross-catchment transferability of model for urban water management in data-poor area. In addition, the trained parameters of network units were found consistent with the prior knowledge of accumulation and transport of NPS pollutants, indicating the deep coupling of neural network and physical process. As a very early case of hybrid AI modelling in urban NPS pollution, this study provided a new perspective on water quality modelling and can help in improving the standards of urban environment governance.

How to cite: Tang, S., Wan, Y., Shang, F., Wang, S., and Jiang, J.: Urban non-point source pollution modelling: A physics-informed neural network approach, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2543, https://doi.org/10.5194/egusphere-egu23-2543, 2023.

 The water distribution network (WDN), a vital component of the water supply system, is an essential urban infrastructure distributing potable water to society. Its design, a non-deterministic polynomial-hard problem, has been a widely studied complex research problem for decades, with various optimization models proposed for its optimal design. Recent advancements in enhancing the computational efficiency of stochastic optimization algorithms by introducing chaotic force have elevated the scope of formulating chaos-directed evolutionary algorithms (EAs). The present study proposes one such approach, the chaos-directed genetic algorithm (CDGA) model, to improve the search mechanism of the genetic algorithm (GA) in solving the complex optimization problem of WDN optimal design.

In one of our recent works, the influence of chaotic maps with high-dimensionality, the Henon and Lorenz maps, are explored and compared to the low-dimensional Logistic map in improving the performance of GA. With the one-dimensional Logistic map demonstrating better computational improvement of GA, the present study considers it for formulating the CDGA model. The Logistic map is a non-linear first-order difference equation. Its dynamics evolve into various possible states of system range without repetition. For the search mechanism of the optimization technique to explore different regions of search space, this particular characteristic forms the most favorable feature. Consequently, by incorporating the chaotic force of the Logistic map into GA’s evolutionary mechanism by replacing every random search phenomenon, the CDGA model is formulated. A novel method of non-sequential allocation of chaotic dynamics is employed to induce chaotic force. Notably, the method is unique, using the same initial characteristics of the Logistic equation, retaining the chaos ergodicity for the evolutionary search.

To demonstrate the computational efficiency of the CDGA model, the enormously studied benchmark problem, the Hanoi network (HN), is considered. HN is a 34-dimensional problem having a complex search space with multiple locally optimal solutions. Defining the WDN optimization problem as the single-objective design framework subjected to linear and non-linear constraints of governing laws, the principal objective is to minimize the investment cost of HN pipes. While minimizing the pipe investment cost, the constraints levied ensure that the HN is hydraulically adequate to deliver the design demands. Thus, the optimization model formulated is the integrated framework with the simulation tool to simulate the WDN's hydraulic conditions. The code for the CDGA model is written in MATLAB R2015a and combined with the simulation software, EPANET 2.0, using the EPANET-MATLAB toolkit.

The computational results demonstrate the convergence precision of the CDGA model over its traditional GA, converging to the optimal cost of 6,081,564 units, the previous best solution reported for HN in the literature. Moreover, it outperforms many stochastic optimization models reported in the literature with computational efficiency in solving HN, particularly simulated annealing, shuffled complex, shuffled frog leaping, ant colony, particle swarm, harmony search, krill herd, and cuckoo search algorithms. Hence, from the results, the study suggests formulating chaos-directed optimization algorithms to improve their traditional model's computational efficiency in solving complex optimization problems.

Keywords: Water distribution network optimal design; Evolutionary algorithms; Genetic algorithm; Chaotic maps; Logistic equation; Hanoi network

How to cite: Poojitha, S. N. and Jothiparakash, V.: Application of Enhanced Search Technique: Chaos-Directed Genetic Algorithm in Optimal Design of Water Distribution Network, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-641, https://doi.org/10.5194/egusphere-egu23-641, 2023.

How to cite: Gute, E., Ickes, L., and Pechlivanidis, I.: Assessing water observation network settings by hydro-geological sub-sampling of a large data set for Sweden, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5491, https://doi.org/10.5194/egusphere-egu23-5491, 2023.

Hydrological models are important tools for management of water resources at the basin scale. However, the outputs of these models might come with significant inaccuracies, often due to uncertainties in model parameters. One of the biggest challenges in working with a hydrological model is that these models require rigorous calibration, validation, and uncertainty analysis. In recent years, there has been an increase in the use of heuristic optimization techniques in water resources research. These techniques can yield more accurate and more reliable modeling results by searching the global optimum of multiple model parameter sets.

This study describes the application of a heuristic optimization method, the Particle Swarm Optimization (PSO), on a hydrological model, the Soil and Water Assessment Tool (SWAT). The model is applied to the Fetrek Stream watershed in western Turkiye, which is under environmental stress due to excessive groundwater abstraction and pollution from numerous wastewater discharges. The model includes data related to 41 point sources, and two inflowing tributaries. The model is configured with 8 sub-basins, and 484 hydrologic response units. Hydrological fluxes are obtained for a 30-year simulation period. The sensitivities of the model parameters and uncertainties of model results are investigated. PSO is used to calibrate sensitive model parameters, followed by a comparison with the calibration outcome using the SUFI-2 (Sequential Uncertainty Fitting) algorithm, which is the usual choice in calibrating SWAT models. The performances of both optimization approaches are evaluated with the Kling-Gupta Efficiency (KGE), the regression coefficient (R²), and the bias percentage  (PBIAS) Model results are presented with their associated prediction uncertainties in the form of the so-called p-factor and r-factor statistics, which represent envelopes of good model solutions. The results show that the PSO approach can achieve satisfactory results on a monthly time-scale thereby offering an alternative calibration with less parameters and a wider interval of the 95% prediction uncertainty.

This study is supported by the PRIMA program under grant agreement No: 2024 Project TRUST (management of industrial Treated wastewater ReUse as mitigation measures to water Scarcity in climaTe change context in two Mediterranean regions). The PRIMA program is supported by the European Union.

How to cite: Çakır, F., Elçi, A., and Somay-Altaş, M.: Calibration of the SWAT Hydrological Model with the Particle Swarm Optimization Technique, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-3428, https://doi.org/10.5194/egusphere-egu23-3428, 2023.

This article reports the findings of a recent study by Das and Chanda (2022), wherein a Bayesian Network (BN) approach was applied to analyze the influence of large-scale climate modes and local hydro-meteorological variables on streamflow and rainfall in four river basins in India. Bayesian Networks (BN) offers a thorough conditional independence structure that can improve comprehension and forecasting of hydroclimatic systems. This served as the main impetus for the work, which explored the relative contributions of large-scale climate modes and local hydro-meteorological variables for the prediction of rainfall and streamflow at the basin scale. Once the conditional independence structure is developed, variables possessing a ‘directed arc’ from the target variable were selected as the potential predictors for developing the prediction models. The results showed that the most important predictors for streamflow were rainfall, u-wind, and soil moisture, while the most important predictors for rainfall were u-wind, air temperature, geo-potential height, precipitable water, vertical velocity, and relative humidity. The analysis also revealed that the influence of large-scale climate modes on the target variables was generally insignificant, except for the Pacific Decadal Oscillation and El-Niño Southern Oscillation. Furthermore, the network structure showed that about 87 and 97% of the initial inputs are redundant. The accuracy of the prediction models are comparable across all of the basins and is higher for rainfall (Refined index of agreement (MD) ranging from 0.61 to 0.81) than for streamflow (MD ranging from 0.61 to 0.78). The study also found that dry, intermediate, and wet months can be satisfactorily classified using two drought indices, the Standardized Drought Index (SDI) for streamflow and the Standardized Precipitation Anomaly Index (SPAI) for rainfall.

Keywords: Large-scale climate modes, local hydro-meteorological variables, Bayesian Networks (BN), Standardized Drought Index (SDI), Standardized Precipitation Anomaly Index (SPAI)

Reference:

Das P, Chanda K (2022) A Bayesian network approach for understanding the role of large-scale and local hydro-meteorological variables as drivers of basin-scale rainfall and streamflow. Stoch Environ Res Risk Assess 2:. https://doi.org/10.1007/s00477-022-02356-2

How to cite: Das, P. and Chanda, K.: Influence of large-scale climate modes and local hydrometeorological factors in predicting basin scale rainfall and streamflow: A Bayesian network approach, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-3826, https://doi.org/10.5194/egusphere-egu23-3826, 2023.

Several long-term gridded datasets have become available in the last decades on a global scale, at increasing spatial and temporal resolution, with low latency times. These datasetshave opened new opportunities to advance Earth Sciences modelling studies at different spatial and temporal scales, especially in poorly gauged areas. However, working with (hundreds of) thousands of raster time series (e.g., for (sub)daily precipitation), usually in different vectorial and raster formats, impose high computational challenges to efficiently analyse all the gridded datasets.

In this work we introduce a new R package for easy processing and analysis of raster time series, to bring the use of gridded data closer to the Earth Sciences community. This package expands the large number of spatial functions provided by the terra package by taking advantage of the time attribute of raster objects. A particular emphasis of the package is exploring and comparing gridded datasets of hydrological variables with different time frequencies (e.g., sub-hourly, hourly, daily, monthly, seasonal, annual).

General purpose functions include temporal subsetting, resampling, cropping, extracting time series for points or polygons, comparing two datasets using summary statistics, and exporting a raster time series as a collection of daily/monthly/annual files, each one of them with several layers of a higher temporal frequency. Also, temporal aggregation is possible from sub-hourly to hourly/daily/weekly/monthly/annual, among others. Most functions can take advantage of multi-core computers and network clusters, to reduce the computational burden.

To illustrate the use of this new package, we compared different state-of-the-art gridded precipitation products (CHIRPSv2, CMORPH v1.0, IMERGv06B, MSWXv1.0, ERA5, ERA5-Land, CR2METv2.5), all of them with different data formats and spatial resolutions, using continental Chile as a case study. However, based on the package's flexibility and ease of use, we hope the broader community of hydro-scientists and water-engineers will use it to visualise the spatio-temporal variation of key hydrological/environmental variables,to carry out time series analysis, to combinedifferent types of models and data sources, and to improve our integrated knowledge of the water cycle.

How to cite: Zambrano-Bigiarini, M. and Bernal Vallejos, S.: A new software for spatio-temporal analysis of gridded data sources, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-16978, https://doi.org/10.5194/egusphere-egu23-16978, 2023.

Satellite precipitation estimation provides crucial information for those places lacking rainfall observations from ground–based sensors, especially in terrestrial or marine areas with complex climatic or topographic conditions. This is the case over much of Western China, including Upper and Middle Lancang River Basin (UMLRB), an extremely important transnational river system in Asia (the Lancang–Mekong River Basin) with complex climate and topography that has limited long–term precipitation records and high–elevation data, and no operational weather radars. In this study, we evaluated three GPM IMERG satellite precipitation estimation (IMERG E, IMERG L and IMERG F) over UMLRB in terms of multi–year average precipitation distribution, amplitude consistency, occurrence consistency, and elevation–dependence in both dry and wet seasons. Results demonstrated that monsoon and solid precipitation mainly affected amplitude consistency of precipitation, aerosol affected occurrence consistency of precipitation, and topography and wind–induced errors affected elevation dependence. The amplitude and occurrence consistency of precipitation were best in wet seasons in the Climate Transition Zone and worst in dry seasons in the same zone. Regardless of the elevation–dependence of amplitude or occurrence in dry and wet seasons, the dry season in the Alpine Canyon Area was most positively dependent and most significant. More significant elevation–dependence was correlated with worse IMERG performance. The Local Weighted Regression (LOWERG) model showed a nonlinear relationship between precipitation and elevation in both seasons. The amplitude consistency and occurrence consistency of both seasons worsened with increasing precipitation intensity and was worst for extreme precipitation cases. IMERG F had great potential for application to hydroclimatic research and water resources assessment in the study area. Further research should assess how the dependence of IMERG’s spatial performance on climate and topography could guide improvements in global precipitation assessment algorithms and the study of mountain landslides, floods, and other natural disasters during the monsoon period.

How to cite: Lu, C.: Assessment of GPM IMERG Satellite Precipitation Estimationunder Complex Climatic and Topographic Conditions, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-16938, https://doi.org/10.5194/egusphere-egu23-16938, 2023.

Achieving improved predictions in ungauged-basins or inferring the effects of climate and land-use changes on streamflow requires hydrologists to first learn the underlying mechanisms behind streamflow generation in gauged-basins. One way to characterize streamflow generation is by quantifying how catchments filter rainfall into streamflow. A simple and popular technique that displays the rainfall-streamflow linkage is the unit hydrograph. Though one could characterize and classify catchments based on their unit hydrographs, this approach implicitly implies that the function and response-time that link rainfall to streamflow are time-invariant. The celerity, and the function that links rainfall to streamflow in a given catchment, could vary from catchment to catchment as well as from season to season. This is primarily due to variations in antecedent wetness, temperature, vegetation transpiration and the ways climatic factors interact with biophysical factors, over time and over space. In this study, we utilize sparse historical functional linear models to quantify the time-variant rainfall-streamflow response function, across hundreds of catchments in North America. The function reflects the temporally varying relationship between rainfall and streamflow and can be used to infer temporally varying response times. We then attempt to relate catchment characteristics such geology, climate, and topography to the characteristics of rainfall-streamflow response function and response time, spatially and temporally.  We argue that our study extracts generalizable and robust process understanding in a novel data-driven manner.

How to cite: Ameli, A., Janssen, J., Meng, S., Cao, J., and Welch, W.: The spatial-temporal variations of rainfall-streamflow linkage across North America: A functional data analysis approach, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-4028, https://doi.org/10.5194/egusphere-egu23-4028, 2023.

The change of streamflow patterns is one of the important information to water resources management. Especially, in the last few years, climate changes have not only caused increases in the intensity and frequency of extreme hydrological events, but also disrupted the monotony of climate which could lead to unexpected consequences and economic losses to our society. However, only a little research has paid attention to the change in patterns or schemes of the hydrological cycle. The issue is even more serious in Taiwan due to the uneven spatial-temporal distribution of rainfall. This research aims to discover the change in patterns in Taiwan. We proposed a HDCE framework which is composed of Hierarchical cluster, Dynamic time warping, Change point detection, and Empirical mode decomposition. With this framework, we apply the hierarchical cluster with different distance matrices, and obtain the optimal clustering number and linkage method according to clustering valid indexes. Dynamic time warping is used as the measure of distance to investigate the pattern of time series By this way, this framework determines the optimal cluster of patterns for the historical inflow data. After clustering, the AMOC (At Most One Change) is used to analyze the structure of the pattern. With AMOC, the change time point in each period is examined under the structure of each cluster. In the end, the empirical mode decomposition is adopted to determine the trend of the pattern change. With the proposed framework, this research applies these schemes to main inflow observation gage stations in Taiwan, and the results demonstrate that the groups and activities of positions of the stations indirectly affect the pattern of the inflow values, instead, the clusters formed are mainly affected by the region and geographical area of the locations. Furthermore, we will explain the results showing the different trends in changing time in each region and the correlation of breaks at each station. In this way, the results of HDCE not only examine the occurrence of droughts but provide information that is useful to develop the strategy to reduce the loss through better water management.

How to cite: Wen, C. and You, J.-Y.: Investigating the change of hydrological patterns of streamflow by using  HDCE method, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-4297, https://doi.org/10.5194/egusphere-egu23-4297, 2023.

The information entropy method, based on mass sampling data processing, has been widely applied in environment monitoring and management. However, previous efforts have been mainly limited to the optimization of stations in discrete space with no association to critical events and their associated temporal scale. In particular, further research on integrating water quality monitoring under critical events (such a spillway accident) and the related cost-benefit analysis of environment management decisions are needed. In this study, we give an entropy-based paradigm of water quality reaction criteria R, which is analogous to the definition of Gibbs free energy (ΔG) in thermodynamics. Then we propose a systematic framework of entropy prisms (HPrisms) with four entropy indexes: dilution index (E), flux index (F), spatial entropy index (G_x) and temporal entropy index (G_t). They describe the pollutants transport process in water bodies from different perspectives, facing different water environmental management decisions. The corresponding reaction criteria of these four entropy indexes for different water quality management scenarios are defined for different spatiotemporal scales where different criteria are applicable. The method has value in emergency monitoring in rivers and lakes, useful for anomaly detection, key point identification and other water environment management scenarios. This study is a generic theoretical framework so far, and we will present specific critical cases for management reaction criteria to find the quantitative relationship between reaction criteria with information entropy indexes.

How to cite: Pang, T., Jiang, J., Alfonso, L., Wang, P., and Zheng, T.: Information entropy for assisting decision-making for critical events in surface water quality management, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9056, https://doi.org/10.5194/egusphere-egu23-9056, 2023.