PhD Thesis of
92. Quantification of model uncertainties in groundwater drought simulations under climate change
Author: Syed Md Touhidul Mustafa
Supervisor(s): Prof. Dr. ir. Marijke Huysmans
Date: 13 February 2018
Groundwater drought can be described as a temporary decrease in groundwater availability over a significant period of time. This PhD (1) addresses the influencing factors on groundwater drought, (2) quantifies the groundwater level prediction uncertainty in climate change impact studies using an ensemble of representative concentration pathways, global climate models, multiple alternative conceptual groundwater flow models and groundwater abstraction scenarios and (3) develops and applies an Integrated fully Bayesian Multi-model Uncertainty Estimation Framework (IBMUEF) to quantify the model uncertainty originating from errors in the model conceptualization, the input data, the parameter values and measurements. Investigating the effect of different climatic and anthropogenic factors on groundwater drought provides essential information for sustainable planning and management of (ground) water resources. In the first part of this study, a multi-step methodology is proposed to understand the influencing factors on groundwater drought. The results show that the evapotranspiration and rainfall deficits are determining meteorological drought, which show a direct relation with groundwater recharge deficits. Land-cover change has a small effect on groundwater recharge but does not seem to be the main cause of groundwater-level decline (depletion) in the study area. The groundwater depth and groundwater-level deficit (drought) is continuously increasing with little correlation to meteorological drought or recharge anomalies. Overexploitation of groundwater for irrigation seems to be the main cause of groundwater-level decline in the study area. Future changes in climate and anthropogenic factors pose additional uncertainties to the supply and management of (ground) water resources. In this study, groundwater level prediction uncertainty in climate change impact studies were quantified using 15 alternative conceptual models, 22 climate model runs for representative concentration pathways 4.5 and 8.5 (in total 44 climate model runs) and 3 groundwater abstraction scenarios. The results confirm that the groundwater level predictions are strongly dependent on the conceptual model structure. A significant decrease in groundwater levels is noticed for all groundwater abstraction scenarios. If the current groundwater abstraction trend continues, the groundwater level would decline about 5 to 6 times faster for the future period 2026-2047 compared to the baseline period (1985–2006). Even with a 30% lower groundwater abstraction, the mean monthly groundwater level would decrease by up to 14 m. The difference in groundwater abstraction scenarios was identified as the dominant source of uncertainty in groundwater level simulation in north-western Bangladesh. The uncertainty due to model conceptualization was also found to be quite significant and higher than that arising from the recharge scenario uncertainty, including the greenhouse gas scenario, climate model and stochastic climate uncertainty contributions. In the last part of this study, we present a flexible Integrated Bayesian Multi-model Uncertainty Estimation Framework (IBMUEF) to explicitly quantify the uncertainty originating from errors in the model conceptualization, the input data, the parameter values and measurement. In the proposed framework, the DiffeRential Evolution Adaptive Metropolis (DREAM) algorithm with updated likelihood function is combined with Bayesian Model Averaging (BMA). Groundwater recharge and groundwater abstraction multipliers are introduced to quantify the uncertainty of the spatially distributed input data of the groundwater model. A novel generalized formal likelihood function is also included for groundwater modelling based on a heteroscedastic error model to extend the applicability of our IBMUEF in situations where residual errors are heteroscedastic. Four alternative conceptual models representing different geological information have been used in this multi-model framework. The proposed framework has been applied for an overexploited aquifer in Bangladesh where groundwater pumping and recharge data are highly uncertain. The results of our study confirm that input uncertainty and model conceptualization have a considerable effect on the model predictions and parameter distributions. The use of groundwater recharge and abstraction multipliers leads to physically more realistic results and improve the reliability and accuracy of the model prediction. We demonstrated that our approach is able to simultaneously quantify the uncertainty originating from model input, parameter and measurement error of a fully distributed groundwater flow model. Our results also confirm that heteroscedasticity is present in the groundwater level error and its consideration along with model parameter uncertainty improves the accuracy of the uncertainty band. Additionally, the approach described serves as a new way to optimize the spatially distributed recharge and abstraction data under uncertain input conditions. It is found that a single conceptual model is unable to represent all the hydrogeologic processes of the system, hence consideration of alternatives conceptual models along with model input uncertainty is mandatory to have a reliable model prediction and confident parameters sets. We conclude that an explicit consideration of conceptual model structural uncertainty along with model input, parameter and measurement uncertainty using the IBMUEF framework improves the accuracy and reliability of the model prediction and related uncertainty bounds.
Author: Befekadu Taddesse Woldegiorgis
Supervisor(s): Prof. Dr. ir. Ann van Griensven, Prof. Dr. ir. Willy Bauwens
Date: 01 December 2017
In the era of high anthropogenic pressures on the quality of waters, a reliable water management strategy is essential. Mathematical models play an essential role in providing the information needed to take efficient measures to improve the water quality status of rivers. The current water quality simulators can be categorized into two broad categories. The first category consists of detailed water quality simulators that represent the physical reality in a detailed way. Water quality models developed using this types of simulators are characterized by long calculation times, and often require intensive input data. As a consequence of the excessive calculation time, they are not suitable for water quality management applications that require multiple simulations. In the second category are water quality simulators that represent the physical reality in a simplified way. Water quality models developed using this types of simulators have very short calculation times. They are therefore suitable for water quality applications that require multiple model runs and long term simulations. This category can further be divided into empirical simulators and conceptual simulators. The empirical simulators rely on input-output relations, and hence the equations do not have a physical meaning. Consequently, they are not suitable for scenario investigations. On the other hand, conceptual water quality simulators use simplified equations that lump some of the physical/chemical processes but that have a physically/chemical meaning. Many conceptual water quality simulators have been used for water quality modelling as alternatives or complements to detailed water quality simulators. However, they have a limited or no applicability for: 1) complex systems that have two-way interactions – such as river-flood plain systems, tidal rivers, and looped sewer systems; 2) the simulation of pollutants that travel attached to fine sediments – such as phosphate and trace metals. These limitations challenge the versatility of the current conceptual water quality simulators. Besides these limitations, the numerical solvers of most conceptual water quality simulators may provide unreliable results for some applications, especially when the computation time step is large.
With an aim of enhancing the applicability of conceptual river water quality simulators, a Conceptual Integrated Tool for Water quality Assessment (CIToWA) has been developed in the framework of this PhD research, hereby addressing the limitations described earlier. The tool is developed to: 1) present a robust solutions scheme 2) enable simulation of size-selective sediment-bound pollutant transport, and 3) enable simulation of complex stream networks using a conceptual approach. It is aims to maintain the computational speed of conceptual models but increases the versatility. In CIToWA, the river system is divided in reaches, conceptual elements that divide the channel longitudinally in to different parts. The discharges and the velocities in these reaches must be provided by external simulation tools. The reaches are further considered to act as Continuously Stirred Tanks Reactors (CSTR), whereby complete mixing is assumed within each reach. The water quality tool integrates several components, developed for specific purposes.
The first component is the simulator of dissolved oxygen, biochemical oxygen demand, the nitrogen compounds, the phosphorus compounds, and the algae.
A quasi-analytical solver is developed to simulate robust solutions of water quality problems, based on the CSTR principles and on ordinary first order reaction equations. The performance of the quasianalytical solver is evaluated by simulating different experiments and testing the robustness of its solution under extreme conditions. The results are also compared with the performance of SWAT (Soil and Water Assessment Tool) – a commonly used water quality simulator - and with high and low order numerical solvers that are commonly used in other conceptual water quality simulators. The comparison shows that the quasi-analytical solver simulates robust solutions -even for large computation time steps- in all the experiments and performs better than all the solvers used in the comparison. This increases the reliability of the solutions as well as the speed of computation of CIToWA.
A method is also developed to simulate complex stream networks and systems characterized by bidirectional interactions. The method uses the direction of the flow to enable pollutant transport in a reverse direction (from downstream to upstream), if needed.
A second component of CIToWA deals with the sediment and sediment-bound pollutants. The sediment transport module considers the sediment in suspension and the sediment on the river bed, taking settlement and resuspension into account. Hereby, the Particle Size Distributions of the sediment (PSD) are taken into account by considering probability density functions. The temporal and spatial evolution of the mean and standard deviation of the PSD –caused by settlement, resuspension, and emissions– are determined by using statistical functions of mixture distributions. The concepts of this novel sediment transport simulator are validated by simulating extensive sediment mobility data of laboratory and field experiments from literature, as well as by a case study on the River Zenne (see below).
CIToWA simulates the sediment-bound pollutant transport using a Langmuir adsorption model. The adsorption simulator integrates the quasi-analytical solver with the fine fraction of the sediment – the mass of the sediment to which pollutants adsorb. The latter module was also tested on a case study on the River Zenne.
An application of CIToWA in the complex river – canal system of the River Zenne (Belgium) illustrates its applicability for real cases. The system is characterized by bifurcations, such as overflows from the river to the canal and some of the reaches experience two-directional flows. The flow in the canal is characterized by high and frequent discontinuities as the flow is controlled by the sluice activities.
A comparison of the simulation results of CIToWA with those of a previous detailed water quality model of the River Zenne shows that CIToWA performs as well or better than the detailed model. The better performance of CIToWA is attributable to the fact that it runs fast (23 seconds to run the model with an hourly time step over the period 2007-2010) and hence automatic parameter optimization can be carried out – one of the key objectives of the research. The latter was indeed not the case for the detailed model, as the latter required 10 days of computation time to run the model for the pre-cited period.
Specifically with regard to the sediment-bound pollutant transport module, the fate of the inorganic phosphate in the River Zenne was considered. The comparison of the simulation results whereby adsorption to the sediments is considered with the simulation results whereby the adsorptiondesorption processes are not taken in to account, shows that the adsorption model outperforms the model without adsorption in the canal – where the sediment is dominated by clay. A comparison of the adsorption model of CIToWA with the detailed water quality model shows that the new tool performs better than the detailed model.
Given the fact that the new tool provides robust simulations in the complex stream network characterized by multiple branching and two-directional flows indicates that CIToWA can be applied to tidal rivers, river-flood plain systems and storm sewers systems characterized by two-directional flows. The plausible performance of CIToWA for the simulation of size selective sediment transport and the association of inorganic phosphate to the fine sediments shows that the new tool can be used to simulate the transport of sediment and pollutants that adsorb to them, such as trace metals. The application on the River Zenne shows that CIToWA has an enhanced applicability, as compared to current conceptual water quality models.
Author: Ting Tang
Supervisor(s): Prof. Dr. ir. Ann van Griensven, Prof. Dr. ir. Piet Seuntjens, Dr. Jan Bronders
Date: 25 October 2017
A wide range of pesticides and their metabolites are frequently detected in both engineered and natural urban water systems, including herbicides, insecticides, fungicides and biocides among others. Their occurrences in systems, such as urban drainage systems, wastewater treatment plant (WWTP) effluent, rivers and drinking water, have raised both ecological and human health concerns. To better control and manage urban pesticides use and occurrence, interest has grown in understanding pesticide transformation and transport over urban areas, mainly with experimental or monitoring approaches.
Modeling is a cost-effective approach to characterize the transformation and transport of urban pollutants from sources to the receiving waters, given sufficient understanding of the targeted system, reasonable simplifications, conceptualization and parameterization. To my knowledge, few tools are available to specifically predict pesticide transport, transformation and occurrences from urban areas and in urban water systems. This can be attributed mainly to the poorly-known pesticide behaviors on urban surfaces, particularly (semi-)impervious surfaces such as asphalt, concrete and facades. The often unknown use by urban residents further increases the difficulty in quantifying and characterizing pesticide loss at an urban catchment scale. The general objective of the PhD research is to develop an urban pesticide runoff model which accounts for the spatial heterogeneity of urban areas and adequately accounts for pesticide behaviors on urban (semi-)impervious surfaces.
Monitoring data from two field studies were used to characterize pesticide behaviors in urban environments: 1) loss and dynamics of pesticides, including herbicides and biocides, of both urban and agricultural origins in a typical mixed land-use catchment on the Swiss Plateau (Mönchaltorfer Aa), and 2) loss of glyphosate (a frequently-used herbicide) and AMPA (aminomethylphosphonic acid, main metabolite of glyphosate) from a typical Belgian residential area (Meerhout). The former case study revealed that the highly diverse urban pesticide sources and complex transport mechanisms make it rather difficult to generalize or simplify urban pesticide transport processes. It showed that urban runoff, discharged into surface water through separate storm sewers, combined sewer overflows or WWTP, can be the dominant force for outdoor pesticide transport into urban surface water depending on rainfall properties. The second case study (Meerhout) investigated specifically pesticide transport via urban runoff and subsequently a separate storm sewer system. Variations in event glyphosate load and loss rate were found predominantly determined by rainfall (amount and intensity). However, multivariate analyses suggest that rainfall and pesticide application cannot adequately explain the concentration variations. This promotes the need to account for additional important factors, such as the surface material and connectivity to the drain inlets of the application sites. Spatially-distributed hydrological models were therefore identified as beneficial for urban pesticide transport modeling to account for the heterogeneous spatial properties and altered urban hydrology.
Pesticide behavior on hard surfaces and building materials is often missing in catchment-scale pesticide models. In order to better understand the behavior, relevant experimental studies and the few available modeling approaches were reviewed. Incorporating the prior knowledge, a conceptual framework was developed to describe pesticide wash-off behaviors on (semi-)impervious hard surfaces, accounting for the physico-chemical interactions between the pesticide and the hard surfaces. Six mathematical representations (model structures) of the framework were implemented and tested using an existing experimental wash-off dataset to find the best compromise amongst the physiochemical validity of the equations, model structural complexity and parameter identifiability. After a collective consideration of model performance, parameter interactions and identifiability, one model structure (non-instantaneous partitioning with empirical exponential wash-off) was recommended as most suitable for urban pesticide runoff modeling, while two more structures were identified as good options given sufficient knowledge of the parameter ranges and that the underlying assumptions are met. The model structures were implemented for the development of an urban pesticide runoff model.
The catchment-scale hydrological model WetSpa-Python (Water and Energy Transfer between Soil, Plants and Atmosphere model, modular version in Python) was extended by adding seven pesticide-related modules: pesticide application, interception, land surface processes, depression, off-site routing/retention, mass balance and output module. The application module accounts for different levels of input data availability by allowing different input formats. This module allows a modeler to include pesticides from both non-point sources and point-sources (e.g., WWTPs) alike. The pesticide interception process accounts for the seasonal variations of pesticide interception capacity resulting from different plant growth stages and foliage coverage. The land surface processes include two subroutines, one for soil and another one for hard surfaces. The hard-surface subroutine takes the developed pesticide wash-off model mentioned above. The depression storage acts as a temporal sink, which delays the transport of on-site mobilized pesticide. The off-site routing/retention module takes care of pesticide routing to the catchment outlet, and the associated off-site retention and mobilization as a function of rainfall intensity.
The extended model, WetSpa-PST, was tested on the Meerhout case study for hydrological and pesticide (glyphosate and AMPA) simulations. WetSpa-PST allows for the quantification of pesticide and metabolite runoff from urban sources into surface waters. It is able to give spatial details on pesticide runoff, and therefore facilitates the identification of critical source areas. WetSpa-PST is the first catchment-scale distributed model that accounts for the physico-chemical interactions between pesticide and hard surfaces, using model parameters specific to the pesticide-surface combination. The strongest feature of WetSpa-PST is its flexibility, including easy coupling with additional processes or other models, flexible spatio-temporal resolution and mathematical representation for pesticide processes of interest. The WetSpa-PST model helps to improve quantitative understanding of the runoff and environmental occurrence of urban pesticides and to define appropriate pesticide management practices in urban areas (e.g., source control, buffer zones, best application timing, etc.).
In summary, this work advanced our understanding of urban pesticide transport and the underlying influencing factors. The developed hard-surface wash-off conceptual framework and WetSpa-PST pesticide runoff model help us to better estimate pesticide from urban sources. This is increasingly relevant in the context of global population growth and urbanization.
Author: Tadesse Alemayehu Abitew
Supervisor(s): Prof. Dr. ir. Ann van Griensven, Prof. Dr. ir. Willy Bauwens
Date: 25 October 2017
Water is an essential natural resource crucial for ecosystem integrity, poverty reduction and economic growth. As such, water is one of the central issues in the announced 2015 United Nations (UN) sustainable development Goals (SDGs), whereby goal 6 states “Ensuring availability and sustainable management of water and sanitation for all by 2030”. Therefore, characterizing the water balance at the river basin scale is important for effective water resources management and informed decision making. In this regard, distributed hydrological models are useful tools to analyse the water availability under changing conditions of land use and climate, among others. Unfortunately, the use of hydrologic models as a tool in water management is often challenged by the lack of reliable hydro-climatological data, particularly in developing countries. However, Earth observation and Land Surface Models provide global datasets that are relevant to hydrological studies. Yet, the reliability of such datasets and their use in hydrologic simulations at the watershed scale remain uncertain. The results from this thesis have demonstrated to what extent public domain climate data and remote sensing products are useful for hydrologic modelling in data scarce basins, and provided insight on how to effectively combine the available data to enhance simulations of evapotranspiration and streamflow using the Mara River basin (Kenya/Tanzania) as a case study. Initially, this research explores and evaluates the potential of Climate Forecast System Reanalysis (CFSR), Water and Global Change (WATCH) climate data to drive the semi-distributed Soil and Water Assessment Tool (SWAT) model. The evaluation of the climate datasets is conducted by comparing them with the available gauge rainfall, temperature and streamflow data in two ways: i) for their potential to calculate reference evapotranspiration using the Penman-Monteith and Hargreaves methods, and ii) as source of rainfall data. The former evaluation also highlights the effects of the method selection for reference evapotranspiration estimation on the simulated water balance components. Our results show that the daily reference evapotranspiration estimates using CFSR and WATCH climate data are sensitive to the reference evapotranspiration estimation method. For the study area, the less data intensive Hargreaves method gives better estimates when using global datasets. Furthermore, unlike streamflow, the simulated actual evapotranspiration (ET) and groundwater losses show significant sensitivity towards the selected reference evapotranspiration calculation method. Here, in addition to the quality of the climate data, the structure of the SWAT model to simulate ET contributes to the observed differences (up to 354 mm yr-1) in the reference and actual plant transpiration. We also found that the daily WATCH and CFSR rainfall are poorly correlated with gauge rainfall (r <0.3) data-preserve-html-node="true" and hence their application for daily hydrologic application is limited. However, these rainfall estimates correlated well (r up to 0.95) at the monthly scale and therefore, their simulated water balance components show a good agreement with gauge-based water balance estimates. This suggests that CFSR and WATCH rainfall can be used as a surrogate for local observations for monthly and seasonal water budget analysis in the region. Because ET is the major outgoing flux in a basin water balance, we also investigate the spatial distribution of ET from a hydrological modelling perspective using the SWAT model and from the surface energy balance perspective using the Operational Simplified Surface Energy Balance (SSEBop) algorithm. Remotely sensed land surface temperature and Global Land Data Assimilation System (GLDAS) climate data are combined to map ET at 1 km/8-day scale using the SSEBop algorithm with modified parameterization. The basin is characterized by considerable variation in land cover, soil and climate and thus our ET modelling results reflects this well. We note a higher variability up to 52% Inter Quartile Range (IQR) of the 8-day median ET for agriculture and herbaceous covers while wetland and evergreen forest covers show a variability of 24% and 29% of the median ET, respectively. Evaluation of the ET estimates using 9-years monthly, gridded global flux tower-based ET data reveals that the estimated ET is able to explain 64% of the variance of the observation-based ET, indicating a good agreement in the ET estimates. The mean annual basin wide water use (i.e. ET) is about 817(±32) mm yr-1, which is 66% of the basin rainfall. The other focus of this thesis research is to improve ET simulation in the SWAT model, which is one of the most widely applied comprehensive models to simulate water yield, sediment yield, and nutrient and pesticide loadings at a river basin scale. Despite the SWAT model popularity, the tool has limitations in simulating vegetation growth cycles of trees and of perennials and annuals in tropical conditions due to temperature dependent dormancy. We therefore improve the vegetation growth module of the SWAT model for tropical ecosystems using the soil moisture index, the major vegetation growth controlling factor in the tropics. The modified SWAT vegetation growth module dynamically triggers the yearly growth cycle based on the soil moisture index. The modification reduces the inconsistencies in the simulated evapotranspiration by the SWAT model from about 12% to less than 2%, suggesting a substantial improvement in the vegetation growth model structure. Furthermore, the simulated 8-day LAI using a calibrated SWAT model with the modified vegetation growth module, correlated well (up to 0.93) with the remote sensing LAI for forest, savanna grassland and shrubland. In general, the improvement on the vegetation growth cycles is reflected with seasonal and spatially consistent LAI and ET simulations, which will be helpful to evaluate the effects of different water management interventions in the tropical region. The last component of this thesis research is to develop a methodological framework for hydrologic models calibration and evaluation in poorly gauged basins. In this regard, a multi-objective framework that combines information about a basin’s hydrologic response characteristics (i.e. streamflow signatures) with the remote sensing-based ET is developed. In a poorly gauged basin with better past monitoring practices, historical streamflow observation can provide useful information on the rainfall-runoff response behaviour. Therefore, this thesis proposes the combination of the streamflow signatures (historical) from different parts of the flow duration curve (FDCs) with the spatially distributed remote sensing-based ET. This framework is implemented to calibrate and evaluate the SWAT model for the Mara Basin. The SWAT model calibrated using the streamflow signatures and the remote sensing-based ET reproduce the observed streamflow signature measures for the high flow regime, the mid-flow segment slope and the low flow regime with biases lower than 4% for the headwater region well. Likewise, the model also shows similar performance for the part of the basin dominated by semi-arid climate with the exception of the low flow regime which shows a relatively higher bias (14%). With respective to ET simulation, the calibrated model reduces the biases in the baseline model by 9.2% at the basin scale. Despite the reduction in biases compared to the baseline model, the model has a positive bias in the simulated ET. This indicates the need for further adjustments on the model parameters. Overall, the rainfall partitioning in the simulated water balance components reflect the climatic and landscape characteristics across several watersheds in the basin. In summary, this PhD research has improved the ET estimates in the Mara Basin which is valuable in understanding the basin’s hydrology. Also, the methodological frameworks developed in this research can be applied elsewhere to support water resources management and planning using spatially distributed information. Yet, a proper selection of data is important, which should be done considering the full modelling framework.
Author: Vicente Iñiguez
Supervisor(s): Prof. Dr. ir. Willy Bauwens, Prof. Dr. ir. Guido Wyseure, Prof. Dr. ir. Felipe Cisneros
Purpose The overall objective of the research is to show how a monitoring set-up for experimental research catchments can be implemented and used to establish the foundations of a hydrological observatory in the southern Ecuadorian Andean region. Such observatory will allow researchers to answer different scientific questions about the hydrological processes in high altitude Andean ecosystems –which are the main water suppliers in the Andean region–. Thus, the first scientific question addressed here is: what is theeffect of vegetation on evapotranspiration and the hydrological regime in tropical Andean mountainous catchment?
Author: Nhu Viet Ha
Supervisor(s):Prof. Dr. Okke Batelaan, Prof. Dr. Marijke Huysmans, Prof. Dr. Pham Ngoc Ho
Land subsidence is one of the most harmful urban geomorphological hazards. It is affected by many factors including rising demands of groundwater supply and urbanization. In Hanoi city (Vietnam), land subsidence occurs at an average rate between 0.6 and 38.2 mmy-1. This causes serious infrastructural damage in the urban area.
In this work, a multi-step methodology for land subsidence simulation and prediction is proposed. It consists of the following steps: (1) construction of a 3D geological-geotechnical model (GGEO), (2) construction of a 3D transient groundwater flow model (GFLOW) and (3) land subsidence assessment using a 3D mechanical model and machine learning models.
The 3D geological-geotechnical model is based on a very extensive database of geological and geotechnical data (1386 drilling records) that was constructed within the framework of this PhD. A geostatistical approach was used to interpolate between the observed depths of all geological layers. The 3D GGEO model simulated the subsurface elevation with a good accuracy (R = 0.81 - -0.94, RMSE = -0.01 - -0.08 m, and ME = 0.59 - -3.92 m).
The 3D transient GFLOW model simulates and predicts groundwater heads up to 2030. It was constructed in Visual MODFLOW based on the subsurface geometry of the 3D GGEO model and other groundwater-related information. The 3D GFLOW model reproduced groundwater heads for the period of 1996-2014 with a reasonably accuracy (R = 0.86 - 0.95, RMSE = 1.76 - 2.5 m, and ME = -0.48 - -0.04 m).
Two types of approaches for detailed assessment and prediction of spatio-temporal land subsidence are applied: a 3D mechanical (MEC) and a machine learning (ML) model. In the 3D MEC model, a Python code was developed using consolidation theory to directly compute time-dependent land subsidence, based on the geotechnical database and simulated heads obtained from the groundwater flow model. As novel techniques in land subsidence prediction, two machine learning models - Support Vector (SVR) and Gaussian Process Regression (GPR) - were applied. By training the algorithms using the WEKA Data Mining tool, spatio-temporal land subsidence was simulated and predicted. These models were able to take into account eight conditional-causative factors, which is more than the 3D mechanical model. Both machine learning models simulate land subsidence with a good accuracy. However, the SVR model (with R = 0.91, RMSE = 2.13 mmy-1, and Values Account For VAF = 82%) slightly outperforms the GPR model (R = 0.87, RMSE = 2.70 mmy-1, and VAF = 76.4%).
All land subsidence models indicate that the most influential factors affecting land subsiding are the thickness of the organic soils and the decline in heads. High subsiding zones (in the west and south of the study area) are related to thick organic soils and strongly declining heads.
The total land subsidence from 1996 to 2030 was predicted by the mechanical model to be between zero and 160 cm. The average land subsidence rates in the period of 2011-2030 were predicted as 0.2 - 29.3 mmy-1 by the SVR model and 1.4 - 28.1 mmy-1 by the GPR model.
On the basis of these results, recommendations about land subsidence in the study area are formulated for engineers, city managers, urban planners, administrators, and local inhabitants.
Author: Khodayar Abdollahi
Supervisor(s): Prof. Dr. Marijke Huysmans, Prof. Dr. Okke Batelaan
The need for improvement of hydrological models often leads to regular program modification, and thus development of hydrological models is a dynamic process. Furthermore, for the case of distributed models due to the spatial and temporal variations of the hydrologic processes, development of this kind of hydrological models remained a difficult task over the last decades. Consequently, application of the Geographic Information Systems (GIS) has become a useful and popular tool in hydrology. On the other hand, previous studies emphasize that hydrological modellers should not only use GIS, but also take advantage of other advances in computing technology. This issue generates a demand for development of hydrological frameworks to conceptualize hydrological processes and more specifically targets hydrological modelling. This study presents the development of an easily reusable hydrological modelling framework ‘Hydrologic and Hydraulic Programming Library’ (H2PL). The developed open-source library is a collection of computational hydrological functions and complex spatial calculations that operate under Windows and .NET frameworks. H2PL offers advantages in the case that a model has to be extended independently, while it requires a high level programming with spatial data support. The functionality of the tools allows users to develop their own models using different scripting languages and a graphical interface, while they can be compiled as standalone executable files. In order to encourage the use by users with minimum programming knowledge, the programming structure of the library has been kept simple. The library was applied for three different cases, the first case was integration with SWAT for coupling purposes; the second application was for the improvement of an existing hydrological model for recharge estimation and distributed water balance calculation; while the last application of the library was the development of a new analogue model for flash-flood prediction.
The second implementation of the library was the updated version of a flexible raster GIS extension called WetSpass (Batelaan and De Smedt, 2001) to simulate water balance elements, such as surface runoff, evapotranspiration and groundwater recharge. Since the model was designed for long-term hydrological modelling, it is well-capable to simulate the dynamics of hydrological processes, but the hydrological parameters were designed for conditions in temperate regions of Europe. The new standalone edition of WetSpass presents a generalized version of the model with monthly time step and improved flexibility.
The last implementation of H2PL was to develop a new hydrological model (RASAM) with emphasis on flash flood processes by applying an electrical flash analogy. The model was based on simplification of the water balance processes into a series resistance-capacitance circuit. Parameters for the circuit were estimated by the Weibull distribution. Application and evaluation of the model was carried out for four flash flood events in the Tarqui Watershed (Equador), and also for a theoretical benchmarking case study for runoff called “Open-Book catchment (the tilted V catchment)”. The applications revealed that the library provides a highly efficient and effective framework for spatiotemporal hydrological modelling.
85 Development of a flexible process-based spatially-distributed hydrological model for urban catchments
Author: Elga Salvadore
Supervisor(s): Prof. Dr. Okke Batelaan
Author: Tomasz Berezowski
Supervisor(s): Prof. Dr. Okke Batelaan, Prof. Dr. Jaroslaw Chormański
Author: Zainab Zomlot
Supervisor(s): Prof. Dr. Okke Batelaan, Prof. Dr. Marijke Huysmans
Author: Pablo Ismael Guzmán Cárdenas
Supervisor(s): Prof. Dr. Okke Batelaan, Prof. Dr. Marijke Huysmans, Prof. Dr. Guido Wyseure
Author: Phan Thi Ngoc Diep
Supervisor(s): Prof. Dr. Okke Batelaan, Prof. Dr. Marijke Huysmans, Prof. Dr. Pham Ngoc Ho
Author: Ine Vandecasteele
Supervisor(s): Prof. Dr. Okke Batelaan, Dr. Luc Feyen
Author: Fabricio Fiengo Perez
Supervisor(s): Prof. Willy Bauwens, Prof. Mark Elskens, Dr. Lieve Sweeck
Author: Nicolas Ghilain
Supervisor(s): Prof. Dr. Okke Batelaan
Author: Fidelis Ndambuki Kilonzo
Supervisor(s): Prof. Willy Bauwens, Prof. Piet N. L. Lens, Prof. Ann Van Griensven, Prof. Joy Obando
Author: Eva Ampe
Supervisor(s): Prof. dr. Okke Batelaan, Prof. dr. Ludwig Triest
Inland waters are valuable ecosystems that provide multiple services such as drinking or irrigation water, fisheries, recreation, transportation and hydropower. Unfortunately, these ecosystems are often under stress. Due to an increase in nutrients from agriculture, forestry and urban development many inland waters are prone to nuisance or harmful algal blooms (HABs). Such blooms, especially cyanobacteria, can be potentially harmful for humans and the local fauna. Careful and frequent monitoring of inland water quality is therefore crucial.
Monitoring of water quality based on point sampling is often limited by the variable spatial distribution of the algae, dissolved material and the suspended sediment in the system. Therefore, remote sensing is useful in providing spatial and temporal information. It can be used to estimate the turbidity or clarity of the water. Three main constituents influence the water clarity: 1) the algae cells or phytoplankton, 2) the non-algal particles of the suspended matter, and 3) the colored fraction of the dissolved organic material. These optically active constituents are important because they influence light penetration through the water column by scattering and absorption. They can be used to derive other water properties or constituents.
The overall objectives of this thesis is to provide new methods for high spectral resolution remote sensing of inland waters. Remote sensing in inland waters is challenging because the optically active constituents may vary independently and have a combined and interacting influence on the signal. Additional confounding factors are, e.g. varying optical properties, spectral influence of neighbouring vegetation, mixed pixel, and bottom effects. Spectral features can also be influenced by noise (e.g. instrument signal-to-noise ratio, sensor saturation).
This thesis starts by analyzing the potential to evaluate phytoplankton size classes from water absorption measurements. Then it focuses on water high spectral resolution reflectance data. Since this signal is influenced by the confounding factors we propose to decompose it using a multiscale wavelet approach. Wavelets can analyze the spectral signal at various feature scales. The thesis applied wavelets in two ways: 1) in a semiempirical approach to estimate chlorophyll-a (CHL), and 2) in a semi-analytical approach where wavelets are used in a bio-optical inversion method.
The phytoplankton size distribution plays an important role in ecosystem modeling since changes in size over space and time significantly influence sinking rates, and the structure and functioning of aquatic ecosystems. In oceans, it has often been modeled using abundance-based size class models. Adaptation of these models for inland waters can be troublesome. To investigate the feasibility of estimating size class from remote sensing in inland waters an experiment was performed on freshwater algae cultures. We evaluated the possibility of estimating the phytoplankton size distribution of the algae with the use of a laser particle sizer. We observed a different phytoplankton size class population structure compared to ocean waters. Yet, it was difficult to perform reliable measurements of the phytoplankton size class distribution. This disabled the application of the traditional abundance-based size class models. However, before drawing general conclusions more investigation is warranted and we recommend that more data is needed.
Semi-empirical techniques are based on a statistical relationship between the water quality parameter and the reflectance spectrum. However, traditional approaches cannot cope easily with changes and shifts in reflectance features caused by the confounding factors. This thesis used continuous wavelet analysis to detect CHL features at various wavelengths and wavelet scales. Using the wavelet decomposition, we built a 2D correlation scalogram between in situ reflectance spectra and the CHL concentration. The wavelet scalogram provides flexibility by identifying informative wavelet regions rather than strict spectral band combinations. It enhanced the spectral information extraction, because wavelets analyze the signal at different scales and synthesize information across bands. Also, we hypothesize that the wavelet features are less sensitive to noise and confounding factors.
Semi-analytical methods for inland waters are based on the inversion of a biooptical model relating the optically active constituents to the reflectance signal. Such a bio-optical model uses in situ measured absorption and scattering properties of the constituents in the water body to model the reflectance signal. However, noise and confounding factors can impede the constituent retrieval. This thesis proposes WaveIN, a wavelet enhanced inversion method, a novel approach specifically designed for inland waters. It performs the model inversion on the wavelet transformed spectra. This thesis applied WaveIN to both simulated and airborne remote sensing data. In general WaveIN improved the constituent retrieval of the three optically active constituents. Wavelets are less sensitive to a vertical shift in the reflectance spectra. Also, WaveIN can avoid contaminated wavelet scales by selecting alternative scales. Additionally, the wavelet decomposition enhances the constituent’s spectral features at specific wavelet scales. However,WaveIN is sensitive to wavelet edge effects, which are an artifact of the wavelet calculation, and it could not resolve the influence of local and seasonal variation in the water body’s specific absorption and scattering properties. Also, WaveIN still requires in situ data for the wavelet scale selection. Future research should therefore improve the wavelet scale selection and move towards a weighted average of informative wavelet scales.
This thesis provides new methods specifically designed for high spectral resolution data of inland water. These methods make optimal usage of the spectra’s contiguous information content. They provide novel tools for water quality monitoring especially in the case of suboptimal data, where the spectral features are influenced by confounding factors. Other than to air-/spaceborne images the methods can readily be applied to point or transect spectral measurements acquired using hand-held spectrometers or spectrometers mounted on drones or ferry-boats.
Author: Olkeba Tolessa Leta
Supervisor(s): Prof. dr. ir. Willy Bauwens
Author: Narayan Kumar Shrestha
Supervisor(s): Prof. dr. ir. Willy Bauwens
The use of modelling tools to represent the reality in a simplified form has been practiced for a long time. Different modelling tools in various forms and implementations exist. While these modelling tools are generally meant for a single purpose, there is an increased need for more holistic solutions to the complex and interrelated real world problems. For a holistic analysis of the water quality management problems, integrated modelling is also imperative in the framework of regulations such as the EU Water Framework Directive (WFD).
Practitioners have adopted different approaches for model integration. While in the hydrological domain a tightly integrated approach was traditionally applied, the popularity of loosely integrated frameworks is growing in recent times. This study used one of the popular loosely integrated frameworks, the Open Modelling Interface (OpenMI), to integrate models for the simulation of different water quantity and quality processes: the hydrology in the river basin, the hydraulics in the river and in the sewers, erosion and sediment transport, the Carbon-Nitrogen-Phosphorus (C-N-P) cycle, the transport of trace metals and the transport and decay bacteria. Existing simulators were adapted to OpenMI standards: for the modelling of upstream rural catchment processes (the Soil and Water Assessment Tool - SWAT); for the modelling of rivers, canals and sewer systems (the Storm Water Management Model - SWMM). New, OpenMI compliant, codes have been developed for sediment, the C-N-P cycle, faecal bacteria, trace metals and the water temperature.
These simulators were then applied on the river Zenne (Belgium) in the scope of a multidisciplinary project called Good Ecological Status of river Zenne (GESZ), funded by Innoviris. The integrated models were calibrated and validated at different locations, using data obtained through different agencies and data measured during GESZ sampling campaigns. The results show that the integrated model simulates the different water quality parameters with reasonable accuracy.
It is also shown that, while the Zenne is relatively clean upstream of Brussels, the situation worsens due to (still) untreated effluents just upstream of Brussels and due to polluted effluents of the Brussels South Waste Water Treatment Plant (WWTP). On the other hand, the impact of the effluents of the WWTP Brussels North was found to be positive for most of the water quality indicators. Another important problem is related to the sporadic combined sewer overflows (CSOs) from the sewer system of Brussels that lead the river water to near anoxic conditions at times. From this, it is clear that the ecological status of the river Zenne in and downstream of Brussels does not meet the EU WFD standards.
Globally, it is concluded that this research have successfully implemented the OpenMI to integrate different models. It was found that the OpenMI based integration is very useful, as it provides the needed flexibility to choose/develop appropriate model components to better represent catchment and river processes. It is believed that such OpenMI integration can be very useful as a decision support tool for integrated river basin management. A big advantage of OpenMI, as the author sees it, is that it allows to check the accuracy of each component of an integrated model and to more easily locate the possible sources of errors in the model. However, the use of such an ‘interface’ obviously causes some calculation time due to the run-time communication between the model components. However, this overhead was found to be minimal. Solutions to optimize the efficiency of the computer hardware and software, such as grid computing or parallel computing, thus need to be explored.
It is believed that with the availability of such integrated models, the overall dynamics of the river with respect to water quantity and quality aspects would be better understood and such information could translate to improving the present condition of the river.
74 Variability of Groundwater Flow as a Factor of Mire Ecohydrological Evolution-Example of Czerwone Bagno
Author: Mateusz Grygoruk
Supervisor(s): Prof. dr. Okke Batelaan, Prof. Dr. ir. Tomasz Okruszko
In this thesis the analysis of ecohydrological processes inducing the evolution of Czerwone Bagno mire is presented. In reference to the 4-year set of data on groundwater levels the variable hydrological dynamics of selected plant alliances of the study area was revealed. A modelling study, in which a quasi 3D groundwater flow model was applied, assessed the directions of groundwater flow within the mire as well as water balance of selected plant alliances in both contemporary and hypothetical historical conditions. Analysis of shallow groundwater level time series is used to assess the actual evapotranspiration of an encroaching mire meadow. This thesis revealed that the construction of canals has entailed the switch in groundwater flow conditions and enhanced the differentiation of mire’s trophic conditions by enhancement of rainwater accumulation in the superficial layers of peat soils in the core zone of the area of research. It is concluded, that the appropriate management of mire meadows can entail the increase of water resources through the limitation of evapotranspiration.
73 Improving Spatiotemporal Urban Runoff Estimation Using Earth Observation Based Approaches
Author: Boud Verbeiren
Supervisor(s): Prof. dr. Okke Batelaan, Prof. dr. Frank Canters
Images of streets, houses, shopping malls, etc. being flooded are a regular recurring phenomenon nowadays. Yearly several European cities are confronted with flood events of varying magnitude. Although the origin of these events often lies in a combination of factors (e.g. extreme rainfall in short time, saturated soils, hydraulic obstructions in rivers, land cover/land-use types and dynamics, climate change, etc.), the specific characteristics of the urban landscape most certainly play a role. Given the huge economic interests and sometimes even lives at stake, it is clear that urban floods at present and for the future are taking a prominent place at global scale.
Therefore a thorough characterization of urban land use and estimation of related hydrological parameters for spatially-distributed modeling is of utmost importance. Hydrological models are indispensable to describe and study hydrological conditions in (urban) catchments. Fully distributed hydrological modeling enable spatial and temporal analysis of various water balance components but needs spatially distributed input data, which are often difficult and/or expensive to collect. In this research Earth observation (EO) is put forward as an important source for providing detailed spatiotemporal information on the characteristics of the Earth surface, useful for hydrological parameterization and modeling.
This study focuses in particular on the impact of urbanization on hydrology in urban river catchments and more specifically aims at improving understanding, parameterization and modeling of runoff in a complex and heterogeneous urban context. Therefore an EO supported hydrological modeling strategy is proposed, aiming to overcome some of the limitations with respect to currently accepted methods for spatiotemporal assessment and modeling of hydrological processes with a focus on urban river catchments.
The EO supported hydrological modeling strategy integrates detailed information on urban land cover, sealed surfaces, vegetation characteristics and dynamics into the rainfall-runoff WetSpa model. The results show the added value of EO derived information for the assessment of urban and vegetation dynamics on hydrology: EO enables area-covering and distributed characterization of the urban landscape. In addition, EO yields catchment- and period-specific information regarding urban land use and vegetation characteristics. EO offers the possibility to increase the temporal resolution for an improved assessment of urban dynamics by using a time series of EO data. The study also demonstrates the benefit of using a change trajectory analysis in an urban context to assure the consistency of the EO time series of urban land use and sealed surface proportions. The EO supported hydrological modeling strategy enables the assessment of urban growth and consequently simulation of the impact on hydrology. The simulations with sub-pixel sealed surface proportions yield considerable, especially spatial, differences in discharge and runoff, confirming the importance of sealed surface cover as a key factor in describing the hydrological dynamics in urban catchments. Vegetation related parameterization (interception storage capacity and root depth) reveals that the impact on the total discharge at the catchment outlet remains rather limited, although there is a considerable effect on the spatial output of smaller components of the water balance. The integration of EO based distributed sub-pixel sealed surface fractions into WetSpa reveals also considerable differences in simulated evapotranspiration (ET) rates and increases as well the spatial detail and variability of evapotranspiration simulation for the urban classes. At this point no decisive answer could be given whether surface energy balance ET estimates (medium and coarse resolution) can be used to validate rainfall-runoff simulations with WetSpa.
72 Calibration of Hydraulic Conductivities in Groundwater Flow Models using the Double Constraint Method and the Kalman Filter
Author: Mustafa Ahmed El-Rawy Ibrahim
Supervisor(s): Prof. dr. Okke Batelaan, Dr. ir. Wouter Zijl
To model groundwater flow, two types of problems must be solved: (i) the forward problem and (ii) the inverse problem (calibration). Forward modeling is based on prior estimates of the model parameters. For instance, in MODFLOW the hydraulic conductivities in the model’s grid blocks have to be specified. Calibration means choosing conductivities in such a way that the heads and fluxes calculated by the model honor the measured heads and fluxes. Modeling in which not only heads and fluxes, but also parameters are determined, is generally called inverse modeling.
Estimation of hydraulic conductivities is one of the main challenges in groundwater modeling. This thesis considers two kinds of conductivity identification: inverse modeling and data assimilation. Inverse modeling is based on a number of spatially distributed observation wells in which the piezometric heads are measured at the same time. Data assimilation is based on time series of measured piezometric heads.
The Double Constraint Method (DCM) was used as a direct inversion technique for estimating hydraulic conductivities for groundwater flow modeling. The DCM is a relatively simple, yet very instructive approach for inverse modeling of hydraulic conductivities. As a first task of the thesis, the DCM is combined with MODFLOW to calibrate a large scale groundwater flow problem (the Kleine Nete catchment). This was done by applying Darcy’s Law to find the hydraulic conductivities from the flux and head boundary conditions. The DCM-MODFLOW combination was developed and tested for the Kleine Nete catchment and the Schietveld area (Belgium). With a priori grid block hydraulic conductivity as input, two runs were made. From the first run, which was based on the known flux boundary conditions, the fluxes were determined. From the second run, which was based on the known head boundary conditions, the head gradients are determined. From these two so-called constraining runs the hydraulic conductivities in the grid blocks are updated using Darcy’s law. A specified anisotropy was obtained by application of a “mixing rule” followed by iterations. Convergence rates were tested for different update and mixing rules (Exponential and Wexler’s rules). Tests show that only a few or even no iterations are necessary to obtain an acceptable accuracy, which makes the calibration procedure numerically efficient. For the Schietveld groundwater model, the sensitivity analysis was caried out based on the hydraulic conductiviy and drains conductance parameters. This analysis showed that hydraulic conductivity is the most sensitive parameter. Model calibration was done using two different techniques (the DCM and UCODE-2005). The measured and modeled results were compared and found to be in good agreement. The results showed that the DCM performs somewhat better than the UCODE. In addition, The DCM gave good results after iteration 4 or 5, while in the case of the UCODE the paramter estimation did not yield acceptable results before 10 iterations (each iteration being very computer time consuming). This result corresponds with theoretical considerations stating that DCM-calibration yields already relevant results after the first stroke or a few iterations, while a gradient method (upon which UCODE is based) has better convergence properties. The two approaches upon which DCM and UCODE are based are different principles. Even when then the modeler starts with grid block conductivities that are homogeneous in a number of zones or layers, the DCM calculates hydraulic conductivities that differ gradually from grid block to grid block within a zone or layer. On the other hand, UCODE is essentially based on zonation. Therefore, it preserves the initially chosen zonation; it does not determine the conductivities as distributed values within a zone. The modeler could increase the number of zones, but the computation time will inhibit the use of too many zones (considering each grid block as a zone is, therefore, impossible). Although, the possibility of zonation may be considered as an advantage of UCODE, in the section on upscaling (section 6) we will propose a complement to DCM that can be used for zonation in a computationally cheaper way.
Data assimilation techniques determine the conductivity pattern from time series of observations, for instance from the time-dependent flow rates and heads. However, also other data related to the flow rates and heads (among which conductivities determined by DCM) were considered as observed data. In this research we considered the Kalman Filter as data assimilation technique. To improve the characterization of hydraulic conductivities, the inverse modeling (DCM) was combined with data assimilation (Kalman Filter). The DCM was applied at different times, for different flow rates and heads. As a consequence, the spatial image of hydraulic conductivities obtained from the DCM may differ for each time. The variations in time of the thus observed hydraulic conductivities reflect the measurement uncertainty in the measured heads and flow rates, as well as the perception uncertainty. A Kalman Filter was used to obtain an estimate of the time-independent “true” hydraulic conductivity, as well as its estimation uncertainty. In this approach only the Kalman Filter’s variances, not the covariances, play a role, which reduces the computational requirements considerably. In the Kleine Nete study we found a considerable decrease (approximately 77%) in the uncertainty of the estimated conductivities compared to the uncertainty obtained from “observations” (i.e., inverse model runs) using the DCM. In addition, the approach could distinguish the regions in which the measured heads influence the calibration from the regions where heads obtained from the monitoring wells have no influence on the calibration.
Finally, in a number of practical cases we have to decrease the number of grid blocks in the MODFLOW model. Models with less grid blocks are simpler to apply, especially when the model has to run many times, as is the case when it is used to assimilate time-dependent data. Grid block reduction requires upscaling, resulting in equal hydraulic conductivities for a cluster of neighboring grid blocks. Such upscaling was done by the DCM as a complement to MODFLOW and applied on the Kleine Nete catchment. As results we identified hydraulic conductivities for the coarse scale and, at the same time, we decreased the number of grid blocks. Different upscaled models were ranked based on estimating the Akaike Information Criterion (AIC), the Bayesian Information Criterion (BIC), and Hannan and Quinn’s criterion (HQ). The results show that the coarser models perform sufficiently well.
71 Simulating Soil Moisture and Climate Change Impacts in a Watershed Through Application of the Distributed Hydrological WetSpa Model
Author: Mohsen Tavakoli
Supervisor(s): Prof. dr. Ir. Florimond De Smedt
70 Study and Development of a Distributed Hydrologic Model, WetSPA, Applied to the DMIP2 Basins in Oklahoma USA
Author: Alireza Safari
Supervisor(s): Prof. dr. Ir. Florimond De Smedt
69 Landslide Susceptibility Mapping in the Himalaya: Case Studies from Nepal
Author: Prabin Kayastha
Supervisor(s): Prof. dr. Ir. Florimond De Smedt, Prof. dr. Megh Raj Dhital
68 Modeling of Hydrological Processes in the Geba river basin, Northern Ethiopia
Author: Haddush Goitom Aforki
Supervisor(s): Prof. dr. Ir. Florimond De Smedt
67 Detailed Groundwater Recharge Estimation for Urban Environments: a Remote Sensing Supported Procedure for Groundwater Risk Management
Author: Juliette Dujardin
Supervisor(s): Prof. dr. Okke Batelaan, Prof. dr. Ir. Florimond De Smedt
66 Quantifying groundwater-surface water interaction by using heat as a natural tracer
Author: Christian Anibas
Supervisor(s): Prof. dr. Okke Batelaan, Prof. dr. Tomasz Okruszko
65 Hydrological Impact Analysis of Land-use and Climate Change
Author: Jef Dams
Supervisor(s): Prof. dr. Okke Batelaan, Prof. dr. Frank Canters
64 Methodology for Estimation of the Change in the Reserve of the Upper Berg Catchment, Western Cape, South Africa
Author: Maher Albhaisi
Supervisor(s): Prof. dr. Okke Batelaan
63 Sensitivity and Uncertainty Analysis in View of the Parameter Estimation of a SWAT Model of the River Kleine Nete, Belgium
Author: Jiri Nossent
Supervisor(s): Prof. dr. ir. Willy Bauwens
Due to the improved knowledge of the physical processes occurring in a watershed, the increased processing capacity of computers and the development of GIS and remote sensing, hydrological models have become very complex during the last decades. In general, these complex environmental models are controlled by a high number of parameters and are therefore often referred to as over-parameterized models. Hence, accurately estimating the values of all these parameters based on a single data set of historical measurements is almost impossible. Therefore, this thesis presents the application of sensitivity analysis, uncertainty analysis and parameter estimation techniques on an environmental model, in order to improve the understanding of the model parameters and their influence on the model output, allow a selection of identifiable parameters, and, moreover, assess the potential of using additional (water quality) data sets for estimating the values of the parameters.
For the evaluation of the applied techniques, a model of the Kleine Nete catchment is built with the Soil and Water Assessment Tool (SWAT), a complex, over-parameterized environmental simulator that can be applied for an integrated modelling of water quantity and quality variables. The model has a daily time step and the simulations are performed for the period 1997-2007.
A first sensitivity analysis technique that is applied is the Latin- Hypercube – One-factor-At-a-Time (LH-OAT) analysis, a screening method that is incorporated in the SWAT simulator. The method appears to be suitable to perform a sensitivity analysis of the simulations of the different output variables of a SWAT model, but the number of intervals used for the Latin-Hypercube sampling should be sufficiently high to obtain converged parameter rankings. Although the method sometimes identifies important parameters as non-influential, the application can enhance the understanding of the model, e.g. on the use of water quality input data.
In order to find an alternative for the LH-OAT technique, the variance based Sobol’ sensitivity analysis is also applied for the simulations with the SWAT model. When computation time requirements are not an issue, the powerful and robust Sobol’ method has the properties of an “ideal” sensitivity analysis technique. However, the results show that the Sobol’ method can be successfully applied even for a limited base sample size. In particular, the convergence of the parameter ranking for total sensitivity effects is relatively fast for most variables, which is promising for factor fixing purposes. In general, it is observed that the use of the first order, second order and total sensitivity indices allows to identify model processes, parameter values and parameter interaction effects. Confidence intervals for the sensitivity indices are inferred by applying bootstrapping.
Furthermore, the application of graphical sensitivity analysis techniques (local scatterplots, global scatterplots, SA repeatability test) shows that these methods are particularly suited to improve the understanding of the SWAT parameters and their influence on the model, and can therefore assist in the interpretation of the results of the other (numerical) sensitivity analysis methods. Nonetheless, they can also be employed as standalone sensitivity analyses.
A last method that is applied for the assessment of the most influential parameters is the singular value decomposition (SVD) of the Jacobian matrix, which is a numerical technique that can be applied to determine the identifiability of the SWAT parameters. This local method requires only a limited number of model evaluations to obtain the number of identifiable parameters and the corresponding right singular vectors. Unfortunately, this result is only valid in the nominal point. The results show that in the selected nominal point, more than 90% of the model output variability is captured by the first two singular values for most of the configurations. It is also shown that the studied SWAT parameter hyperspace approximates an orthogonal base.
In general, the results of the various sensitivity and identifiability analyses show that both flow and water quality simulations with the considered model are mainly influenced by variations of water quantity parameters, and more in particular by the curve number value. Sediment and nutrients are transported to the river by different flow components. The sensitivity analyses are able to clearly identify this physical background and the corresponding model processes. As a consequence, it may be concluded that water quality simulations have the potential to provide additional input for the optimization of the flow parameters. Overall, sediment and nutrient related parameters only have a significant impact in particular configurations. Surprisingly however, nutrient related parameters do also have some influence on flow predictions.
The Shuffled Complex Evolution (SCE-UA) method and the Differential Evolution Adaptive Metropolis (DREAM and DREAM(ZS)) algorithm are applied for the single objective parameter estimation process for flow simulations with the SWAT model. Moreover, DREAM and DREAM(ZS) are also employed for multi-objective optimizations – by combining flow and water quality variables - and for uncertainty analysis purposes.
The results show that, although the performances acquired after applying the SCE-UA (and an additional manual) optimization are high, the model efficiencies based on the simulations with the parameter values inferred by the single objective DREAM(ZS) analysis, are even better. With the latter configuration, a performance of more than 85% is achieved for both the calibration and validation period. However, the DREAM(ZS) analysis also requires more model runs than the SCE-UA to obtain the results.
Additionally, it is noted that the uncertainty on the parameters and the model predictions after applying the DREAM(ZS) analysis is unrealistically low. However, when a higher relative measurement error is considered for log transformed flow simulations, the parameter and predictive uncertainties become realistic and overall, the observed flow lies mostly within the 95% confidence interval of the model predictions. For both configurations, it is possible to identify influential and noninfluential parameters.
Although the considered measurement error for flow simulations in the different multi-objective optimizations was too high, the potential of applying a multi-objective calibration with water quality variables in order to reduce the uncertainty on the model parameters is illustrated, especially for high measurement errors on the water quality variables.
62 Environmental Resources Sustainability Indicators: an Integrated Assessment Model for Tanzania
Author: Aloyce S. Hepelwa
Supervisor(s): Prof. dr. ir. Willy Bauwens, Prof. Kassim Kulindwa
61 Planning and Design of Rainwater Harvesting Infrastructure in the Middle East
Author: Muamaraldin Ghassan Mhanna
Supervisor(s): Prof. dr. ir. Willy Bauwens
60 Thermal Subpixel Estimation in Urban Areas with Spaceborne Sensing
Author: Wiesam A. A. Essa
Supervisor(s): Prof. dr. Okke Batelaan, Dr. J. van der Kwast