RESEARCH ARTICLE Hydrologically Informed Machine Learning

10.1029/2019WR026933
for Rainfall-Runoff Modeling: A Genetic
Key Points: Programming-Based Toolkit for
• This paper presents a novel
machine learning algorithm, which Automatic Model Induction
is guided through the incorporation
of existing hydrological knowledge
• Proposed machine learning Jayashree Chadalawada1 , H. M. V. V. Herath1 , and Vladan Babovic1
algorithm is based on evolutionary
1 Department of Civil and Environmental Engineering, National University of Singapore, Singapore
computation approach using
genetic programming
• In the present case, the building
blocks of flexible hydrological Abstract Models of water resources systems are conceived to capture the underlying environmental
modeling frameworks represent
elements of the background
dynamics occurring within watersheds. All such models can be regarded as working hypotheses, differing
knowledge in the aspects of process representation and conceptualization. Most of the associated efforts in the water
resources research community is dedicated to development of new models that perform well under specific
Correspondence to:
atmospheric conditions and catchment properties. In this context, flexible modeling frameworks are
V. Babovic, gaining importance as they facilitate the model building process by providing the model building blocks,
[email protected] whereby the hydrologist is free to assemble the model for task at hand. Such flexible models have high
degree of transferability, which in turn aid in progressing toward a unified hydrological theory at
Citation: catchment scale. However, in cases without sufficient insights regarding a catchment characteristics
Chadalawada, J., Herath, H. M. V. V., and/or lack of expert's knowledge, one may have to try a large number of model configurations based
& Babovic, V. (2020). Hydrologically
on available model building blocks to construct an appropriate model for the catchment of interest.
informed machine learning for
rainfall-runoff modeling: A genetic Undoubtedly, this may be time consuming and computationally intensive. This paper proposes a novel
programming-based toolkit for model building algorithm, which uses the full potential of flexible modeling frameworks by searching the
automatic model induction.
model space and inferring suitable model configurations relying on machine learning. Proposed machine
Water Resources Research, 56,
e2019WR026933. https://doi.org/10. learning algorithm is based on evolutionary computation approach using genetic programming (GP).
1029/2019WR026933 State-of-art GP applications in rainfall-runoff modeling so far used the algorithm as a short-term
forecasting tool that generates an expected future time series very similar to neural networks application.
Received 10 DEC 2019 In this case, the proposed algorithm develops a physically meaningful rainfall-runoff model. Although at
Accepted 7 MAR 2020 the moment we learn models using two flexible modeling frameworks (SUPERFLEX and FUSE), the
Accepted article online 15 MAR 2020 model induction toolkit can be armed with any internal coherence building blocks. The model induction
capabilities of the proposed framework have been evaluated on the Blackwater River basin, Alabama,
United States. The model configurations evolved through the model induction toolkit are consistent with
the fieldwork investigations and previously reported research findings.

1. Introduction
The main idea behind the development of hydrological models is to interpret the available hydrologic mea-
surements to understand functioning of watershed and to improve the decision making processes related
to hydrologic problems. The field of hydrological modeling evolved from simple rational methods to fully
process-based distributed models over the last two centuries (Beven, 2012). Regardless, the researchers are
seeking novel methods to provide a better representation of rainfall-runoff phenomena as no model type
can be considered to be universally applicable for a wide range of problems.

1.1. Fixed Models Versus Flexible Models

Design of conceptual models traditionally begins with a perceptual model derived from insights gained on
the basis of fieldwork and experience, proceeding through a mathematical formulation of the hypothesized
structure to the numerically robust implementation in a computer code (Fenicia et al., 2011). Conceptual
hydrological modeling can be broadly classified into single- and multiple-hypothesis (often referred to as
flexible) modeling approaches.

©2020. American Geophysical Union. Development of models with fixed structure is based on the identification of a general model structure that is
All Rights Reserved. physically realistic and applicable to a reasonably wide range of catchments and climatic conditions. Several

CHADALAWADA ET AL. 1 of 23
 Water Resources Research 10.1029/2019WR026933

models that belong to this class are NAM (Nielsen & Hansen, 1973), TOPMODEL (Beven et al., 1995), HBV
(Bergström & Singh, 1995), GR4J (Edijatno et al., 1999; Perrin et al., 2003), and so forth. A fixed model
is computationally efficient due to processing speed and standardization, making it easier to interpret the
dependencies between its parameters and catchment properties, benefiting in terms of model interpretation
and regionalization. However, considering diversity of catchment properties such as area, geology, climate,
and land use, it is unlikely that a model with fixed structure can be equally successfully applied to wide
range of catchments. For example, GR4H model (Perrin et al., 2003) is reported to perform well for wet
catchments with little seasonality, whereas it fails to capture the shift in hydrograph response from wet to dry
conditions (Kavetski & Fenicia, 2011). The main limitation of models with fixed structure is the need to add
specialized modules to adjust constitutive functions governing the fluxes in order to address the ensemble
of processes across range of catchments. This is intuitively obvious and supported by research findings. For
example, incorporation of an additional store (reservoir unit) into an existing conceptual model improves
the streamflow predictions (Buytaert & Beven, 2011).
Testing the performance of a number of models with fixed structures on any single catchment would be
an alternative, yet a cumbersome process. As a generic remedy to the challenge several flexible hydrolog-
ical modeling frameworks have been developed. Recent literature (Weiler & Beven, 2015) argues in favor
of a need to focus on the development of community model, instead of efforts in developing new models.
Flexible models can be configured in a controlled manner to suit any specific catchment and climatic con-
ditions. The analysis and comparison of different model variants can help interpret dominant processes,
suggest improved flux representations, and approximate structural uncertainties (Fenicia et al., 2014). The
challenges in the implementation of this approach include trade-offs between model flexibility, complexity,
comprehensiveness, and computational cost. Depending on the scale and underlying principles, flexi-
ble frameworks can be classified into two categories: model-interfacing frameworks and multicomponent
frameworks.
Model-interfacing frameworks are the ones in which the existing hydrological models are combined to build
larger-scale integrated environmental models. Framework for Understanding Structural Errors (FUSE)
(Clark et al., 2008) and Rainfall-Runoff Modelling Toolbox (RRMT) (Wagener et al., 2001) belong to this
category. Multicomponent frameworks are the ones with finer-scale flexibility in hypothesizing the repre-
sentation of individual fluxes/processes within a catchment. Such frameworks are expected to offer broader
coverage of the model space and deeper insight into underlying processes. Identification of as many model-
ing decisions as possible, higher granularity of model components and their representation in a consistent
manner are the key requirements of a good multicomponent modeling framework. Structure for Unify-
ing Multiple Modelling Alternatives (SUMMA) (Clark et al., 2015, 2015) and SUPERFLEX (Fenicia et al.,
2011; Kavetski & Fenicia, 2011) are two popular multicomponent flexible modeling frameworks. Hydro-
logical model building using flexible modeling frameworks can be done using statistical approaches like
Bayesian Total Error Analysis (BATEA) (Renard et al., 2011) in SUPERFLEX applications (Fenicia et al.,
2014, 2016; van Esse et al., 2013) or using machine learning techniques like data mining in Automatic Model
Configuration Algorithm (AMCA) (Vitolo, 2015).
1.2. SUPERFLEX
SUPERFLEX is a flexible modeling framework that enables hydrologists to hypothesize different model
structures by adding and removing reservoir units, lag functions, and connection elements. Here, the reser-
voir units are used to conceptualize the storage elements and their release of water (e.g., interception,
groundwater, snow effect, and soil moisture). Merging and splitting of fluxes are done by the junction ele-
ments (e.g., Hortonian flow and evaporation). Idea of having lag functions is to account for the delays due
to flow transmission (e.g., channel routing). Further, user has the flexibility of choosing various constitu-
tive functions to describe storage discharge relationships of reservoirs and characteristics of lag functions.
For more details about SUPERFLEX framework, please refer to Fenicia et al. (2011) and Kavetski and
Fenicia (2011), whereas SUPERFLEX applications can be found in van Esse et al. (2013) and Fenicia et al.
(2014, 2016).
1.3. FUSE
FUSE approach (Clark et al., 2008) facilitates the modeler to construct unique model structures by com-
bining model building decisions from four existing hydrological models, namely, PRMS, SACRAMENTO,
TOPMODEL, and ARNO-VIC models. FUSE hypothesizes on basis of a two-layered moisture store and does

CHADALAWADA ET AL. 2 of 23
 Water Resources Research 10.1029/2019WR026933

not resolve spatial variability, thus reducing the number and complexity of modeling hypotheses. Model
building decisions consist of the selection of unsaturated and saturated zone architectures and the selection
of flux equations to represent surface runoff, interflow, evapotranspiration, baseflow, and percolation (Clark
et al., 2010). These model building decisions represent the building blocks for FUSE framework.

2. Machine Learning in Water Resources

In water resources in general, phenomena are governed by physical laws and involve relationships that have
amorphous boundaries and complex latent variables (Karpatne et al., 2018). Majority of the processes are
studied by simulating them using physics-based mathematical models that run on computers. Running such
a physics-based models for large real-world systems is computationally intensive and difficult to generalize
due to number of reasons. The main data sources for water resources modeling originate from observations
on land, space, water bodies, surveys, lab experiments, and relationships simulated over years of system-
atic research. The data collected include huge number of potential variables, from multiple sources and at
multiple resolutions in space and time, with varying degrees of skewness and uncertainty. With the aim
of overcoming the above-mentioned challenges, several research communities: Hydroinformatics, Climate
Informatics, American Geophysical Union, Earth and Space Sciences Informatics to name the few, have
been focusing on the application of artificial intelligence (AI) techniques such as artificial neural networks
(ANNs), support vector machines (SVMs), and evolutionary algorithms (EAs) for the purpose of extract-
ing information from raw data. The following section provides few highlights of major machine learning
techniques that have shown promising performance in applications in hydrology and water resources.
ANN (McClelland & Rumelhart, 1986) is a computing and information processing system motivated by the
functional units of the human brain, namely, neurons. ANNs are data-driven models that can learn func-
tional relationships among data even if the underlying relationships are nonlinear and hard to describe.
Neural networks are used for classification, clustering, vector quantification, pattern association, function
approximation, control, optimization, and search. ANNs have been widely applied in modeling of nonlin-
ear hydrological relationships (Minns & Hall, 1996; Abrahart & See, 2007). A routine for error correction of
numerical models was developed using ANN by Sun et al. (2012). Streamflow estimation has been carried
out using ANN by Nourani et al. (2009) and Humphrey et al. (2016). In Elshorbagy et al. (2002), ANNs were
coupled with chaos theory to estimate the missing data in streamflows. ANN was used to predict water table
elevations by Nayak et al. (2006) and Gholami et al. (2015). A comprehensive list of early ANN applications
in different branches of hydrology can be found in Govindaraju (2000). ANNs are powerful in situations
where conventional statistical and mathematical methods are inadequate, parameters are highly interde-
pendent, and data are intrinsically noisy, incomplete, or error prone. The limitations of ANN are, overfitting,
the network architecture has to be predetermined by the user and requires cautious usage of the resultant
model, as it should not be used beyond the ranges of the data for which it was trained. Despite the wide
usage of ANNs in many scientific fields, it is often criticized as being a “black box” due to the inability to
understand how the model makes its predictions. Hence, ANNs are quite useful in forecasting by making
accurate predictions but provide little or no insights regarding the underlying physical phenomenon.
SVM algorithm (Vapnik, 1992, 1995) is originally developed to solve classification problems but can be
extended to the domain of regression problems (Yu et al., 2004; Elshorbagy et al., 2010). When SVM is used
for regression problems (so-called support vector regression—SVR), the attempt is to fit the error within a
certain threshold. The goal of SVR is to find a function that has at the most deviation from the targets for all
the training data. Kernel functions are used to map the data into higher-dimensional spaces. Studies includ-
ing the recent Sedighi et al. (2016) use SVR in rainfall-runoff modeling. SVR's performance is sensitive to
choice of kernel (linear kernel, polynomial kernel, RBF kernel, and sigmoid kernel), which if appropriately
selected can result in better prediction accuracy.
Model trees (MTs) are regression trees where the parameter space is split into areas (subspaces) and for each
of them a linear regression model is built, which represents an “expert approximation” for that subspace.
The algorithm used to build MTs is M5 (Quinlan, 1992). MT uses linear regression functions at the leaves,
which can predict continuous numeric attributes. The resulting model is analogous to a piece wise linear
function. The M5 algorithm builds a tree by splitting the data based on the values of predictive attributes
(inputs) (Elshorbagy et al., 2010). M5 chooses attributes that minimize intrasubset variation in the class val-
ues of instances that go down each branch. After examining all possible splits (attributes and split values)

CHADALAWADA ET AL. 3 of 23
 Water Resources Research 10.1029/2019WR026933

by exhaustive search, M5 chooses the one that maximizes standard deviation reduction (SDR). Successful
applications of M5 MT in rainfall-runoff modeling can be found in Shortridge et al. (2016) and Rahmani and
Zarghami (2015). MT learns efficiently and can tackle tasks with very high dimensionality, up to hundreds of
variables. It performs very well with linear and semilinear data, which cover a wide range of hydrological sit-
uations. The main advantage of MT is that results are transparent and convenient to use but its performance
given nonlinear data is questionable.
Random forest (RF) is another machine learning technique that can be used to solve both regression and
classification problems. As the name suggests, a RF is a collection of large number of individual decision
trees. Each individual decision tree makes its own prediction, and the prediction resulting from the most
number of trees becomes the final prediction of the RF. Successful RF implementation requires low cor-
relations between the predictions made by individual trees. Feature randomness and bagging are used in
RF to ensure the low correlation between individual decision trees (Yiu, 2019). In Naghibi et al. (2016), RF
was used to generate groundwater spring potential maps, and in Booker and Woods (2014), RFs were used
to predict hydrological variables in ungauged catchments. In another research (Vermeulen & Van Niekerk,
2017), prediction of location and extent of salt-affected areas was carried out using RF.

2.1. Genetic Programming

Genetic programming (GP) (Koza, 1992) is an evolutionary computing technique, which has been used suc-
cessfully in many fields including hydrology and water resources. Evolutionary computation is a general
purpose search algorithm inspired by the theory of biological evolution (Goldberg, 1989). Applications of
evolutionary computation techniques can be found in various fields of water resources engineering rang-
ing from automatic calibration of hydrodynamic model using genetic algorithm (Babovic et al., 1994),
rainfall-runoff modeling (Babovic & Keijzer, 2002), reservoir operations (Giuliani et al., 2015), groundwa-
ter systems (Datta et al., 2014; Fallah-Mehdipour et al., 2013), and water quality studies (Savic & Khu,
2005). Recent applications include multiobjective evolutionary polynomial regression for modeling aquifer
response to rainfall (Doglioni et al., 2015; Giustolisi & Savic, 2006), baseflow prediction (Meshgi et al.,
2014), and modular streamflow model generation (Meshgi et al., 2015) using GP, and short- and long-term
streamflow prediction using Gene expression programming (Karimi et al., 2016).
The main characteristic of GP is that it generates equations and identifies parameters in symbolic form.
The functional operators (mathematical function) and terminal sets (independent variables and physical
or randomly generated constants) are building blocks of GP. The initial set of individual equations can be
generated in a multitude of randomized ways. Once initialized, each equation is evaluated using so-called
objective function. The selection process ensures that the individuals with better fitness values (accuracy)
have higher chance of breeding into the next generation. At each generation, new sets of individuals are
evolved by applying genetic operators (mainly, crossover and mutation), and they form the basis for the next
generation. The individuals evolve as the GP algorithm minimizes the error between model and observed
system responses (Babovic & Keijzer, 2000). Most of the earlier studies on GP applications in hydrology listed
in Oyebode and Adeyemo (2014) and Mehr et al. (2018) involve the use of GP for short-term hydrologic
prediction and forecast, for example, hourly rainfall-runoff modeling (Babovic & Keijzer, 2002), weather
prediction (Bautu & Bautu, 2006), prediction of soil moisture content (Elshorbagy & El-Baroudy, 2009), and
real-time runoff forecasting (Khu et al., 2001).

2.2. Physics Informed Machine Learning

Scientific applications and model building rely on existing background knowledge about domain in ques-
tion. As mentioned above, in most of the GP applications in water resources, GP has been used as a
short-term forecasting tool rather than a model induction tool. Simultaneously, it should be recognized
that there is an enormous amount of knowledge and understanding of physical processes that should not
just be thrown away. Physical meaningfulness and interpretability of models are very important for prun-
ing large spaces of inconsistent solutions, for deeper understanding of the system and knowledge transfer.
Recent studies demonstrate that theory-guided AI is beginning to be pursued in several scientific disciplines
(Faghmous et al., 2014; Karpatne et al., 2018; Keijzer et al., 2001; Babovic & Keijzer, 2005; Singh et al., 2017).
Also, successful AI implementation specific to a research domain involves innovative problem formula-
tion, human insights-based augmentation, and automation of steps in learning process. Hence, we strongly
believe that the most appropriate way forward is to combine the already existing body of knowledge with

CHADALAWADA ET AL. 4 of 23
 Water Resources Research 10.1029/2019WR026933

machine learning techniques in a controlled manner to improve the meaningfulness and interpretability of
the induced models.
Following the same idea, some researchers have developed the augmented versions of ANN and GP by incor-
porating the current domain knowledge in the field of water resources. For example, augmented version of
neural networks called physics-informed neural networks (PNN) was used to learn the governing partial dif-
ferential equations from time series data by Brunton et al. (2016), Rudy et al. (2017), and Raissi et al. (2017).
More theory-guided ANN applications in hydrological modelling can be found in Fleming et al. (2014), Xu
et al. (2019), and Wang et al. (2019). A physics-constrained neural network model was used for estimat-
ing rainfall extremes in Cannon (2018). Fleming and Goodbody (2019) developed a water supply forecast
(WSF) technique by incorporating physical hydrometeorological process knowledge into machine learning
algorithms. As it has been highlighted in the recent review paper by Mcgovern et al. (2019), despite the few
explainable AI applications in hydrometeorological analysis and prediction in past (Cannon & Mckendry,
2002; Fleming, 2007), more researchers have recently begun to focus on the interpretability of AI models
by developing theory-guided ML techniques while resulting a rapid growth in their applications includ-
ing recent applications such as optimal process-oriented hydroclimatic model building and multiple-source
data fusion for hydrologic process simulation (Snauffer et al., 2018; Solander et al., 2019).
In the context of GP, several studies focused on augmenting the performance of GP by incorporation of addi-
tional objective measures relating to dimensional correctness and hydrological theories such that GP evolves
optimal and physically interpretable models. Earlier studies like Babovic & Keijzer, 1999 (1999, 2002) and
Keijzer and Babovic (2002) introduced declarative and preferential biases in the process of model generation
using GP, to minimize the number of physical violations and evolve dimensionally correct equations. It is
reported that implementation of declarative bias results in reduction of search space that can aid faster con-
vergence, while application of preferential bias in a multiobjective search achieves the best balance among
satisfying the defined constraints, fitting the data, regularization, and parsimony of the induced expressions.
In Babovic et al. (2001), dimensionally aware GP is used for the induction of hydraulic formulae from data.
Another study (Babovic, 2009) demonstrated that incorporation of high level concepts into GP algorithm aid
in evolving equations describing sediment transport processes, with performance equal or superior to that
of human experts. In Selle and Muttil (2011), GP is used to test the structure of hydrological models in order
to gain insight about the dominant processes in hydrological systems. In Havlíček et al. (2013), time series
functions (say, lag, simple moving average, and cumulative sum) are incorporated into function library of GP
in addition to the basic algebraic operators, to automatically estimate input lag and improve rainfall-runoff
forecasts. In another recent work (Chadalawada & Babovic, 2017a), GP optimizes the number of reservoir
units (tanks), connections, and outflows, thereby evolving the optimal tank model configuration (Sugawara,
1979) for the catchment of interest based on field observations.
In this contribution, GP is used to its full potential in rainfall-runoff modeling not only as short-term fore-
cast but also as fully fledged models. The domain knowledge is included through incorporation of process
knowledge by adding model building blocks from prevailing flexible rainfall-runoff modeling frameworks
into the GP function set. According to the taxonomy used in Karpatne et al. (2017), the proposed algorithm
can be classified as a hybrid Theory-Guided Data Science (hybrid TGDS) model. Both machine learning
and physics-based model simulation modeling communities in water resources science and engineering are
two major but, historically, quite separate in their modeling approaches. Hence, the interested readers are
encouraged to read the fundamental information in each approaches in Beven (2012) for hydrological pro-
cess simulation modeling and Hsieh (2009) for machine learning applications to environmental science.
More details about the proposed model induction algorithm is given in sequel.

3. Machine Learning Rainfall-Runoff Model Induction Toolkit

The key contribution of this paper is introduction of a novel hydrologically informed machine learning
framework for rainfall-runoff modeling. The intention with this work is not to cover details of any specific
flexible modeling framework or to compare them but to use the model building blocks developed within
two available flexible frameworks in order to generate hydrological models using machine learning based
on GP. The model induction toolkit is armed with model building blocks from two flexible modeling frame-
works (SUPERFLEX and FUSE). However, perhaps more importantly, the framework is readily integratable
with any other flexible modeling framework. The proposed algorithm is applied on Blackwater River basin,

CHADALAWADA ET AL. 5 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B1. Representation of a SUPERFLEX submodel within ML-RR-MI Toolkit (a: Functional representation;
b: parse tree representation; c: 2-D array representation, x1 to x56 are constants between zero and one).

Alabama, United States, by using SUPERFLEX and FUSE model building blocks separately and the results
compared with fieldwork investigations and previous research findings. This toolkit aims to facilitate hydro-
logical modelers to use the full potential of flexible modeling frameworks by building and testing large
number of hypothesizes on a catchment of interest. Proposed tool will be specially helpful in situations
where there is lack of fieldwork insights and human expert's knowledge.
A Machine Learning Rainfall-Runoff Model Induction (ML-RR-MI) Toolkit has been implemented using
GP framework in R programming language (R Core Team, 2013). In addition to SUPERFLEX and FUSE
building block libraries, the ML-RR-MI toolkit includes objective function library containing most of the
commonly used objective functions. The multiobjective optimization scheme in model induction toolkit is
based on Non-dominated Sorting Genetic Algorithm-II (NSGA-II) (Deb et al., 2002). As the optimization is
based on the Pareto optimality concept, the final output of ML-RR-MI toolkit is a set of model configurations
that are not dominated by each other on the selected objective criteria. In this context, the user can select
an optimal model configuration for the catchment of interest based on any selection criterion that may be
based on his/her experience and knowledge in hydrological modeling and/or available fieldwork insights
of the catchment. In this study, a general model selection module has been implemented based on testing
fitness, flow duration curve (FDC) signatures, visual inspection of hydrographs, and complexities of induced
models. Basic steps of the automatic model induction module (GP approach) of the ML-RR-MI toolkit are
given below.
1. Random generation of initial population using the function set, independent variables, and constants.
Ramped half and half method is used to generate individuals (models) of initial population of size N.
Depth of individuals (Tree depth) are bounded by a user defined maximum initial tree depth. Independent
run number = j, where j = 1: maximum number of independent runs.
2. Fitness evaluation of individuals in a multiobjective framework using parallel computation. Calculation
of nondomination rank and crowding distance of each individual. This becomes the parent population
for the next generation.

CHADALAWADA ET AL. 6 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B2. Time series of forcings and observed streamflow of Blackwater catchment

3. Creation of a mating pool using crowded tournament selection method with a user defined tournament
size. Generation number = k, where k = 1: maximum number of generations.
4. Apply genetic operators (crossover and mutation) to create child population of size N. Probabilities of
each generic operator are defined by the user.
5. Fitness evaluation of individuals in child population in a multiobjective framework using parallel
computation.
6. Combine parent population and child population to create combined population (size = 2N). Calculation
of nondomination rank and crowding distance of each individual in combined population.
7. Selection of individual for the next parent population of size N using elitism. Generation number = k + 1.
8. Repeat Steps 3 to 7 until generation number equals to the maximum number of generations.
9. Save best individuals (nondomination rank = 1) to a separate file. Independent run number = j + 1.
10. Repeat Steps 1 to 9 until the independent run number equals to the maximum number of
independent runs.
More details about the model selection module used in this study are given in section 4.3. The parallel
computation has been implemented on model simulations to reduce the computational time.

CHADALAWADA ET AL. 7 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B3. Workflow diagram of ML-RR-MI Toolkit

3.1. SUPERFLEX in ML-RR-MI

The model building blocks available in SUPERFLEX library consist of the reservoir units (interception
reservoir: IR; snow reservoir: WR; unsaturated soil reservoir: UR; combined reservoir: CR; riparian zone
reservoir: RR; fast reacting reservoir: FR; and slow reacting reservoir: SR), lag functions (three lag functions
associated with RR, FR, and SR), and connection elements. Further, there are different constitutive func-
tions (power function, reflected power function, Monad-type kinetics, reflected hyperbolic fuction, tessier
function, and modified logistic curve function), which can be used to represent storage-outflow relation-
ships of reservoirs. Full details about the above building blocks can be found in Fenicia et al. (2016). In
the initialization stage of the ML-RR-MI, different SUPERFLEX submodels are randomly generated in a
guided manner. For example, same reservoir type is not allowed to appear multiple times within one model
configuration, and meaningful order of reservoirs is followed. A SUPERFLEX submodel generated through
ML-RR-MI toolkit consists of three types of model parameters. They are calibration parameters like dis-
charge coefficients and smoothing parameters; option parameters, which are used to define the constitutive
functions to be used; and switching parameters, which decide the presence of each reservoir unit and lag
function within the model configuration.
Figure B1 illustrates an example SUPERFLEX submodel (individual) representation within ML-RR-MI algo-
rithm. Figure B1a shows the functional representation of the model, whereas Figure B1b presents the parse
tree representation. Model is represented as a 2-D array (Figure B1c) within the algorithmic calculations to
facilitate the application of genetic operators (crossover and mutation). Values of the last seven arguments
of the model (here x50 to x56 ) decide the presence of each seven reservoir type within the model architecture
(e.g., [1 0 0 1 0 1 0] represents a model architecture consists with a WR, an UR, and a SR).

3.2. FUSE in ML-RR-MI

Similarly, when the FUSE library is used, algorithm creates individuals constructed from FUSE building
blocks. Here, the model building blocks are the upper layer architecture (single state, separate tension stor-
age, and cascading buckets), lower layer architecture (single state without evapotranspiration, single state
with evapotranspiration, and parallel baseflow reservoirs with evapotranspiration), and the equations to rep-
resent surface runoff (unsaturated zone Pareto, unsaturated zone linear, and saturated zone topographic),
percolation (drainage above field capacity, gravity drainage, and saturated zone control), and evapotranspi-
ration (root weighting and sequential evaporation model) along with the presence or absence of routing and
interflow. Full details about FUSE building blocks are given in Clark et al. (2008, 2010). Here also, a FUSE

CHADALAWADA ET AL. 8 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B4. SUPERFLEX submodels (IR: interception reservoir; RR: riparian reservoir; UR: unsaturated soil reservoir;
FR: fast reacting reservoir; SR: slow reacting reservoir; CR: combined reservoir; sCR and uCR: saturated and
unsaturated zones of CR; Exx , Pxx , and Qxx : evaporation, precipitation, and discharge from respective reservoir; Qbxx and
q
Qxx : baseflow and routed flow through respective reservoir; LRR , LFR : lag functions associated with RR and FR).

submodel generated through ML-RR-MI toolkit consists of the calibration parameters; option parameters,
which decide the upper and lower layer architecture and the equations to be used to express surface runoff,
percolation, and evaporation; and the switching parameters, which will turn off or on the interflow and rout-
ing within the model configuration. Therefore, as the genetic operators are applied, both model structure
and model parameters are optimized simultaneously.

4. Application of ML-RR-MI
4.1. Study Area and Data
The Blackwater River basin near Bradley, Alabama, was selected to test the capabilities of proposed
ML-RR-MI toolkit. Blackwater River basin is one of the Hydrologic Benchmark Network (HBN) station in
Eastern United States (Station 02369800). Blackwater catchment has an area of 227 km2 with land cover
consisting of 89% forested land, 10% agricultural land, and 1% residential areas. Basin consists of two major
landform types, pine hills (75%) and bays (25%). Pine hill regions have a loam or sandy clay loam subsoil,
which are highly permeable in nature. On the other hand, bay areas consist of mucky peat subsoil, which
have a very low permeability and hence tend to retain water year-round. The stream network within the
catchment has a channel length about 21 km, and the main stream is classified as a perennial river. The
climate of the region is affected by the moist tropical air from the Gulf of Mexico, which causes long, hot
summers (Average daily temperature 26.2 ◦ C) and cool, short winters (Average daily temperature 9.1 ◦ C)
with negligible snowfall. The annual average precipitation is about 150 cm and is fairly uniformly spread
throughout the year. More details about the Blackwater River basin is available in Mast and Turk (1999).
Daily data consisting of precipitation (P), potential evaporation (E), mean temperature (T), and dis-
charge (Q) of Blackwater catchment over a 11-year period (01/01/2004 to 31/12/2014) were used for the
model simulations. First-year data were used for model spin-up, and the next 5-year data were used
for model calibration. The remaining 5 years were split into two periods of having 3 and 2 years of duration
for model validation and model testing. The time series plot of forcing terms (P, E, and T) and streamflows
(Q) of Blackwater catchment is shown in Figure B2.

CHADALAWADA ET AL. 9 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B5. Hydrographs of SUPERFLEX submodels (zoomed sections show hydrograph segments from each calibration, validation, and testing periods).

4.2. Objective Functions

Identification of the best performing model for a given catchment is not a trivial matter. The identification
is closely connected to the quality and explanatory power of the objective function employed (Chadalawada
& Babovic, 2017b). Therefore, ML-RR-MI toolkit is armed with a wide range of objective functions used and
beneficial for hydrological processes. In this study, the automatic model generation is carried out using a list
of objective functions described below:
4.2.1. Volumetric Efficiency
Volumetric efficiency (VE) is a performance measure related to the water balance of model simulations
(Criss & Winston, 2008). The efficiency has a range from −∞ to 1 (best).

|∑ |
|N |
| (Qo − Qs )|
| t t |
| t=1 |
VE = 1 − | |, (1)
∑
N
(Qot )
t=1

where, Qot , Qst and N denote observed discharge, simulated discharge and number of time steps respectively.
4.2.2. Kling-Gupta Efficiency (KGE)
Kling-Gupta Efficiency is a combined measure of correlation, standard deviation and mean squared error
(Gupta et al., 2009). KGE has its optimum at unity and is sensitive to the flow variability.

√
KGE = 1 − (r − 1)2 + (𝛼 − 1)2 + (𝛽 − 1)2 (2)

where, r is the linear correlation coefficient between simulated and observed values, 𝛼 denotes measure of
relative variability in the simulated and observed values, and 𝛽 (bias) represents the ratio between mean
𝜎 𝜇
observed and mean simulated flows (𝛼 = 𝜎 s , 𝛽 = 𝜇 s ).
o o

4.2.3. Nash-Sutcliffe Efficiency (NSE)

Nash-Sutcliffe Efficiency (Nash & Sutcliffe, 1970) is the most common objective measure which has been
used to assess the predictive power of hydrological models. The efficiency measure is sensitive to the high

CHADALAWADA ET AL. 10 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B6. FDCs of SUPERFLEX submodels.

flows of the hydrograph and has a range from −∞ to 1 (best).

∑
N
(Qot − Qst )2
t=1
NSE = 1 − (3)
∑
N
(Qot − Qot )2
t=1

where Qot denotes the mean of observed discharge values.

4.2.4. Log Nash-Sutcliffe Efficiency
Log Nash-Sutcliffe Efficiency (logNSE) (Krause et al., 2005) is the Nash-Sutcliffe efficiency of the natural
logarithm transformation of flows and is sensitive to the low flows of the hydrograph. logNSE also has its
optimum at unity.

∑
N
(ln Qot − ln Qst )2
t=1
logNSE = 1 − (4)
∑
N
(ln Qot − ln Qot )2
t=1

VE, KGE, NSE, and logNSE were used in the multiobjective optimization framework of ML-RR-MI toolkit,
as they are sensitive to different flow regimes and characteristics of simulated and observed flows.
4.3. Model Identification and Selection
ML-RR-MI toolkit was applied on the Blackwater catchment data to induce representative model config-
urations for the catchment. SUPERFLEX and FUSE model building blocks were used separately however

CHADALAWADA ET AL. 11 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B7. Density histogram plots of normalized parameter values with satisfactory model performances—SUPERFLEX MI (density = frequency/(1,000*class
interval), number of classes = 40, class interval = parameter range/40).

with the same multiobjective configuration. For SUPERFLEX building blocks, daily rainfall, potential
evaporation, and mean temperature values were used as input variables, whereas for FUSE, daily rain-
fall and potential evaporation values were used. Figure B3 presents the workflow diagram of the proposed
ML-RR-MI toolkit. Settings of the model induction toolkit is given in Table B1.
Output of the ML-RR-MI consists of a set of SUPERFLEX and FUSE submodels, which are not dominated
by each other in terms of training fitness of four objective functions. In the current study, two model con-
figurations (one using SUPERFLEX and the other using FUSE) were selected to capture the catchment
dynamics of Blackwater River basin. Fitness values of each objective function on validation data are calcu-
lated, and Pareto optimal models are reidentified using both training and validation fitness values. Then, the
relative performances in capturing the observed FDC of those Pareto optimal models are calculated using
the standardized signature index sum (SIS) values (Ley et al., 2016). The model configurations with nega-
tive SIS values while delivering good performance with respect to many flow characteristics are considered
to perform better than the average in capturing observed FDC:

|xsa | − xa
Zsa = , (5)
𝜎a

CHADALAWADA ET AL. 12 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B8. FUSE submodels (p: precipitation; e: evaporation; qsx : surface runoff; qb : baseflow; qif : interflow; Zuz and
Zlz : depth of upper layer and lower layer; S1 and S2 : total water content in upper layer and lower layer; q12 : percolation;
S1F and S1T : free water content and tension water content in upper layer; S2T , S2FA , and S2FB : tension water content, free
water content in the primary, and secondary base flow reservoirs; qA b
and qBb : base flow from the primary and secondary
reservoirs; 𝜃 wlt , 𝜃 f ld and 𝜃 sat : soil moisture at wilting point, field capacity, and saturation).

SISs = ZsFHV + ZsFMV + ZsFMS + ZsFLV , (6)

where |xsa | denotes the absolute value of signature index, where s indicates the structure, a denotes the type
of FDC signature index (i.e., FHV: Flow Exceeding Probability < 2%; FMV: 2% < Flow Exceeding Proba-
bility <20%; FMS: 20% < Flow Exceeding Probability < 70%; FLV: Flow Exceeding Probability > 70%), and
x its value; xa and 𝜎 a denote the mean and standard deviation of |xsa | for all s, Z denotes the standardized
value (Z score). Model configurations having negative SIS values for both calibration and validation peri-
ods were considered in the final stage of optimal model selection. Finally, the visual inspection of simulated
hydrographs and model complexities in terms of number of model parameters were considered to select the
optimal model configuration for the catchment of interest.
4.4. Parameter Sensitivity and Uncertainty Analysis
Parameter sensitivity and uncertainty analysis of the selected SUPERFLEX and FUSE submodels was car-
ried out by changing one model parameter at a time and keeping remaining model parameters at their
calibrated values. This allows for analysis of the distribution of parameter values that provide satisfactory
model performances. Here, the NSE was used to measure the model performances. The parameter val-
ues with efficiencies greater than 0.6 were considered as satisfactory model performances. For each model
parameter, 1,000 parameter values that give satisfactory model performances were identified and then used
to draw the density histogram plots (Y axis represents the density, and X axis represents the normalized
parameter range). The shape of the distribution of each histogram was used to identify the sensitivity and
the degree of uncertainty associated with the model parameter of interest. The 90% uncertainty bounds were
calculated as proposed in Generalized Likelihood Uncertainty Estimation (GLUE) (Beven & Binley, 1992)
by running 1,000 model simulations with the parameter values identified in the above step for both SUPER-
FLEX and FUSE submodels. Here again, NSE was used as the objective function, and parameter sets with
NSE values greater than 0.6 were identified as the behavioral models and used to calculate the uncertainty
bounds. Finally, the percentage of observed discharge values that lie between the uncertainty bounds (5%
and 95% confidence limits) was calculated and used as an indicator to see whether the parameter uncertainty
alone can compensate for the total output uncertainty.

5. Results
5.1. SUPERFLEX Building Blocks
Following the methodology described on section 4.3, ML-RR-MI algorithm came up with four
SUPERFLEX submodels as the final stage model configurations. As illustrated in Figure B4, Model I (MI)

CHADALAWADA ET AL. 13 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B9. Hydrographs of FUSE submodels (zoomed sections show hydrograph segments from each calibration,
validation, and testing periods).

has a four-reservoir configuration including an unsaturated soil reservoir (UR), a fast reacting reservoir (FR),
a riparian reservoir (RR), and a slow reacting reservoir (SR) along with two lag functions associated with
RR and FR. Model II (MII) differs from MI by an interception reservoir (IR) instead of RR. Both Model III
(MIII) and Model IV (MIV) have five-reservoir configurations. MIII consists of an IR, an UR, a RR, a FR,
and a SR. MIII incorporates two lag functions linked by RR and FR. MIV differs from MIII by the inclusion
of a combined reservoir (CR) instead of IR.
Simulated hydrographs of each submodel for calibration, validation, and testing periods are shown in
Figure B5, whereas associated FDCs are given in Figure B6 along with the observed hydrograph. All four
submodels were able to capture the discharge signature of the catchment reasonably well. It can be noted
that in all four SUPERFLEX submodels, at some instances the model simulations underestimate the peak
discharges. Simulated FDCs closely follow the observed FDC in high and medium flows and tend to deviate
at low flows. As the model predictions are very much similar for all four submodels, the additional com-
plexities of MIII and MIV (five-reservoir configurations) may not be supportable. Hence, either MI or MII
(four-reservoir configurations) can be selected to represent the catchment dynamics of Blackwater River
basin. In terms of number of model parameters, the order of model complexities in both MI and MII is pretty
much similar (MI has 20 model parameters and MII has 21 model parameters). Finally, since submodel MI
produces more negative SIS values for both training and validation periods than MII (FDC of MI is closer
to the observed FDC in Figure B6) was identified as the optimal SUPERFLEX model configuration for the
Blackwater catchment. The performance matrices of all four submodels with respect to the selected four
objective functions are provided in Table B2.
The total outflow of MI (Qt ) consists of three flow components: the flow routed through RR (QRR ), the flow
q
routed through FR (QFR ), and the flow routed through SR (QSR ). Evaporation originates from UR, FR, and
SR. Model parameter details of all four SUPERFLEX submodels are given in the appendix (Tables B4 and
q q
B1). The constitutive functions that govern the flow through UR (QUR ), FR (QFR ), and SR (QSR ) of MI are
modified logistic curve function, power function, and linear function, respectively. The uncertainty and

CHADALAWADA ET AL. 14 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B10. FDC Signatures of FUSE submodels.

sensitivity analysis was carried out as described in section 4.4, and the results are presented in Figure B7. The
percentage of observed values that lie in between 90% uncertainty bounds of MI is 70.15. Such high value
indicates that the parameter uncertainty alone is capable of compensating for the other sources of errors
and hence can account for the total output uncertainty of MI (Boukezzi et al., 2016). Model parameters with
narrowly distributed histograms (Figure B7) were identified as the sensitive model parameters of MI. Out
of the total 20 model parameters, there are eight significantly sensitive model parameters. They are D_R
(fraction of inflow to RR), Smax_UR (maximum reservoir capacity of UR), Beta_E_UR (smoothing factor for
Pot.E.T. of UR), K_Qb_UR (percolation coefficient of UR), K_Qq_SR (K in Q = K*S of SR), Ce (evaporation
multiplication factor), Tlag (base of rising limb of LRR and LFR ), and D_S (fraction of precipitation to FR).
Experimental investigations conducted in Blackwater catchment reveal that the main river channel of the
catchment is a perennial type river. This suggests that there might be a stable groundwater component
among the catchment runoff dynamics. Furthermore, 75% of the basin area is characterized as pine hills
with a highly permeable soil type, which may facilitate the deep percolation. So it is reasonable that SR is
included in the model configuration of MI as SR represents the slow runoff components like groundwater
flow. Additionally, the SR in MI is a linear reservoir and linear models are considered to have better repre-
sentation of slow dynamics of groundwater reservoirs (Fenicia et al., 2006). Since the remaining 25% of the
basin area consists of wetlands with poorly drained soil strata, it can be expected to have fast runoff com-
ponents as well among the catchment runoff dynamics. Having a FR within the model architecture enables
MI to capture these fast runoff components like saturation excess overland flow. In addition to that, MI is
capable of capturing the threshold like behaviors and nonlinearity in catchment dynamics well, as the con-
q
stitutive function that governs the flow routed through the UR (QUR ) is modified logistic curve function.
Other than UR, FR, and SR, MI is consisted with a RR in its model configuration. The concept behind hav-
ing a RR is to account for the runoff generation due to the precipitation on water bodies. The Blackwater
River basin does not only consist of wetlands but also of dense network of stream channels. So it is rea-
sonable to have a RR included in MI model configuration. On top of that, model parameter D_R, which
controls the fraction of precipitation goes to the RR, has been identified as a sensitive parameter and hence
shows the importance of having a RR in the model configuration of MI. Since the Blackwater River basin
is a comparatively large catchment, it can be expected to have a delayed response of runoff with respect to
precipitation forcing. The two lag functions associated with RR and FR may help MI to better represent this
q
delayed response. Although the link between input precipitation to the UR and the outflow from UR (QUR )
represents the infiltration excess overland flow, the associated model parameters (P_ED_max and m_P_ED)

CHADALAWADA ET AL. 15 of 23
 Water Resources Research 10.1029/2019WR026933

Figure B11. Density histogram plots of normalized parameter values with satisfactory model performances—FUSE MI (density = frequency/(1,000*class
interval), Number of classes = 40, class interval = parameter range/40).

are totally insensitive to the model performances of MI. This suggests that there is a negligible contribution
from the infiltration excess overland flow toward the total runoff generation and it is quite reasonable as the
majority of basin area consists of a highly permeable soil strata.

5.2. FUSE Building Blocks

The same procedure as described in section 5.1 was followed for FUSE building blocks as the model inven-
tory in ML-RR-MI toolkit. Two FUSE submodels (Figure B8) were identified as the final stage model
configurations to represent Blackwater catchment dynamics. Model I (MI) has the similar upper and lower
layer architecture as in the FUSE parent model SACREMENTO. Model II (MII) differs from MI by having
an upper layer architecture similar to FUSE parent model ARNO-VIC. Simulated hydrographs of each sub-
model have been compared with the observed hydrograph in Figure B9, and the simulated and observed
FDCs are shown in Figure B10. As it was seen with SUPERFLEX submodels, both FUSE submodels are
also able to capture the catchment's discharge signature reasonably well. Simulated FDCs of both models
are following the observed FDC very closely in all flow regimes. The model complexities in terms of num-
ber of model parameters are similar in both MI and MII (14 model parameters each). FUSE submodel MI

CHADALAWADA ET AL. 16 of 23
 Water Resources Research 10.1029/2019WR026933

Table B1
ML-RR-MI Toolkit Settings
Parameter Value
Independent runs 20
Population size 2,000
Number of generations 50
Initialization Ramped half and half
Function set SUPERFLEX, FUSE, +, -, /, *
Variables independent/dependent P, E, T/ Q
Constant range - min/max 0/1
Tree depth - initial/max 2/4
Selection method Tournament (size = 4)
Probability of crossover 0.7
Probability of mutation 0.5/0.3/0.3/0.5
Constant/node/separation/tree
Number of parallel processors 20
Parallelization stage Model simulation stage

was selected as the optimal model configuration to represent the Blackwater catchment dynamics, as it gave
more negative SIS values both in calibration and validation periods than MII.
Performance matrices of submodel MI and MII are given in Table B3. Model parameter details for both mod-
els are given in appendix (Table B2). The upper layer architecture of model MI consists of a dual storage
(tension and free storage) reservoir, and the lower layer is consisted with a tension reservoir and two parallel
tanks (similar to the FUSE parent model SACRAMENTO). Surface runoff from the model MI is generated
through the saturation excess mechanism and calculated as done in FUSE parent model TOPMODEL. Per-
colation of model MI is through the gravity drainage and governed by the moisture content in the lower layer
(similar to the FUSE parent model SACRAMENTO). Evaporation occurs in both soil layers of the model MI
sequentially where the demand is first satisfied by the upper layer and the balance is satisfied by the lower
layer (similar to the FUSE parent models SACRAMENTO and PRMS). Other than the surface runoff and
baseflow, model MI also consists of an interflow that makes the total model outflow to have three parallel
flow components. The total model outflow is routed using the two-parameter Gamma distribution to incor-
porate the time delay in runoff. The analysis of uncertainty and sensitivity results are shown in Figure B11.
As it can be seen, out of 14 model parameters, five model parameters can be identified as significant model
parameters of model MI. They are maxwatr_2 (maximum total storage in lower soil layer), fprimqb (frac-
tion storage in first baseflow reservoir), iflwrte (interflow rate), qbrate_2b (baseflow depletion rate second
reservoir), and timedelay (time delay in runoff). The percentage of observed values that lie between 90%
uncertainty bounds of model MI is 75.85. Therefore, it is safe to assume that parameter uncertainty alone is
sufficient to compensate total output uncertainty of model MI.
As it was seen with the total outflow of SUPERFLEX submodel MI, the total outflow of FUSE submodel MI
also consists of three parallel runoff components: baseflow (qAb and qBb ), interflow (qif ), and surface runoff
(qsx ). As reflected from the calibrated values of the relevant model parameters and the sensitivity analysis
(maxwatr_1 < maxwatr_2, low baseflow exponent of 1.15, and most of the sensitive parameters identified

Table B2
Performance Matrices of SUPERFLEX Submodels
VE KGE NSE logNSE
Model Cal. Val. Test Cal. Val. Test Cal. Val. Test Cal. Val. Test
SUPERFLEX MI 0.620 0.670 0.595 0.823 0.870 0.780 0.721 0.801 0.569 0.699 0.676 0.566
SUPERFLEX MII 0.589 0.579 0.556 0.849 0.859 0.732 0.717 0.745 0.536 0.563 0.467 0.477
SUPERFLEX MIII 0.623 0.627 0.557 0.806 0.872 0.760 0.722 0.780 0.560 0.665 0.580 0.516
SUPERFLEX MIV 0.613 0.655 0.563 0.841 0.869 0.741 0.707 0.770 0.510 0.699 0.664 0.566

CHADALAWADA ET AL. 17 of 23
 Water Resources Research 10.1029/2019WR026933

Table B3
Performance Matrices of FUSE Submodels
VE KGE NSE logNSE
Model Cal. Val. Test Cal. Val. Test Cal. Val. Test Cal. Val. Test
FUSE MI 0.644 0.700 0.686 0.863 0.900 0.835 0.731 0.818 0.685 0.755 0.772 0.758
FUSE MII 0.649 0.712 0.684 0.849 0.903 0.831 0.747 0.828 0.659 0.739 0.777 0.754

are associated with lower layer reservoir), model MI's runoff generation is predominantly governed by the
groundwater flow. This is quite meaningful due to the fact that 75% of the basin area of Blackwater catchment
consists of highly permeable soil column and hence the deep percolation may contribute to have a significant
groundwater storage. FUSE submodel MI also consists of a linear groundwater reservoir, which is considered
to have a better representation of slow dynamics of groundwater reservoirs. The surface runoff of model MI is
due to the saturation excess overland flow, which was identified as a possible runoff generation mechanism
in Blackwater River basin due to the presence of bays and wetlands (25% of land area). It is interesting to
note that both SUPERFLEX submodel MI and FUSE submodel MI are having lag functions in their model
configurations and the associated time delays are pretty close to each other (SUPERFLEX MI: 1.51 days and
FUSE MI: 1.49 days).

6. Summary
The powerful nature of any flexible modeling framework is that it provides the full freedom to the modeler
to build his/her own model using the model building blocks of the modeling framework. At the same time,
use of flexible frameworks requires lot of time and demands high computational efforts and hence may hold
the hydrologist from benefiting from the full potential provided by the flexible modeling framework. This
paper introduces a novel approach (ML-RR-MI) for hydrological model building using a machine learning
technique, GP. ML-RR-MI differs from the rest of machine learning applications in rainfall-runoff modeling,
as it not only generates the runoff predictions but also develops a physically meaningful hydrological model

Table B4
Model Parameter Details of SUPERFLEX Submodels
Calibrated value
Parameter (unit) Description (parameter range) MI MII MIII MIV
Beta_Qq_UR (−) Runoff coefficient parameter (0.001–10) 8.59 10 10 8.99
Smax_UR (mm) Maximum reservoir capacity (0.1–10,000) 910 1443 1473 1127
Beta_E_UR (−) Smoothing factor for Pot.E.T. (0.01–10) 0.01 0.01 0.01 0.01
SiniFr_UR (−) State initial factor (0–1) 0.30 0.24 0.22 0.24
K_Qb_UR (1/t) Percolation coefficient (0.000001–2) 1e–06 1e–06 1e–06 2.3e–05
mu_Qq_UR (−) Modified logistic curve parameter (0.1–1) 0.33 0.19 0.17 0.32
K_Qq_FR (mmalpha /t) K in Q = K ∗ Salpha (0.0001–10) 0.49 0.37 5.29 0.49
K_Qb_FR (1/t) Controls flow from FR to SR (0.001–10) 3.15 3.15 3.15 3.15
m_E_FR (−) Smoothing factor for Pot.E.T. (0.01–2) 1.98 1.41 1.41 1.85
alpha_Qq_FR (−) alpha in Q = K ∗ Salpha (0.1–10) 7.97 7.64 6.56 7.64
D_F (−) Fraction of inflow to Qb from Qq (0–1) 0.36 0.42 0.46 0.28
Smax_IR (mm) Maximum reservoir capacity of IR (0.1–20) — 0.1 0.1 —
m_QE_IR (−) Smoothing factor for Pot.E.T. (0.001–1) — 0.362 0.362 —
D_I (−) Constant fraction of inflow to IR (0–1) — 0.067 0.114 —
K_Qq_RR (1/t) K in the outflow Q=K*(S) from RR (0.05–4) 1.68 — 1.43 2.46
D_R (−) Fraction of inflow to RR (0–1) 0.045 — 0.023 0.048
Smax_CR (mm) Maximum reservoir capacity of CR (0.1–10,000) — — — 443
Umax_uCR (−) Threshold parameter for scaled storage of unsaturated zone (0.8–0.9) — — — 0.9
Smin_uCR (mm) Minimum storage of unsaturated zone (0.001–1) — — — 0.79
Beta_Qq_uCR (−) Parameter of constitutive equation of uCR (0.001–10) — — — 9.06

CHADALAWADA ET AL. 18 of 23
 Water Resources Research 10.1029/2019WR026933

Table B5
Model Parameter Details of SUPERFLEX Submodels
Calibrated value
Parameter (unit) Description (parameter range) MI MII MIII MIV
K_Qb_uCR (1/t) K in the outflow Q = K*(S) of uCR (1e–06 to 1) — — — 0.8
Sevmax_CR (mm) Threshold for evaporation contribution (0.01–1e04) — — — 10,000
Beta_E_CR (−) Parameter of constitutive equation of evaporation (0.01–10) — — — 1.719
Beta_Qq_sCR (−) Parameter of constitutive equation of saturated zone (0.01–10) — — — 0.01
K_Qb_sCR (1/t) K in the outflow Q=K*(S) of sCR (1e–06 to 2) — — — 1.486
K_Qd_sCR (1/t) K in the drain outflow Q=K*(S) of sCR (0–4) — — — 0.305
option_f (−) Constitutive function of FR (1–2) 1:Power 1 1 1
option_u (−) Constitutive function of UR (1–4) 4:Modified logistic 4 4 4
option_c (−) Constitutive function of CR (1–2) — — — 1:Power
option_i (−) Constitutive function of IR (1–2) — 1 1 —
Lag_RR (−) Presence of lag function (0–1) 1:Present 0 1 1
Lag_FR (−) Presence of lag function (0–1) 1:Present 1 1 1
Lag_SR (−) Presence of lag function (0–1) 0:Absent 0 0 0
D_C (−) Fraction of flow to sCR from outflow of uCR (0–1) — — — 0.08
Ce (−) Evaporation multiplication factor (0.1–3) 0.92 1.36 1.36 0.87
P_ED_max (mm/t) Infiltration excess threshold (0.1–1e07) 3.6e6 0.1 0.1 3.4e06
m_P_ED (mm/t) Infiltration excess flow smoothing factor (0.001–10) 10 10 10 10
Tlag (t) Base of rising limb (1–10) 1.51 1.48 1.48 1.50
D_S (−) Fraction of precipitation to FR (0–1) 0.038 0.081 0.041 0.033

that helps the hydrologist to better understand the catchment dynamics. This approach considers the model
space and automatically identifies the suitable model configurations for a catchment of interest. Modeler's
hydrological knowledge and experience and/or the available experimental investigations can be coupled
here to select an optimal model out of the induced model configurations.
This research is part of a broader effort aiming to develop a methodology for automated model induction
built from data and using existing background knowledge. In the present case, the building blocks of flexible
hydrological modeling frameworks represent elements of the background knowledge. The ultimate objective
of the effort is to learn meaningful models that can be interpreted by the domain experts.
Simplistic application of data-driven black-box machine learning techniques may lead to development of
accurate yet meaningless models, which do not satisfy basic hydrological insights and may have severe diffi-
culties with interpretation. Yet it is the value associated with interpretability of resulting models that utilize
elements of background knowledge is what distinguishes our approach from other methods. The approach
makes it possible to rely on the resulting models with more than just statistical confidence. It is only in this
way that we can take full advantage of machine-aided knowledge discovery and advance our understanding
of physical processes.
We are only beginning to develop effective ways of combining the strengths of human cognition with
those of computational discovery systems. However, it is fairly easy to anticipate a more widespread use of
science-informed machine learning in the process of scientific discovery and automated model building.

Appendix A: Parameter Details of SUPERFLEX Submodels

Tables B4 and B5 present the model parameter details of SUPERFLEX submodels and their calibrated values.

Appendix B: Parameter Details of FUSE Submodels

Table B6 presents the model parameter details of FUSE submodels and their calibrated values.

CHADALAWADA ET AL. 19 of 23
 Water Resources Research 10.1029/2019WR026933

Table B6
Model Parameter Details of FUSE Submodels
Calibrated value
Parameter (unit) Description (parameter range) MI MII
maxwatr_1 (mm) Maximum total storage in upper soil layer (25–500) 195.3 101.9
maxwatr_2 (mm) Maximum total storage in lower soil layer (50–5,000) 802.8 944.9
fracten (−) Fraction total storage as tension storage (0.05–0.95) 0.40 0.75
fprimqb (−) fraction storage in 1st baseflow reservoir (0.05–0.95) 0.05 0.95
sacpmlt (−) SAC model percolation multiplier for dry soil layer (1–250) 250 -
sacpexp (−) SAC model percolation exponent for dry soil layer (1–5) 1.63 -
percrte mm/day) epercolation rate (0.01–1,000) - 423.6
percexp (−) percolation exponent (1–20) - 5.23
percfrac (−) fraction of percolation to tension storage (0.05–0.95) 0.04 0.31
iflwrte (mm/day) Interflow rate (0.01–1,000) 82.9 1000
qb_powr (−) Baseflow exponent (1–10) 1.15 7.67
qbrate_2a (1/day) baseflow depletion rate 1st reservoir (0.001–0.25) 0.083 0.016
qbrate_2b (1/day) baseflow depletion rate 2nd reservoir (0.001–0.25) 0.005 0.007
loglamb (m) Mean value of the log-transformed topographic index (5–10) 6.12 5.68
tishape (−) Shape parameter for the topo index gamma distribution (2–5) 3.65 3.54
timedelay (days) Time delay in runoff (0.01–5) 1.49 1.52
arch1 (−) Architecture of upper layer (1–3) 2:Dual storage 1:Single state
arch2 (−) Architecture of lower layer (1–4) 2:tension reservoir plus two parallel tanks 2
qsurf (−) Runoff equation (1–3) 3:TOPMODEL parameterization 3
qperc (−) Percolation equation (1–3) 2:Gravity drainage 3:Saturated zone control
esoil (−) Evaporation equation (1–2) 2:Sequential 2
qintf (−) Interflow (1–2) 2:Allowed 2
q_tdh (−) Routing (1-d-2) 2:Allowed 2

Acknowledgments References
We are grateful to Dr. Fabrizio Fenicia
Abrahart, R., & See, L. (2007). Neural network modelling of non-linear hydrological relationships. Hydrology and Earth System Sciences
for sharing his expertise in
Discussions, 11(5), 1563–1579.
SUPERFLEX framework and Dr.
Babovic, V. (2009). Introducing knowledge into learning based on genetic programming. Journal of Hydroinformatics, 11(3-4), 181–193.
Vojtech Havlicek, Faculty of
Babovic, V., & Keijzer, M. (1999). Forecasting of river discharges in the presence of chaos and noise. In J. Marsalek (Ed.), Coping with floods:
Environmental Sciences, Czech
Lessons learned from recent experiences, nato arw series. Dordrecht: Kluwer.
University of Life Sciences, Prague,
Babovic, V., & Keijzer, M. (2000). Genetic programming as a model induction engine. Journal of Hydroinformatics, 2(1), 35–60.
for his assistance in R programming.
Babovic, V., & Keijzer, M. (2002). Rainfall runoff modelling based on genetic programming. Hydrology Research, 33(5), 331–346.
Hydrometeorological data set of
Babovic, V., & Keijzer, M. (2005). Determining equations for vegetation induced resistance using genetic programming. Proceedings of the
the Blackwater River basin was
7th annual conference on genetic and evolutionary computation (pp. 1999–2006). Washington DC, USA: ACM.
downloaded from CAMELS data
set (Newman et al., 2015). Babovic, V., Keijzer, M., Aguilera, D. R., & Harrington, J. (2001). An evolutionary approach to knowledge induction: Genetic programming
in hydraulic engineering. Proceedings of the world water and environmental resources congress (pp. 64). Orlando, Florida.
Babovic, V., Zhengyi, W., & Larsen, L. (1994). Calibrating hydrodynamic models by means of simulated evolution. In 1st conference on
hydroinformatics (pp. 1025). Delft, Netherlands.
Bautu, A., & Bautu, E. (2006). Meteorological data analysis and prediction by means of genetic programming. In Proceedings of the 5th
workshop on mathematical modeling of environmental and life sciences problems (pp. 35–42). Constanta, Romania.
Bergström, S., & Singh, V. (1995). The HBV model. In V. P. Singh (Ed.), Computer Models of Watershed Hydrology (443–476). Highlands
Ranch: Water Resources Publications.
Beven, K. (2012). Down to basics: Runoff processes and the modelling process. In Rainfall-runoff modelling: the primer, pp. 1–22.
Beven, J. K., & Binley, M. A. (1992). The future of distributed models: Model calibration and uncertainty prediction. Hydrological Processes,
6, 278–298.
Beven, K., Lamb, R., Quinn, P., Romanowicz, R., Freer, J. (1995). Topmodel. In V. P. Singh (Ed.), Computer models of watershed hydrology
(pp. 627–668). Colorado: Water Resource Publications.
Booker, D. J., & Woods, R. A. (2014). Comparing and combining physically-based and empirically-based approaches for estimating the
hydrology of ungauged catchments. Journal of Hydrology, 508, 227–239.
Boukezzi, Z. L., Boukezzi, L., & Errih, M. (2016). Uncertainty analysis of HEC-HMS model using the glue method for flash flood forecasting
of Mekerra watershed, Algeria. Arabian Journal of Geosciences, 9(20), 1–12.
Brunton, S. L., Proctor, J. L., & Kutz, J. N. (2016). Discovering governing equations from data by sparse identification of nonlinear dynamical
systems. In Proceedings of the National Academy of Sciences, 113(15), 3932–3937.

CHADALAWADA ET AL. 20 of 23
 Water Resources Research 10.1029/2019WR026933

Buytaert, W., & Beven, K. (2011). Models as multiple working hypotheses: Hydrological simulation of tropical alpine wetlands. Hydrological
Processes, 25(11), 1784–1799.
Cannon, A. J. (2018). Non-crossing nonlinear regression quantiles by monotone composite quantile regression neural network, with
application to rainfall extremes. Stochastic Environmental Research and Risk Assessment, 32(11), 3207–3225.
Cannon, A. J., & Mckendry, I. G. (2002). A graphical sensitivity analysis for statistical climate models: Application to indian monsoon
rainfall prediction by artificial neural networks and multiple linear regression models. International Journal of Climatology, 22(13),
1687–1708. https://doi.org/10.1002/joc.811
Chadalawada, J., & Babovic, V. (2017a). Genetic programming for automatic hydrological modelling. In EGU general assembly conference
abstracts, 19, (pp. 15990). Vienna, Austria.
Chadalawada, J., & Babovic, V. (2017b). Review and comparison of performance indices for automatic model induction. Journal of
Hydroinformatics, 21, 13–31.
Clark, M. P., Hilary, K. M., Daniel, B. G. C., Kavetski, D., & Woods, R. A. (2010). Hydrological field data from a modeller's perspective: Part
2: Process-based evaluation of model hypotheses. Hydrological Processes, 25(4), 523–543.
Clark, M. P., Nijssen, B., Lundquist, J. D., Kavetski, D., Rupp, D. E., Woods, R. A., et al. (2015). A unified approach for process-based
hydrologic modeling: 1. Modeling concept. Water Resources Research, 51(4), 2498–2514. htts://doi.org/10.1002/2015wr017198
Clark, M. P., Nijssen, B., Lundquist, J. D., Kavetski, D., Rupp, D. E., Woods, R. A., et al. (2015). The structure for unifying multiple modeling
alternatives (SUMMA), Version 1.0: Technical description, NCAR Tech. Note NCAR/TN-5141STR.
Clark, M. P., Slater, A. G., Rupp, D. E., Woods, R. A., Vrugt, J. A., Gupta, H. V., et al. (2008). Framework for Understanding Structural Errors
(FUSE): A modular framework to diagnose differences between hydrological models. Water Resources Research, 44, W00B02. https://
doi.org/10.1029/2007WR006735
Criss, R. E., & Winston, W. E. (2008). Do Nash values have value? Discussion and alternate proposals. Hydrological Processes, 22(14), 2723.
Datta, B., Prakash, O., & Sreekanth, J. (2014). Application of genetic programming models incorporated in optimization models for
contaminated groundwater systems management, Evolve-a bridge between probability, set oriented numerics, and evolutionary computa-
tion v (pp. 183–199). New York, NY, USA: Springer.
Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions
on Evolutionary Computation, 6(2), 182–197.
Doglioni, A., Galeandro, A., & Simeone, V. (2015). Evolutionary data-driven modelling of Salento shallow aquifer response to rainfall,
Engineering geology for society and territory-volume 3 (pp. 281–285). Switzerland: Springer.
Edijatno, Nascimento, N. D. O., Yang, X., Makhlouf, Z., & Michel, C. (1999). GR3J: A daily watershed model with three free parameters.
Hydrological Sciences Journal, 44(2), 263–277.
Elshorbagy, A., Corzo, G., Srinivasulu, S., & Solomatine, D. (2010). Experimental investigation of the predictive capabilities of data driven
modeling techniques in hydrology—Part 1: Concepts and methodology. Hydrology and Earth System Sciences, 14(10), 1931–1941.
Elshorbagy, A., & El-Baroudy, I. (2009). Investigating the capabilities of evolutionary data-driven techniques using the challenging
estimation of soil moisture content. Journal of Hydroinformatics, 11(3-4), 237–251.
Elshorbagy, A., Simonovic, S. P., & Panu, U. S. (2002). Estimation of missing streamflow data using principles of chaos theory. Journal of
Hydrology, 255(1–4), 123–133.
Faghmous, J. H., Banerjee, A., Shekhar, S., Steinbach, M., Kumar, V., Ganguly, A. R., & Samatova, N. (2014). Theory-guided data science
for climate change. Computer, 47(11), 74–78.
Fallah-Mehdipour, E., Haddad, O. B., & Mariño, M. (2013). Prediction and simulation of monthly groundwater levels by genetic
programming. Journal of Hydro-environment Research, 7(4), 253–260.
Fenicia, F., Kavetski, D., & Savenije, H. H. G. (2011). Elements of a flexible approach for conceptual hydrological modeling: 1. Motivation
and theoretical development. Water Resources Research, 47, W11510. https://doi.org/10.1029/2010WR010174
Fenicia, F., Kavetski, D., Savenije, H. H. G., Clark, M. P., Schoups, G., Pfister, L., & Freer, J. (2014). Catchment properties, function, and
conceptual model representation: Is there a correspondence? Hydrological Processes, 28(4), 2451–2467.
Fenicia, F., Kavetski, D., Savenije, H. H., & Pfister, L. (2016). From spatially variable streamflow to distributed hydrological models: Analysis
of key modeling decisions. Water Resources Research, 52, 954–989. https://doi.org/10.1002/2015WR017398
Fenicia, F., Savenije, H. H. G., Matgen, P., & Pfister, L. (2006). Is the groundwater reservoir linear? Learning from data in hydrological
modelling. Hydrological and Earth System Sciences, 10, 139–150.
Fleming, S. W. (2007). Artificial neural network forecasting of nonlinear Markov processes. Canadian Journal of Physics, 85(3), 279–294.
https://doi.org/10.1139/p07-037
Fleming, S. W., Bourdin, D. R., Campbell, D., Stull, R. B., & Gardner, T. (2014). Development and operational testing of a super-ensemble
artificial intelligence flood-forecast model for a Pacific Northwest River. JAWRA Journal of the American Water Resources Association,
51(2), 502–5125. https://doi.org/10.1111/jawr.12259
Fleming, S. W., & Goodbody, A. G. (2019). A machine learning metasystem for robust probabilistic nonlinear regression-based forecasting
of seasonal water availability in the us west. IEEE Access, 7, 11,9943–11,9964.
Gholami, V., Chau, K. W., Fadaee, F., Torkaman, J., & Ghaffari, A. (2015). Modeling of groundwater level fluctuations using dendrochronol-
ogy in alluvial aquifers. Journal of Hydrology, 529, 1060–1069.
Giuliani, M., Castelletti, A., Pianosi, F., Mason, E., & Reed, P. M. (2015). Curses, tradeoffs, and scalable management: Advancing
evolutionary multiobjective direct policy search to improve water reservoir operations. Journal of Water Resources Planning and
Management, 142(2), 4,015,050. https://doi.org/10.1061/(asce)wr.1943-5452.0000570
Giustolisi, O., & Savic, D. A. (2006). A symbolic data-driven technique based on evolutionary polynomial regression. Journal of Hydroin-
formatics, 8(3), 207–222.
Goldberg, D. E. (1989). Genetic algorithms in search, optimization, and machine learning. Boston: Addison-Wesley.
Govindaraju, R. S. (2000). Artificial neural networks in hydrology. II: Hydrologic applications. Journal of Hydrologic Engineering, 5(2),
124–137.
Gupta, H. V., Kling, H., Yilmaz, K. K., & Martinez, G. F. (2009). Decomposition of the mean squared error and NSE performance criteria:
Implications for improving hydrological modelling. Journal of Hydrology, 377(1), 80–91.
Havlíček, V., Hanel, M., Máca, P., Kuráž, M., & Pech, P. (2013). Incorporating basic hydrological concepts into genetic programming for
rainfall-runoff forecasting. Computing, 95(1), 363–380.
Hsieh, W. W. (2009). Machine learning in the environmental sciences. Cambridge, United Kingdom: Cambridge University Press.
Humphrey, G. B., Gibbs, M. S., Dandy, G. C., & Maier, H. R. (2016). A hybrid approach to monthly streamflow forecasting: Integrating
hydrological model outputs into a Bayesian artificial neural network. Journal of Hydrology, 540, 623–640.

CHADALAWADA ET AL. 21 of 23
 Water Resources Research 10.1029/2019WR026933

Karimi, S., Shiri, J., Kisi, O., & Shiri, A. A. (2016). Short-term and long-term streamflow prediction by using ‘wavelet-gene expression’
programming approach. ISH Journal of Hydraulic Engineering, 22(2), 148–162.
Karpatne, A., Atluri, G., Faghmous, J. H., Steinbach, M., Banerjee, A., Ganguly, A., et al. (2017). Theory-guided data science: A new
paradigm for scientific discovery from data. IEEE Transactions on Knowledge and Data Engineering, 29(10), 2318–2331.
Karpatne, A., Ebert-Uphoff, I., Ravela, S., Babaie, H. A., & Kumar, V. (2018). Machine learning for the geosciences: Challenges and
opportunities. IEEE Transactions on Knowledge and Data Engineering, 31(8), 1544–1554.
Kavetski, D., & Fenicia, F. (2011). Elements of a flexible approach for conceptual hydrological modeling: 2. Application and experimental
insights. Water Resources Research, 47, W11511. https://doi.org/10.1029/2011WR010748
Keijzer, M., & Babovic, V. (2002). Declarative and preferential bias in GP-based scientific discovery. Genetic Programming and Evolvable
Machines, 3, 41–79.
Keijzer, M., Babovic, V., Ryan, C., o'Neill, M., & Cattolico, M. (2001). Adaptive logic programming. In Proceedings of the 3rd annual
conference on genetic and evolutionary computation, San Francisco, USA, 42–49.
Khu, S. T., Liong, S. Y., Babovic, V., Madsen, H., & Muttil, N. (2001). Genetic programming and its application in real-time runoff forecasting.
Journal of the American Water Resources Association, 37(2), 439–451.
Koza, J. R. (1992). Genetic programming: on the programming of computers by means of natural selection, vol. 1: MIT press.
Krause, P., Boyle, D., & Bäse, F. (2005). Comparison of different efficiency criteria for hydrological model assessment. Advances in
Geosciences, 5, 89–97.
Ley, R., Hellebrand, H., Casper, M. C., & Fenicia, F. (2016). Comparing classical performance measures with signature indices derived
from flow duration curves to assess model structures as tools for catchment classification. Hydrology Research, 47(1), 1–14.
Mast, M. A., & Turk, J. T. (1999). Environmental characteristics and water quality of Hydrologic Benchmark Network stations in the eastern
united states. U.S. Geological Survey Circular, 1173A, 158.
McClelland, J. L., & Rumelhart, D. E. (1986). Parallel distributed processing: Explorations in the microstructure of cognition. Cambridge,
Massachusetts: The MIT Press.
Mcgovern, A., Lagerquist, R., Gagne, D. J., Jergensen, G. E., Elmore, K. L., Homeyer, C. R., & Smith, T. (2019). Making the black box more
transparent: Understanding the physical implications of machine learning. Bulletin of the American Meteorological Society, 2175–2199.
Mehr, A. D., Nourani, V., Kahya, E., Hrnjica, B., Sattar, A. M. A., & Yaseen, Z. M. (2018). Genetic programming in water resources
engineering: A state-of-the-art review. Journal of Hydrology, 566, 643–667.
Meshgi, A., Schmitter, P., Babovic, V., & Chui, T. F. M. (2014). An empirical method for approximating stream baseflow time series using
groundwater table fluctuations. Journal of Hydrology, 519, 1031–1041.
Meshgi, A., Schmitter, P., Chui, T. F. M., & Babovic, V. (2015). Development of a modular streamflow model to quantify runoff contributions
from different land uses in tropical urban environments using genetic programming. Journal of Hydrology, 525, 711–723.
Minns, A. W., & Hall, M. J. (1996). Artificial neural networks as rainfall-runoff models. Hydrological Sciences Journal, 41(3), 399–417.
Naghibi, S. A., Pourghasemi, H. R., & Dixon, B. (2016). GIS-based groundwater potential mapping using boosted regression tree,
classification and regression tree, and random forest machine learning models in Iran. Environmental Monitoring and Assessment,
188(44), 1–27.
Nash, J. E., & Sutcliffe, J. V. (1970). River flow forecasting through conceptual models Part I—A discussion of principles. Journal of
hydrology, 10(3), 282–290.
Nayak, P. C., Satyaji Rao, Y. R., & Sudheer, K. P. (2006). Groundwater level forecasting in a shallow aquifer using artificial neural network
approach. Water Resources Management, 20, 77–90. https://doi.org/10.1007/s11269-006-4007-z
Newman, A. J., Clark, M. P., Sampson, K., Wood, A., Hay, L. E., Bock, A., et al. (2015). Development of a large-sample watershed-scale
hydrometeorological dataset for the contiguous USA: Dataset characteristics and assessment of regional variability in hydrologic model
performance. Hydrology and Earth System Sciences, 19, 209–223.
Nielsen, S. A., & Hansen, E. (1973). Numerical simulation of the rainfall-runoff process on a daily basis. Hydrology Research, 4(3), 171–190.
Nourani, V., Komasi, M., & Mano, A. (2009). A multivariate ANN-wavelet approach for rainfall-runoff modeling. Water Resources
Management, 23, 2877–2894. https://doi.org/10.1007/s11269-009-9414-5
Oyebode, O. K., & Adeyemo, J. A. (2014). Genetic programming: Principles, applications and opportunities for hydrological modelling.
World Academy of Science, Engineering and Technology International Journal of Environmental, Chemical, Ecological, Geological and
Geophysical Engineering, 8(6), 348–354.
Perrin, C., Michel, C., & Andréassian, V. (2003). Improvement of a parsimonious model for streamflow simulation. Journal of Hydrology,
279(1), 275–289.
Quinlan, J. R. (1992). Learning with continuous classes, 5th australian joint conference on artificial intelligence (vol. 92, pp. 343–348).
R Core Team (2013). R: A language and environment for statistical computing, R Foundation for Statistical Computing, Vienna, Austria.
Retrieved from http://www.R-project.org/
Rahmani, M. A., & Zarghami, M. (2015). The use of statistical weather generator, hybrid data driven and system dynamics models for water
resources management under climate change. Journal of Environmental Informatics, 25(1), 23–35.
Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2017). Physics informed deep learning (Part II): Data-driven discovery of nonlinear partial
differential equations. arXiv preprint arXiv:1711.10566.
Renard, B., Kavetski, D., Leblois, E., Thyer, M., Kuczera, G., & Franks, S. W. (2011). Toward a reliable decomposition of predictive
uncertainty in hydrological modeling: Characterizing rainfall errors using conditional simulation. Water Resources Research, 47, W11516.
https://doi.org/10.1029/2011WR010643
Rudy, S. H., Brunton, S. L., Proctor, J. L., & Kutz, J. N. (2017). Data-driven discovery of partial differential equations. Science Advances,
3(4), p.e1602614.
Savic, D., & Khu, S.-T. (2005). Evolutionary computing in hydrological sciences. In: Anderson, MG (ed.), Encyclopedia of Hydrological
Sciences (pp. 331–348). New York: Wiley.
Sedighi, F., Vafakhah, M., & Javadi, M. R. (2016). Rainfall-runoff modeling using support vector machine in snow-affected watershed.
Arabian Journal for Science and Engineering, 41(10), 4065–4076.
Selle, B., & Muttil, N. (2011). Testing the structure of a hydrological model using genetic programming. Journal of Hydrology, 397(1), 1–9.
Shortridge, J. E., Guikema, S. D., & Zaitchik, B. F. (2016). Machine learning methods for empirical streamflow simulation: A comparison
of model accuracy, interpretability, and uncertainty in seasonal watersheds. Hydrology and Earth System Sciences, 20(7), 2611–2628.
Singh, A. P., Medida, S., & Duraisamy, K. (2017). Machine-learning-augmented predictive modeling of turbulent separated flows over
airfoils. AIAA Journal, 55(7), 2215–2227.
Snauffer, A. M., Hsieh, W. W., Cannon, A. J., & Schnorbus, M. A. (2018). Improving gridded snow water equivalent products in British
Columbia, Canada: Multi-source data fusion by neural network models. The Cryosphere, 12(3), 891–905.

CHADALAWADA ET AL. 22 of 23
 Water Resources Research 10.1029/2019WR026933

Solander, K. C., Bennett, K. E., Fleming, S. W., & Middleton, R. S. (2019). Estimating hydrologic vulnerabilities to climate change using
simulated historical data: A proof-of-concept for a rapid assessment algorithm in the colorado river basin. Journal of Hydrology: Regional
Studies, 26, 100,642.
Sugawara, M. (1979). Automatic calibration of the tank model/l'étalonnage automatique d'un modèle à cisterne. Hydrological Sciences
Journal, 24(3), 375–388.
Sun, Y., Babovic, V., & Chan, E. S. (2012). Artificial neural networks as routine for error correction with an application in singapore regional
model. Ocean Dynamics, 62(5), 661–669.
van Esse, W., Perrin, C., Booij, M., Augustijn, D., Fenicia, F., Kavetski, D., & Lobligeois, F. (2013). The influence of conceptual model
structure on model performance: A comparative study for 237 french catchments.
Vapnik, V. N. (1992). Principles of risk minimization for learning theory. Advances in Neural Information Processing Systems, 4, 831–838.
Vapnik, V. N. (1995). The nature of statistical learning theory, vol. 2. New York: Springer Verlag.
Vermeulen, D., & Van Niekerk, A. (2017). Machine learning performance for predicting soil salinity using different combinations of
geomorphometric covariates. Geoderma, 299, 1–12.
Vitolo, C. (2015). Exploring data mining for hydrological modelling.
Wagener, T., Boyle, D. P., Lees, M. J., Wheater, H. S., Gupta, H. V., & Sorooshian, S. (2001). A framework for development and application
of hydrological models. Hydrology and Earth System Sciences Discussions, 5(1), 13–26.
Wang, N., Zhang, D., Chang, H., & Li, H. (2019). Deep learning of subsurface flow via theory-guided neural network. In Presentation at the
american Geophysical Union Fall meeting, San Francisco, CA.
Weiler, M., & Beven, K. (2015). Do we need a community hydrological model? Water Resources Research, 51, 7777–7784. https://doi.org/
10.1002/2014WR016731
Xu, T., Longyang, Q., Tyson, C., Zeng, R., Neilson, B. T., & Tarboton, D. G. (2019). Hybrid physically-based and deep learning modeling of
a snow dominated mountainous karst watershed. In Presentation at the American Geophysical Union fall meeting, San Francisco, CA.
Yiu, T. (2019). Understanding random forest, how the algorithm works and why it is so effective. https://towardsdatascience.com/
understanding-random-forest-58381e0602d2
Yu, X., Liong, S. Y., & Babovic, V. (2004). EC-SVM approach for real-time hydrologic forecasting. Journal of Hydroinformatics, 6(3), 209–223.

CHADALAWADA ET AL. 23 of 23