- Email: [email protected]
Flexibility on storage-release based distributed hydrologic modeling with object-oriented approach
Journal of Hydrology 540 (2016) 17–25
Contents lists available at ScienceDirect
Journal of Hydrology journal homepage: www.elsevier.com/locate/jhydrol
Technical Note
Flexibility on storage-release based distributed hydrologic modeling with object-oriented approach Kwangmin Kang a,b,⇑, Venkatesh Merwade c, Jong Ahn Chun d, Dennis Timlin a a
Crop Systems and Global Change Laboratory, Agricultural Research Service (ARS), U.S. Department of Agriculture (USDA), Beltsville, MD 20705, United States School of Agricultural and Natural Science, University of Maryland, College Park, MD 21658, United States c School of Civil Engineering, Purdue University, West Lafayette, IN 47907, United States d Climate Research Department, APEC Climate Center, Busan 48058, Republic of Korea b
a r t i c l e
i n f o
Article history: Received 1 April 2016 Received in revised form 31 May 2016 Accepted 2 June 2016 Available online 9 June 2016 This manuscript was handled by Geoff Syme, Editor-in-Chief Keywords: Modeling framework Object-oriented Flexibility Extensibility STORE DHM
s u m m a r y With the availability of advanced hydrologic data in public domain such as remote sensed and climate change scenario data, there is a need for a modeling framework that is capable of using these data to simulate and extend hydrologic processes with multidisciplinary approaches for sustainable water resources management. To address this need, a storage-release based distributed hydrologic model (STORE DHM) is developed based on an object-oriented approach. The model is tested for demonstrating model flexibility and extensibility to know how to well integrate object-oriented approach to further hydrologic research issues, e.g., reconstructing missing precipitation in this study, without changing its main frame. Moreover, the STORE DHM is applied to simulate hydrological processes with multiple classes in the Nanticoke watershed. This study also describes a conceptual and structural framework of objectoriented inheritance and aggregation characteristics under the STORE DHM. In addition, NearestMP (missing value estimation based on nearest neighborhood regression) and KernelMP (missing value estimation based on Kernel Function) are proposed for evaluating STORE DHM flexibility. And then, STORE DHM runoff hydrographs compared with NearestMP and KernelMP runoff hydrographs. Overall results from these comparisons show promising hydrograph outputs generated by the proposed two classes. Consequently, this study suggests that STORE DHM with an object-oriented approach will be a comprehensive water resources modeling tools by adding additional classes for toward developing through its flexibility and extensibility. Ó 2016 Elsevier B.V. All rights reserved.
1. Introduction An object-oriented approach to hydrologic modeling increases model flexibility and reduces efforts when adapting the model for a new application, area and algorithm. Rather than replacing old code that already works, the model code can be extended using the object-oriented characteristic of inheritance (Kiker et al., 2006). An object-oriented approach allows building an incremental model that can be adapted to different watershed conditions (Wang et al., 2005). In spite of many advantages, object-oriented approach has found only limited applications in hydrologic modeling (Band et al., 2000; Kralisch et al., 2005). Band et al. (2000) introduced an object-oriented approach to simulate hydrologic processes, specifically infiltration excess overland flow and described a spatial
⇑ Corresponding author at: Crop Systems and Global Change Laboratory, Agricultural Research Service (ARS), U.S. Department of Agriculture (USDA), Beltsville, MD 20705, United States. http://dx.doi.org/10.1016/j.jhydrol.2016.06.009 0022-1694/Ó 2016 Elsevier B.V. All rights reserved.
object-oriented framework for modeling watershed system to include hydrological and ecosystem fluxes. Chen and Beschta (1999) developed 3-dimensional distributed hydrological model OWLS (the Object Watershed Link Simulation model) for dynamic hydrologic simulation and applied it to the Bear Brook watershed in Maine. Garrote and Becchi (1997) employs object oriented programming techniques with distributed hydrologic models for real-time flood forecasting. Boyer et al. (1996) presents an object-oriented method to simulate a rainfall-discharge relationship using a lumped hydrologic model. The above applications used object-oriented approach and achieved reasonable results for hydrologic simulations. However, object-oriented approach is not enough comprehensively discussed in the hydrologic literature and no general guideline exists for implementing them in hydrologic models (Wang et al., 2005; Kiker et al., 2006; Kang and Merwade, 2011). Realization of this approach requires a modular structure, and it allows different sub-models to be interconnected depending on the
18
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
hydrologic system. Another important aspect that must be considered in new development of hydrologic models is creating software elements that can be adapted in future projects. If a modular structure provides reusable components, both regarding development time and reliability of the software produced, the modular structure will be a useful water resources modeling platform (Goodchild et al., 1993). This also will reduce development and maintenance cost for the project. In recent years, object-oriented based hydrologic models are increasingly used in water resources research, and also the modeling paradigm in water resources is changing to the object-oriented approach (Wang et al., 2005; Kang and Merwade, 2011). Creating an incremental watershed model can be made by object-oriented design methods and using an object-oriented programming language (Wang et al., 2005), and this model can be applied to various watershed conditions. Streamflow forecasting in hydrologic modeling is the most fundamental stage, but it cannot be easily realized due to its complicated physical process and time varying system. To implement these troublesome in streamflow forecasting, hydrologic inputs in each sub-basin mechanism under the entire basin are needed (Kang and Merwade, 2011). However, researchers are confronted with the difficulty to get realistic hydrologic inputs because inputs and outputs in hydrologic model have temporal variability and non-linear response. Several studies investigated on assessing the impacts of hydrologic inputs for hydrologic simulation (Ogden and Julien, 1993). Kang and Merwade (2014) addressed that even though precipitation input is main driver in hydrologic simulation, utilizing an effective precipitation uncertainty in hydrologic simulation is still lack of research. The primary objective of this study is to investigate how to make an efficient model frame that can raise hydrologic modeling extensibility. The motivation to address this question is that large number of advanced technology data and multidisciplinary research approaches lead to additional computation platform that trigger wasting development cost. The STORE DHM (Kang and Merwade, 2011) is used for describing the general idea of objectoriented hydrologic model framework. Expandable user options for reconstructing missing precipitation are also described to show an object-oriented flexibility in the STORE DHM. The following section contains a general description of the study area in this research. Specific object-oriented approach in hydrology is explained in Section 3. Section 4 proposes STORE DHM object classes for hydrologic simulation. Statistic methodologies, nearest neighborhood regression (NNR) and Kernel Function (KF), are described for estimating missing precipitation in Section 5. Finally, Sections 6 and 7 present and discuss the results and conclusions, respectively.
2. Study area The Nanticoke watershed is located in the Chesapeake Bay which spans more than 165,759 km2. The bay connects and encompasses parts of six states – Maryland, Delaware, New York, Pennsylvania, Virginia and West Virginia. It has a great variety of ecological species, and almost 18 million people live in the Chesapeake Bay (Fig. 1a). The Nanticoke watershed covers an area of 2142 km2 and is characterized by an elevation ranging from 11 to 85 m above sea level (Fig. 1b), while the main river course is 10.7 km long. This region has typically four seasons with an average temperature of 1.5 °C in winter and 28 °C in summer. The average annual precipitation in the watershed was about 1080 mm from 2000 to 2015. Rainfall occurs normally from December to May (wet season), while during the dry season, from June to November, rainfall is generally concentrated in a few events. The general rainfall events
in all seasons are unevenly distributed and have high intensity for a short time. These rainfall characteristics exert strong influence on the flow regimes and erosion, subsequently on sediment and nutrient delivery. The hourly precipitation data and 15 min streamflow data for the study site were obtained from the National Climate Data Center and the USGS Instantaneous Data Archive, respectively. The streamflow values for four events include both base flow and surface runoff. This study used the straight line base flow separation method for retrieving surface runoff hydrographs from streamflow. Land cover types in the Nanticoke watershed are presented as 49% agriculture, 38% forest and wetland, 9% water, and 4% urban development.
3. Object-orientated in hydrology According to Bian (2007), object orientation involves three levels of abstraction: object oriented analysis, object oriented design, and object oriented programming. Object oriented analysis involves conceptual representation of the world including the facts and relationships about a situation. In hydrology, this would mean the conceptual representation of a watershed as a set of objects including streams and corresponding catchments. Object oriented design uses the conceptual representation from object oriented analysis to create a formal model of objects, their properties, events, and relationships. Object oriented programming involves the implementation of objects and their events to accomplish a certain task. Object orientation relies on two basic principles: encapsulation and composition. Encapsulation considers that the world is composed of objects, and that each object has an identity, properties and behavior. The properties of an object are defined by its attributes (e.g., length, area), and the behavior is represented by methods. While the value of an attribute can define the state of an object, a method can change the state of an object, and that change is referred to as an event. For example, a river object will have properties such as length and slope, methods such as RouteFlow and ComputeStorage, and routing a hydrograph through the river (by using RouteFlow method) is an event. The principle of composition describes how objects are related through relationships including inheritance, aggregation and association. In object orientation, all objects belong to object classes, and all classes are hierarchal. A sub-class is a kind of this own super-class through inheritance and inherits all properties and methods from the super class, but also may have its own additional properties and methods. An object can also be a part of another object through aggregation, and can simultaneously maintain relationships with other objects through association. For example, an AlluvialRiver class can be a sub-class of River super class (inheritance), a River class can be a part of RiverNetwork class (aggregation) and River class is related with Watershed class through streamflow (relationship). Past studies that used object orientation for hydrologic modeling include Whittaker et al. (1991) who used object-oriented approach to model infiltration excess overland flow. Boyer et al. (1996) used object oriented approach to develop a lumped rainfall–runoff model and used object orientation to combine remote sensing and hydrologic data to develop a forecast model. Garrote and Becchi (1997), Band et al. (2000), and Wang et al. (2005) proposed object oriented frameworks for modeling hydrologic processes at watershed scale. Most of these studies used object orientation to model hydrologic processes using the concepts of inheritance and aggregation. Recently, Richardson et al. (2007) proposed a prototype geographically based object framework for linking hydrologic and biochemical processes in the sub-surface. However, the process objects were loosely coupled with geographic objects, thus there is leaving an opportunity for a tightly coupled geographically based object oriented modeling.
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
19
(a)
(b) Fig. 1. Study area location: (a) represents Chesapeake Bay; and (b) represents Nanticoke Watershed including weather station and stream gauge.
Relatively recent advances in GIS have enabled the adaption of object orientation in storing and handling geospatial and temporal hydrologic data in research and practice. For example, Arc Hydro (Maidment, 1993) uses an object oriented approach to represent hydrologic environment through feature, object and relationship classes within a geodatabase. In ArcHydro, a HydroEdge (stream) is sub class of generic Polyline super class (inheritance), and is a part of HydroNetwork (aggregation). HydroEdge is related to Watershed (which itself is a sub-class of Polygon super class) through a common identifier (HydroID). Thus, ArcHydro uses object orientation to develop a physical representation of hydrography by using GIS objects. Thus, by knowing the HydroID of any geographic feature, it is possible to trace the flow of water by using points, lines, and polygons at multiple scales including at continental scale. The National Hydrography Dataset (NHD) available for the entire United States from the United States Geological Survey (USGS) also uses the object oriented (or geodatabase) design to provide data to its users.
The geodatabase approach to hydrology data overcomes several practical issues which are associated with storing and handling heterogeneous multi-scale data by providing a relational data model. Besides overcoming the data issues, the geodatabase approach provides an opportunity to exploit the potential of object oriented approach to overcome the limitations of scale and parameterization in distributed modeling of hydrologic processes. For example, if a polygon representing a watershed in GIS is treated as a hydrologic object that has some properties (e.g. area and slope) and methods (to compute runoff and route flow), then multiple watersheds can be linked and executed in parallel to scale-up the modeling domain from one single watershed to larger (national or continental) scales. Similarly, the availability of increasing GIS layers to represent soil, land use and topography at multiple scales, can enable parameterization of hydrologic processes (or watershed methods) through GIS tools, which is not possible with most existing models that do not explicitly work within a GIS environment. This research builds on past studies to create a prototype tightly
20
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
hydrologic processes to create excess rainfall and runoff hydrograph by implementing specific sub-classes (ExRain and Hydrograph) as shown in Fig. 2. ExRain implements specific techniques such as SCS curve number (through SCS sub-class) and Green– Ampt (through GreenAmpt sub-class) to compute excess rainfall using the rainfall input. Hydrograph class implements specific techniques such as storage release to compute runoff hydrograph from excess rainfall. TopoGrid implements sub-classes to create terrain attributes such as flow direction, flow accumulation, stream network and catchment by using topography data (DEM). HydroArea can work on flow transformation by implementing specific techniques such as SCS dimensionless unit hydrograph (through SCSUnit sub-class) and Clark unit hydrograph (through ClarkUnit sub-class) for vector data. Also, HydroLine can work on river routing by implementing specific techniques such as Kinematic wave river routing (through KinematicWave sub-class) and Muskingum river routing (through Muskingum sub-class). HydroTable implements three tables: TSTable, ParameterTable and PairTable. TSTable class processes time series table (e.g., rainfall and streamflow time series); ParameterTable process es parameter values linked to geographic features (e.g., Manning’s n values for different land cover types); and PairTable processes paired data such as stage-discharge rating curves.
coupled object-oriented hydrologic model in GIS to simulate hydrologic processes using geospatial inputs.
4. STORE DHM object classes The prototype modeling approach presented in this research is developed by using Visual Basic and ArcObjects within ArcGIS, and is referred to as the STORE DHM. The new conceptual hydrologic model is based on storage-release with grid travel time approach. In addition, the study uses an object-oriented programming approach which can provide useful data handling and model flexibility in hydrologic applications. An object-oriented hydrologic model framework is implemented in ArcGIS by creating hydrologic modeling objects as shown in Fig. 2. The hydrologic modeling object framework implements both principles of aggregation (represented with diamond) and inheritance (represented with triangle arrow) as shown in Fig. 2. HydroShed is the highest level class that includes the following six classes: (1) HydroGrid (to process gridded hydrologic information such as topography and rainfall), (2) ParameterGrid (to process gridded hydrologic parameters such as Mannings n), (3) HydroArea (to process vector hydrologic data for lakes and rivers), (4) HydroCatchment (to process vector hydrologic data for catchments or sub-watersheds), (5) HydroLine (to process vector hydrologic data for streams), and (6) HydroTable (to process tabular data). As displayed in Fig. 2 and Table 1, those classes dealing with raster, vector and tables are implemented in this research. ProcessGrid (to implement hydrologic processes) and TopoGrid (to implement terrain processes) are two sub-classes of the HydroGrid class. ProcessGrid can work on gridded data. It implements
*
HydroArea
1 1
-Attribute
Precipitation is the most crucial part in hydrologic modeling. Malfunctions in running hydrologic modeling can occur due to discontinuous time series precipitation inputs (Lee and Kang, 2015). As a part of practical test with flexibility of object-oriented in the
HydroShed
HydroCatchment
-Attribute
-Attribute
+Method()
+Method()
5. Reconstructing missing precipitation
1
1
*
+Method()
1 1
HydroLine
*
HydroGrid
-Attribute
-Attribute
*
+Method()
+Method() *
HydroTable ProcessGrid
TopoGrid -Attribute
-Attribute
+Method()
+Method()
ExRain
*
Hydrograph
ParameterGrid
-Attribute
-Attribute
+Method()
+Method()
TSTable
ParameterTable
PairTable
-Attribute
-Attribute
-Attribute
-Attribute
-Attribute
+Method()
+Method()
+Method()
+Method()
+Method()
SCS
GreenAmpt
-Attribute
-Attribute
+Method()
+Method()
Fig. 2. Object model diagram for STORE DHM (aggregation and inheritance characteristic in object-oriented represents with diamond and arrow, respectively). ⁄main object classes.
21
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25 Table 1 Types of object classes in STORE DHM. Class
Characteristic
Type
Implementation
HydroShed
Highest
Highest
Core class
HydroGrid ParameterGrid HydroArea HydroCatchment HydroLine HydroTable
Aggregation
Sub-class Sub-class Sub-class Sub-class Sub-class Sub-class
Process Process Process Process Process Process
ProcessGrid TopoGrid ExRain Hydrograph SCS GreenAmpt TSTable ParameterTable PairTable NesrestMP KernelMP
Inheritance
Under Under Under Under Under Under Under Under Under Under Under
Implement Implement Implement Implement
HydroCatchment
*
1
HydroGrid HydroGrid ProcessGrid ProcessGrid ExRain ExRain HydroTable HydroTable HydroTable TSTable TSTable
HydroArea
-Attribute +Method() 1 1
*
1
1
1 HydroLine *
TopoGrid
-Attribute +Method()
*
-Attribute +Method()
-Attribute +Method()
hydrologic processes terrain processes excess rainfall calculation runoff calculation
HydroShed
-Attribute +Method()
HydroGrid
gridded information gridded parameters vector data for rivers vector data for catchments vector data for streams tabular data
-Attribute +Method()
ProcessGrid
KinematicWave
Muskingum
-Attribute +Method()
-Attribute +Method()
-Attribute +Method()
ParameterGrid ExRain
Hydrograph
-Attribute +Method()
GreenAmpt -Attribute +Method()
-Attribute +Method()
RadarBias
HydroTable
-Attribute +Method()
*
SCS
TSTable
-Attribute +Method()
-Attribute +Method()
NearestMP -Attribute +Method()
-Attribute +Method()
ParameterTable -Attribute +Method()
*
-Attribute +Method()
PairTable -Attribute +Method()
KernelMP -Attribute +Method()
Fig. 3. Updated object model diagram for STORE DHM without changing its main frame (Red boxes represent updated classes). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
STORE DHM, this study proposed reconstructing missing precipitation data in hydrologic modeling: (1) nearest neighborhood regression (the NearestMP class); and (2) kernel function for missing precipitation (the KernelMP class) under the TSTable class (Fig. 3).
5.1. Nearest neighborhood regression A nearest neighborhood regression (NNR) typically uses a Markov chain to determine the missing precipitation based on proba-
22
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
5.2. Kernel function
Table 2 Manning’s n for Nanticoke watershed. Land use
Range
Initial value
After calibration
Agricultural Forest Developed Water
0.030–0.500 0.035–0.160 0.011–0.035 0.025–0.033
0.220 0.090 0.020 0.030
0.056 0.063 0.033 0.028
bility distributions. NNR so far has been successfully applied for estimating precipitation. NNR also has the advantage of computation efficiency (Sharif and Burn, 2006). It calculates missing precipitation with cumulative probability distribution of nearest neighborhood values and then selects the rank of the distance for each of the nearest value. A missing precipitation is estimated by NNR as follows: P 1X M¼ xi wi ; P i¼1
i¼1 Ri
P 1X xi Kðui Þ; P i¼1
ð3Þ
ui ¼
Ni ; 0:5P þ 1
ð4Þ
ð1Þ
P P Ni ¼ ; . . . ; ; 2 2
;
ð2Þ
where M is missing precipitation, P is the number of the nearest neighborhood, and xi is the sample data at ith nearest value and wi is the weighting for close neighbor at ith. R is the Pearson correlation coefficient that is a measure of the linear relationship between two random variables. The R values are converted into weights by using the weighting formula that is suggested by Dumedah and Caulibaly (2011). Estimation of the missing variable is computed as the weighted sum, and this procedure is operating in NearestMP class under TSTable class (Fig. 3).
6. Results 6.1. Update object-oriented framework Fig. 3 represents how to implement user optional precipitation inputs with object-oriented approach in the STORE DHM. Near-
25
7 Obs. Calibrated Sim.
Runoff (m3/s)
20 15 10
6
5
Obs. Sim.
5 4 3 2 1
0
0 0
0
100
20
40
15min/unit
15min/unit
(a)
(b)
60
20
40 35
Runoff (m3/s)
ð5Þ
where ui is the Nth nearest values which correspond to xi (positive means the right side and negative means the left side). K (ui) represents kernel function at ui. Estimating a missing precipitation data with KF, KernelMP, is implementing under the TSTable class (Fig. 3).
Runoff (m3/s)
wi ¼ PP
M¼
Obs. Sim.
30 25
Obs. Sim.
15
Runoff (m3/s)
Ri
While a NNR method is widely used in filling the missing precipitation data for hydrologic modeling, it has shown troublesome statistic theorem due to event based precipitation data having limited nearest sample data to make a normal distribution (Lee and Kang, 2015). General mechanism of weather data during short term rainfall event would not follow normal distribution trend. It also would not have enough nearest sample data around missing precipitation. As using of neighborhood mean value in kernel function (KF), it can overcome troublesome statistic theorem experienced in the NNR method (Lee and Kang, 2015). Missing precipitation data are calculated by the following KF formula:
10
20 15 10
5
5 0
0 0
100
200
0
50
100
15min/unit
15min/unit
(c)
(d)
150
Fig. 4. Hydrographs with STORE DHM simulations. (a), (b), (c) and (d) present Event 1, 2, 3 and 4, respectively.
23
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25 Table 3 STORE DHM results for Nanticoke watershed. Event
STORE DHM simulations 3
1 2 3 4
Observed runoff 2
Peak flow (m /s)
Time to peak (h)
R
22.48 5.9 34.76 17.6
7.50 3.75 19.75 11.50
0.95 0.94 0.94 0.95
ENS
Peak flow (m3/s)
Time to peak (h)
0.92 0.93 0.93 0.82
22.78 6.14 35.22 17.44
6.75 3.5 22.50 11.75
7
25
6 Event 1
Runoff (m3/s)
Runoff (m3/s)
20 15 10
5
Event 2
4 3 2
5 1 0
0 0
0
100
40
20
35
18
Runoff (m3/s)
Runoff (m3/s)
40
16
Event 3
30
20
60
15min/unit
15min/unit
25 20 15 10
Event 4
14 12 10 8 6 4
5
2
0
0 0
100
200
15min/unit Observed
0
100
15min/unit Simulation with missing precipitation
Fig. 5. STORE DHM simulations with synthetic missing precipitations.
estMP and KernelMP classes are created by an inheritance characteristic of object-oriented model and the two classes implement the estimation of missing precipitation values under the TSTable class. An object-oriented data class such as TSTable is a bridge between data handling, e.g., NearestMP and KernelMP, and calculating hydrologic processes (ProcessGrid). The first step executing NearestMP and KernelMP is to load time series rainfall data in the TSTable. Then, the TSTable will automatically scan for whether time series data is continuous or discontinuous. If the data has any missing rainfall data, it can be transferred to the NearestMP or KernelMP to reconstruct missing data depending on user selection. The NearestMP and KernelMP illustrates the ‘‘is-kind-of” relationships to the TSTable throughout inheritance characteristic in object-oriented framework. Each entity and attribute in NearestMP and KernelMP is associated with HydroTable which is an upper level class for all tabular data. For example, KF equations are associated with TSTable to generate time series precipitation data and it can also be associated with the HydroTable as an input table for the ProcessGrid. Likewise, KinematicWave and RadarBias classes (Fig. 3), updated classes in the STORE DHM, are implemented for runoff routing under the HydroLine and radar precipitation bias correction under the ProcessGrid, respectively (Kang
and Merwade, 2014). But, these classes are not presented their internal methodologies in this paper to avoid inconsistency of the theoretic topic. 6.2. STORE DHM simulation Instead of using theoretical calculation for getting geographical parameters, the STORE DHM used spatially distributed grid data in ArcGIS. Thus, calibration simulation only for manning’s n is executed for the first modeling step. After getting the calibrated manning’s n from the first event (Table 2), the model is applied to three additional three events at Nanticoke watershed. Fig. 4 presents hydrograph results from application of the model. Qualitative and quantitative comparison of each model hydrograph, peak flow and time to peak, with observed hydrograph show reasonable agreement. Coefficient of determination (R2) and Nash Sutcliffe efficiency coefficient (ENS) are used for model performance measures, and STORE DHM showed 0.94 average R2 and 0.90 average ENS. Model simulations with calibrated manning’s show that the peak flow and time-to-peak are underestimated by 6.4% and 5.7%, respectively. Table 3 presents a summary of model results and error statics for STORE DHM simulations. Overall, comparison
24
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
30
7
25
6 Event 2
5
Runoff (m3/s)
Runoff (m3/s)
Event 1 20 15 10
4 3 2
5
1
0
0 0
50
100
0
150
20
40
15min/unit 40
20
35
18
Runoff (m3/s)
Runoff (m3/s)
16
Event 3
30
60
15min/unit
25 20 15 10
Event 4
14 12 10 8 6 4
5
2
0
0 0
100
0
200
50
100
15min/unit Observed
150
15min/unit Simulation with NNR
Simulation with KN
Fig. 6. STORE DHM simulation with reconstructed missing precipitations. Table 4 Validation results using missing and reconstructed precipitation. Event
1 2 3 4
Simulation with missing prec.
Simulation with NNR
Simulation with KN
Peak flow (m3/s)
Time to peak (h)
R2
ENS
Peak flow (m3/s)
Time to peak (h)
R2
ENS
Peak flow (m3/s)
Time to peak (h)
R2
ENS
18.22 5.92 29.17 14.27
6.75 3.75 19.75 12.00
0.81 0.91 0.91 0.95
0.86 0.87 0.81 0.64
25.57 6.58 36.35 17.58
7.00 3.75 19.50 12.50
0.82 0.92 0.91 0.94
0.85 0.88 0.88 0.84
23.62 8.81 31.59 16.74
7.00 3.75 19.75 11.75
0.88 0.93 0.91 0.95
0.91 0.93 0.86 0.86
12
of the model output and observed data looks reasonable for the Nanticoke watershed (Fig. 4).
NNR
KN
10
Making a synthetic missing precipitation data from four selected event periods is the first step, and then STORE DHM simulations with missing precipitation are performed in Fig. 5 which are showed under estimated hydrographs comparing with observed hydrographs due to less quantity of precipitation. Fig. 6 and Table 4 presents results from application of the model with NearestMP and KernelMP classes for estimating missing precipitation in Nanticoke watershed. The results from STORE DHM with missing precipitation are compared with both NNR and KN reconstructed missing precipitation (Table 4). Results in NNR are showed 18% over estimation of the peak flow. The time-to-peak is underestimated in event 4. Fig. 7 presents scatter plots of reconstructed missing precipitation both using NNR and KF. Estimated missing precipitation with NNR showed over estimation than observation precipitation. On the other hand, estimated missing precipitation with KF showed well estimation to observation precipitation (Fig. 7).
Estimated precipitation
6.3. Additional precipitation options 8
6
4
2
0
0
2
4
6
8
10
12
Observed precipitation Fig. 7. Comparison of NNR and KN reconstructed precipitation data for Nanticoke watershed.
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
7. Conclusions This research is focused on developing a prototype objectoriented hydrologic modeling framework called the STORE DHM. This research presents an object-oriented approach using objectoriented design techniques and object-oriented language (Visual Basic) to the description and simulation of a watershed based hydrologic processes. The STORE DHM allows flexibility and extensibility to investigate future hydrologic issues without changing its main framework because it uses characteristics of inheritance and aggregation through object-oriented hydrologic approach. It would be a flexible and dynamic tool that is capable of simulating a wide array of management practices such as water contamination, supplying, flooding and drought from watershed scale to continental scale. The user necessaries have been incorporated in development of the STORE DHM to generate further investigation on multidisciplinary water resources researches. For example, nutrient contaminations from surface runoff in crop areas will be investigated by using of the Agricultural Policy Environmental Extender (APEX) model with linking class under the STORE DHM framework for making a sustainable water resources management. Model extensibility is an inescapable factor that builds up sustainable water resources management. Unlike a hierarchy model which is all handled by one core frame, the STORE DHM, composed of an objectoriented approach, can be easily updated and modified without changing the main frame. Each object can be worked individually by its attribute and method, and also it can be part of an entire model procedure through inheritance and aggregation characteristics. Furthermore, inheritance and aggregation of object-oriented approach in the STORE DHM is contributing to save the unnecessary computation demand, and also enhances modeling efficiency. The STORE DHM framework involves computing rainfall, volumetric flow rate, and travel time to the basin outlet by combining steady state uniform flow approximation with Manning’s equation (Kang and Merwade, 2011). While other travel time based distributed hydrologic models that route the flow from each cell to the watershed outlet along a common flow path at each time step (which can lead to improper water balance), the STORE DHM considers the flow contribution of neighboring cells by using continuity equation with storage release concept (which can lead to proper water balance). Expendable user precipitation options, NearestMP and KernelMP, are used for evaluating the STORE DHM flexibility with object-oriented approach. Lee and Kang (2015) used SWAT (Soil Water Assessment Tools) to evaluate better statistical scheme for reconstructing missing precipitation and they found that SWAT simulation with KF approach for reconstructing missing precipitation showed a better performance than simulation with NNR approach in Imha watershed. Likewise, STORE DHM simulation through the KernelMP class shows better simulation accuracy than simulation with the NearestMP class. Overall results from these comparisons show promising hydrograph outputs generated by the proposed two classes. These simulations are not always true for different events and areas. Therefore, additional object classes and study areas for the wide array of time series precipitation researches are necessary in the STORE DHM for further research. As the first part of the broader sustainable water resources modeling framework, this research develops the grid based hydrologic model for hydrologic modeling as modular development. The modeling framework presented in this research is operated within ArcGIS environment such that all the steps from extracting information from geospatial data to running model simulations are exe-
25
cuted in an ArcGIS environment. The approach from data to model output within a single environment is attractive from a practical point of view. Expanding of the STORE DHM will be essential for robustness of the hydrologic model. For example, linkage between STORE DHM and water quality model (e.g., SWAT) can predict effect of storm event on water quality assessment. Similarly, linkage between GCMs (General Circulation Models), which are the most advanced tools for estimating future climate change, and GHISMO will provide a foundation to assess the impact of climate change on rainfall-runoff prediction. The conceptualization and characterization of this coupling strategy can be extended to a sustainable water resources management and decision supporting tool. It the prototype of the object-oriented hydrologic framework, the STORE DHM, is successfully used to develop a coupled open source and platform for seamless hydrologic components, it may lead to change a hydrologic modeling paradigm to objectoriented approach through its flexible modeling schemes. Acknowledgements We would like to the editor (Dr. Geoff Syme) for constructive suggestion on previous STORE DHM paper. The first author would also like to thank proofreading from Darae Kang. References Band, L.E., Tague, C.L., Brun, S.E., Tenenbaum, D.E., Frernandez, R.A., 2000. Modeling watersheds as spatial object hierarchies: structure and dynamics. Trans. GIS 4 (3), 181–196. Bian, L., 2007. Object-oriented representation of environmental phenomena: is everything best represented as an object. Ann. Assoc. Am. Geogr. 97 (2), 267– 281. Boyer, J.F., Berkhoff, C., Servat, E., 1996. Object-oriented programming for a simulation of the rainfall-discharge relationship. In: Balkema, A.A. (Ed.), Proceedings of Hydroinformatics 96. Balkema, A.A. Zurich, Switzerland, pp. 299–305. Chen, H., Beschta, R., 1999. Dynamic hydrologic simulation of the Bear Crook Watershed in Maine (BBWM). Environ. Monit. Assess. 55, 53–59. Dumedah, G., Caulibaly, P., 2011. Evaluation of statistical methods for infillinf missing values in high-resolution soil moisture data. J. Hydrol. 400 (1–2), 95– 102. Garrote, L., Becchi, I., 1997. Object-oriented software for distributed rainfall-runoff models. J. Comput. Civ. Eng. 11, 190–194. Goodchild, M., Parks, B., Steyaert, L., 1993. Environmental Modeling with GIS. Oxford University Press, Oxford. Kang, K., Merwade, V., 2011. Development and application of a storage-release based distributed hydrologic modeling using GIS. J. Hydrol. 403, 1–13. Kang, K., Merwade, V., 2014. The effect of spatially uniform and non-uniform precipitation bias correction methods on improving NEXRAD rainfall accuracy for distributed hydrologic modeling. Hydrol. Res. 45 (1), 23–42. Kiker, G.A., Clark, D.J., Martinez, C.J., Schulz, R.E., 2006. A java based, object-oriented modeling system for southern African hydrology. Trans. ASABE 49 (5), 1419– 1433. Kralisch, S., Krause, P., David, O., 2005. Using the object modeling system for hydrological model development and application. Adv. Geosci. 4, 75–81. Lee, H., Kang, K., 2015. Interpolation of missing precipitation data using kernel estimations for hydrologic modeling. Adv. Meteorol. 935868, 1–13. Maidment, D.R., 1993. GIS and hydrologic modeling. In: Goodchild, M., Parks, B., Steyaert, L. (Eds.), Environmental Modeling with GIS. Oxford University Press, New York, USA. Ogden, F.L., Julien, P.Y., 1993. Runoff sensitivity to temporal and spatial rainfall variability at runoff plane and small basin scales. Water Resour. Res. 29 (8), 2589–2597. Richardson, M.C., Branfireun, B.A., Robinson, V.B., Graniero, P.A., 2007. Towards simulating biogeochemical hot spots in the landscape: a geographic objectbased approach. J. Hydrol. 342 (1–2), 97–109. Sharif, M., Burn, D.H., 2006. Simulating climate change scenarios using an improved K-nearest neighbor model. J. Hydrol. 325, 179–196. Wang, J., Hassett, J.M., Endreny, T.A., 2005. An object oriented approach to the description and simulation of watershed scale hydrologic processes. Comput. Geosci. 31 (4), 425–435. Whittaker, A.D., Wolfe, M.L., Godbole, R., Van Alem, G.J., 1991. Object-oriented modeling of hydrologic processes. Appl. Nat. Resour. Manage. 5 (4), 49–58.
Contents lists available at ScienceDirect
Journal of Hydrology journal homepage: www.elsevier.com/locate/jhydrol
Technical Note
Flexibility on storage-release based distributed hydrologic modeling with object-oriented approach Kwangmin Kang a,b,⇑, Venkatesh Merwade c, Jong Ahn Chun d, Dennis Timlin a a
Crop Systems and Global Change Laboratory, Agricultural Research Service (ARS), U.S. Department of Agriculture (USDA), Beltsville, MD 20705, United States School of Agricultural and Natural Science, University of Maryland, College Park, MD 21658, United States c School of Civil Engineering, Purdue University, West Lafayette, IN 47907, United States d Climate Research Department, APEC Climate Center, Busan 48058, Republic of Korea b
a r t i c l e
i n f o
Article history: Received 1 April 2016 Received in revised form 31 May 2016 Accepted 2 June 2016 Available online 9 June 2016 This manuscript was handled by Geoff Syme, Editor-in-Chief Keywords: Modeling framework Object-oriented Flexibility Extensibility STORE DHM
s u m m a r y With the availability of advanced hydrologic data in public domain such as remote sensed and climate change scenario data, there is a need for a modeling framework that is capable of using these data to simulate and extend hydrologic processes with multidisciplinary approaches for sustainable water resources management. To address this need, a storage-release based distributed hydrologic model (STORE DHM) is developed based on an object-oriented approach. The model is tested for demonstrating model flexibility and extensibility to know how to well integrate object-oriented approach to further hydrologic research issues, e.g., reconstructing missing precipitation in this study, without changing its main frame. Moreover, the STORE DHM is applied to simulate hydrological processes with multiple classes in the Nanticoke watershed. This study also describes a conceptual and structural framework of objectoriented inheritance and aggregation characteristics under the STORE DHM. In addition, NearestMP (missing value estimation based on nearest neighborhood regression) and KernelMP (missing value estimation based on Kernel Function) are proposed for evaluating STORE DHM flexibility. And then, STORE DHM runoff hydrographs compared with NearestMP and KernelMP runoff hydrographs. Overall results from these comparisons show promising hydrograph outputs generated by the proposed two classes. Consequently, this study suggests that STORE DHM with an object-oriented approach will be a comprehensive water resources modeling tools by adding additional classes for toward developing through its flexibility and extensibility. Ó 2016 Elsevier B.V. All rights reserved.
1. Introduction An object-oriented approach to hydrologic modeling increases model flexibility and reduces efforts when adapting the model for a new application, area and algorithm. Rather than replacing old code that already works, the model code can be extended using the object-oriented characteristic of inheritance (Kiker et al., 2006). An object-oriented approach allows building an incremental model that can be adapted to different watershed conditions (Wang et al., 2005). In spite of many advantages, object-oriented approach has found only limited applications in hydrologic modeling (Band et al., 2000; Kralisch et al., 2005). Band et al. (2000) introduced an object-oriented approach to simulate hydrologic processes, specifically infiltration excess overland flow and described a spatial
⇑ Corresponding author at: Crop Systems and Global Change Laboratory, Agricultural Research Service (ARS), U.S. Department of Agriculture (USDA), Beltsville, MD 20705, United States. http://dx.doi.org/10.1016/j.jhydrol.2016.06.009 0022-1694/Ó 2016 Elsevier B.V. All rights reserved.
object-oriented framework for modeling watershed system to include hydrological and ecosystem fluxes. Chen and Beschta (1999) developed 3-dimensional distributed hydrological model OWLS (the Object Watershed Link Simulation model) for dynamic hydrologic simulation and applied it to the Bear Brook watershed in Maine. Garrote and Becchi (1997) employs object oriented programming techniques with distributed hydrologic models for real-time flood forecasting. Boyer et al. (1996) presents an object-oriented method to simulate a rainfall-discharge relationship using a lumped hydrologic model. The above applications used object-oriented approach and achieved reasonable results for hydrologic simulations. However, object-oriented approach is not enough comprehensively discussed in the hydrologic literature and no general guideline exists for implementing them in hydrologic models (Wang et al., 2005; Kiker et al., 2006; Kang and Merwade, 2011). Realization of this approach requires a modular structure, and it allows different sub-models to be interconnected depending on the
18
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
hydrologic system. Another important aspect that must be considered in new development of hydrologic models is creating software elements that can be adapted in future projects. If a modular structure provides reusable components, both regarding development time and reliability of the software produced, the modular structure will be a useful water resources modeling platform (Goodchild et al., 1993). This also will reduce development and maintenance cost for the project. In recent years, object-oriented based hydrologic models are increasingly used in water resources research, and also the modeling paradigm in water resources is changing to the object-oriented approach (Wang et al., 2005; Kang and Merwade, 2011). Creating an incremental watershed model can be made by object-oriented design methods and using an object-oriented programming language (Wang et al., 2005), and this model can be applied to various watershed conditions. Streamflow forecasting in hydrologic modeling is the most fundamental stage, but it cannot be easily realized due to its complicated physical process and time varying system. To implement these troublesome in streamflow forecasting, hydrologic inputs in each sub-basin mechanism under the entire basin are needed (Kang and Merwade, 2011). However, researchers are confronted with the difficulty to get realistic hydrologic inputs because inputs and outputs in hydrologic model have temporal variability and non-linear response. Several studies investigated on assessing the impacts of hydrologic inputs for hydrologic simulation (Ogden and Julien, 1993). Kang and Merwade (2014) addressed that even though precipitation input is main driver in hydrologic simulation, utilizing an effective precipitation uncertainty in hydrologic simulation is still lack of research. The primary objective of this study is to investigate how to make an efficient model frame that can raise hydrologic modeling extensibility. The motivation to address this question is that large number of advanced technology data and multidisciplinary research approaches lead to additional computation platform that trigger wasting development cost. The STORE DHM (Kang and Merwade, 2011) is used for describing the general idea of objectoriented hydrologic model framework. Expandable user options for reconstructing missing precipitation are also described to show an object-oriented flexibility in the STORE DHM. The following section contains a general description of the study area in this research. Specific object-oriented approach in hydrology is explained in Section 3. Section 4 proposes STORE DHM object classes for hydrologic simulation. Statistic methodologies, nearest neighborhood regression (NNR) and Kernel Function (KF), are described for estimating missing precipitation in Section 5. Finally, Sections 6 and 7 present and discuss the results and conclusions, respectively.
2. Study area The Nanticoke watershed is located in the Chesapeake Bay which spans more than 165,759 km2. The bay connects and encompasses parts of six states – Maryland, Delaware, New York, Pennsylvania, Virginia and West Virginia. It has a great variety of ecological species, and almost 18 million people live in the Chesapeake Bay (Fig. 1a). The Nanticoke watershed covers an area of 2142 km2 and is characterized by an elevation ranging from 11 to 85 m above sea level (Fig. 1b), while the main river course is 10.7 km long. This region has typically four seasons with an average temperature of 1.5 °C in winter and 28 °C in summer. The average annual precipitation in the watershed was about 1080 mm from 2000 to 2015. Rainfall occurs normally from December to May (wet season), while during the dry season, from June to November, rainfall is generally concentrated in a few events. The general rainfall events
in all seasons are unevenly distributed and have high intensity for a short time. These rainfall characteristics exert strong influence on the flow regimes and erosion, subsequently on sediment and nutrient delivery. The hourly precipitation data and 15 min streamflow data for the study site were obtained from the National Climate Data Center and the USGS Instantaneous Data Archive, respectively. The streamflow values for four events include both base flow and surface runoff. This study used the straight line base flow separation method for retrieving surface runoff hydrographs from streamflow. Land cover types in the Nanticoke watershed are presented as 49% agriculture, 38% forest and wetland, 9% water, and 4% urban development.
3. Object-orientated in hydrology According to Bian (2007), object orientation involves three levels of abstraction: object oriented analysis, object oriented design, and object oriented programming. Object oriented analysis involves conceptual representation of the world including the facts and relationships about a situation. In hydrology, this would mean the conceptual representation of a watershed as a set of objects including streams and corresponding catchments. Object oriented design uses the conceptual representation from object oriented analysis to create a formal model of objects, their properties, events, and relationships. Object oriented programming involves the implementation of objects and their events to accomplish a certain task. Object orientation relies on two basic principles: encapsulation and composition. Encapsulation considers that the world is composed of objects, and that each object has an identity, properties and behavior. The properties of an object are defined by its attributes (e.g., length, area), and the behavior is represented by methods. While the value of an attribute can define the state of an object, a method can change the state of an object, and that change is referred to as an event. For example, a river object will have properties such as length and slope, methods such as RouteFlow and ComputeStorage, and routing a hydrograph through the river (by using RouteFlow method) is an event. The principle of composition describes how objects are related through relationships including inheritance, aggregation and association. In object orientation, all objects belong to object classes, and all classes are hierarchal. A sub-class is a kind of this own super-class through inheritance and inherits all properties and methods from the super class, but also may have its own additional properties and methods. An object can also be a part of another object through aggregation, and can simultaneously maintain relationships with other objects through association. For example, an AlluvialRiver class can be a sub-class of River super class (inheritance), a River class can be a part of RiverNetwork class (aggregation) and River class is related with Watershed class through streamflow (relationship). Past studies that used object orientation for hydrologic modeling include Whittaker et al. (1991) who used object-oriented approach to model infiltration excess overland flow. Boyer et al. (1996) used object oriented approach to develop a lumped rainfall–runoff model and used object orientation to combine remote sensing and hydrologic data to develop a forecast model. Garrote and Becchi (1997), Band et al. (2000), and Wang et al. (2005) proposed object oriented frameworks for modeling hydrologic processes at watershed scale. Most of these studies used object orientation to model hydrologic processes using the concepts of inheritance and aggregation. Recently, Richardson et al. (2007) proposed a prototype geographically based object framework for linking hydrologic and biochemical processes in the sub-surface. However, the process objects were loosely coupled with geographic objects, thus there is leaving an opportunity for a tightly coupled geographically based object oriented modeling.
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
19
(a)
(b) Fig. 1. Study area location: (a) represents Chesapeake Bay; and (b) represents Nanticoke Watershed including weather station and stream gauge.
Relatively recent advances in GIS have enabled the adaption of object orientation in storing and handling geospatial and temporal hydrologic data in research and practice. For example, Arc Hydro (Maidment, 1993) uses an object oriented approach to represent hydrologic environment through feature, object and relationship classes within a geodatabase. In ArcHydro, a HydroEdge (stream) is sub class of generic Polyline super class (inheritance), and is a part of HydroNetwork (aggregation). HydroEdge is related to Watershed (which itself is a sub-class of Polygon super class) through a common identifier (HydroID). Thus, ArcHydro uses object orientation to develop a physical representation of hydrography by using GIS objects. Thus, by knowing the HydroID of any geographic feature, it is possible to trace the flow of water by using points, lines, and polygons at multiple scales including at continental scale. The National Hydrography Dataset (NHD) available for the entire United States from the United States Geological Survey (USGS) also uses the object oriented (or geodatabase) design to provide data to its users.
The geodatabase approach to hydrology data overcomes several practical issues which are associated with storing and handling heterogeneous multi-scale data by providing a relational data model. Besides overcoming the data issues, the geodatabase approach provides an opportunity to exploit the potential of object oriented approach to overcome the limitations of scale and parameterization in distributed modeling of hydrologic processes. For example, if a polygon representing a watershed in GIS is treated as a hydrologic object that has some properties (e.g. area and slope) and methods (to compute runoff and route flow), then multiple watersheds can be linked and executed in parallel to scale-up the modeling domain from one single watershed to larger (national or continental) scales. Similarly, the availability of increasing GIS layers to represent soil, land use and topography at multiple scales, can enable parameterization of hydrologic processes (or watershed methods) through GIS tools, which is not possible with most existing models that do not explicitly work within a GIS environment. This research builds on past studies to create a prototype tightly
20
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
hydrologic processes to create excess rainfall and runoff hydrograph by implementing specific sub-classes (ExRain and Hydrograph) as shown in Fig. 2. ExRain implements specific techniques such as SCS curve number (through SCS sub-class) and Green– Ampt (through GreenAmpt sub-class) to compute excess rainfall using the rainfall input. Hydrograph class implements specific techniques such as storage release to compute runoff hydrograph from excess rainfall. TopoGrid implements sub-classes to create terrain attributes such as flow direction, flow accumulation, stream network and catchment by using topography data (DEM). HydroArea can work on flow transformation by implementing specific techniques such as SCS dimensionless unit hydrograph (through SCSUnit sub-class) and Clark unit hydrograph (through ClarkUnit sub-class) for vector data. Also, HydroLine can work on river routing by implementing specific techniques such as Kinematic wave river routing (through KinematicWave sub-class) and Muskingum river routing (through Muskingum sub-class). HydroTable implements three tables: TSTable, ParameterTable and PairTable. TSTable class processes time series table (e.g., rainfall and streamflow time series); ParameterTable process es parameter values linked to geographic features (e.g., Manning’s n values for different land cover types); and PairTable processes paired data such as stage-discharge rating curves.
coupled object-oriented hydrologic model in GIS to simulate hydrologic processes using geospatial inputs.
4. STORE DHM object classes The prototype modeling approach presented in this research is developed by using Visual Basic and ArcObjects within ArcGIS, and is referred to as the STORE DHM. The new conceptual hydrologic model is based on storage-release with grid travel time approach. In addition, the study uses an object-oriented programming approach which can provide useful data handling and model flexibility in hydrologic applications. An object-oriented hydrologic model framework is implemented in ArcGIS by creating hydrologic modeling objects as shown in Fig. 2. The hydrologic modeling object framework implements both principles of aggregation (represented with diamond) and inheritance (represented with triangle arrow) as shown in Fig. 2. HydroShed is the highest level class that includes the following six classes: (1) HydroGrid (to process gridded hydrologic information such as topography and rainfall), (2) ParameterGrid (to process gridded hydrologic parameters such as Mannings n), (3) HydroArea (to process vector hydrologic data for lakes and rivers), (4) HydroCatchment (to process vector hydrologic data for catchments or sub-watersheds), (5) HydroLine (to process vector hydrologic data for streams), and (6) HydroTable (to process tabular data). As displayed in Fig. 2 and Table 1, those classes dealing with raster, vector and tables are implemented in this research. ProcessGrid (to implement hydrologic processes) and TopoGrid (to implement terrain processes) are two sub-classes of the HydroGrid class. ProcessGrid can work on gridded data. It implements
*
HydroArea
1 1
-Attribute
Precipitation is the most crucial part in hydrologic modeling. Malfunctions in running hydrologic modeling can occur due to discontinuous time series precipitation inputs (Lee and Kang, 2015). As a part of practical test with flexibility of object-oriented in the
HydroShed
HydroCatchment
-Attribute
-Attribute
+Method()
+Method()
5. Reconstructing missing precipitation
1
1
*
+Method()
1 1
HydroLine
*
HydroGrid
-Attribute
-Attribute
*
+Method()
+Method() *
HydroTable ProcessGrid
TopoGrid -Attribute
-Attribute
+Method()
+Method()
ExRain
*
Hydrograph
ParameterGrid
-Attribute
-Attribute
+Method()
+Method()
TSTable
ParameterTable
PairTable
-Attribute
-Attribute
-Attribute
-Attribute
-Attribute
+Method()
+Method()
+Method()
+Method()
+Method()
SCS
GreenAmpt
-Attribute
-Attribute
+Method()
+Method()
Fig. 2. Object model diagram for STORE DHM (aggregation and inheritance characteristic in object-oriented represents with diamond and arrow, respectively). ⁄main object classes.
21
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25 Table 1 Types of object classes in STORE DHM. Class
Characteristic
Type
Implementation
HydroShed
Highest
Highest
Core class
HydroGrid ParameterGrid HydroArea HydroCatchment HydroLine HydroTable
Aggregation
Sub-class Sub-class Sub-class Sub-class Sub-class Sub-class
Process Process Process Process Process Process
ProcessGrid TopoGrid ExRain Hydrograph SCS GreenAmpt TSTable ParameterTable PairTable NesrestMP KernelMP
Inheritance
Under Under Under Under Under Under Under Under Under Under Under
Implement Implement Implement Implement
HydroCatchment
*
1
HydroGrid HydroGrid ProcessGrid ProcessGrid ExRain ExRain HydroTable HydroTable HydroTable TSTable TSTable
HydroArea
-Attribute +Method() 1 1
*
1
1
1 HydroLine *
TopoGrid
-Attribute +Method()
*
-Attribute +Method()
-Attribute +Method()
hydrologic processes terrain processes excess rainfall calculation runoff calculation
HydroShed
-Attribute +Method()
HydroGrid
gridded information gridded parameters vector data for rivers vector data for catchments vector data for streams tabular data
-Attribute +Method()
ProcessGrid
KinematicWave
Muskingum
-Attribute +Method()
-Attribute +Method()
-Attribute +Method()
ParameterGrid ExRain
Hydrograph
-Attribute +Method()
GreenAmpt -Attribute +Method()
-Attribute +Method()
RadarBias
HydroTable
-Attribute +Method()
*
SCS
TSTable
-Attribute +Method()
-Attribute +Method()
NearestMP -Attribute +Method()
-Attribute +Method()
ParameterTable -Attribute +Method()
*
-Attribute +Method()
PairTable -Attribute +Method()
KernelMP -Attribute +Method()
Fig. 3. Updated object model diagram for STORE DHM without changing its main frame (Red boxes represent updated classes). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
STORE DHM, this study proposed reconstructing missing precipitation data in hydrologic modeling: (1) nearest neighborhood regression (the NearestMP class); and (2) kernel function for missing precipitation (the KernelMP class) under the TSTable class (Fig. 3).
5.1. Nearest neighborhood regression A nearest neighborhood regression (NNR) typically uses a Markov chain to determine the missing precipitation based on proba-
22
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
5.2. Kernel function
Table 2 Manning’s n for Nanticoke watershed. Land use
Range
Initial value
After calibration
Agricultural Forest Developed Water
0.030–0.500 0.035–0.160 0.011–0.035 0.025–0.033
0.220 0.090 0.020 0.030
0.056 0.063 0.033 0.028
bility distributions. NNR so far has been successfully applied for estimating precipitation. NNR also has the advantage of computation efficiency (Sharif and Burn, 2006). It calculates missing precipitation with cumulative probability distribution of nearest neighborhood values and then selects the rank of the distance for each of the nearest value. A missing precipitation is estimated by NNR as follows: P 1X M¼ xi wi ; P i¼1
i¼1 Ri
P 1X xi Kðui Þ; P i¼1
ð3Þ
ui ¼
Ni ; 0:5P þ 1
ð4Þ
ð1Þ
P P Ni ¼ ; . . . ; ; 2 2
;
ð2Þ
where M is missing precipitation, P is the number of the nearest neighborhood, and xi is the sample data at ith nearest value and wi is the weighting for close neighbor at ith. R is the Pearson correlation coefficient that is a measure of the linear relationship between two random variables. The R values are converted into weights by using the weighting formula that is suggested by Dumedah and Caulibaly (2011). Estimation of the missing variable is computed as the weighted sum, and this procedure is operating in NearestMP class under TSTable class (Fig. 3).
6. Results 6.1. Update object-oriented framework Fig. 3 represents how to implement user optional precipitation inputs with object-oriented approach in the STORE DHM. Near-
25
7 Obs. Calibrated Sim.
Runoff (m3/s)
20 15 10
6
5
Obs. Sim.
5 4 3 2 1
0
0 0
0
100
20
40
15min/unit
15min/unit
(a)
(b)
60
20
40 35
Runoff (m3/s)
ð5Þ
where ui is the Nth nearest values which correspond to xi (positive means the right side and negative means the left side). K (ui) represents kernel function at ui. Estimating a missing precipitation data with KF, KernelMP, is implementing under the TSTable class (Fig. 3).
Runoff (m3/s)
wi ¼ PP
M¼
Obs. Sim.
30 25
Obs. Sim.
15
Runoff (m3/s)
Ri
While a NNR method is widely used in filling the missing precipitation data for hydrologic modeling, it has shown troublesome statistic theorem due to event based precipitation data having limited nearest sample data to make a normal distribution (Lee and Kang, 2015). General mechanism of weather data during short term rainfall event would not follow normal distribution trend. It also would not have enough nearest sample data around missing precipitation. As using of neighborhood mean value in kernel function (KF), it can overcome troublesome statistic theorem experienced in the NNR method (Lee and Kang, 2015). Missing precipitation data are calculated by the following KF formula:
10
20 15 10
5
5 0
0 0
100
200
0
50
100
15min/unit
15min/unit
(c)
(d)
150
Fig. 4. Hydrographs with STORE DHM simulations. (a), (b), (c) and (d) present Event 1, 2, 3 and 4, respectively.
23
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25 Table 3 STORE DHM results for Nanticoke watershed. Event
STORE DHM simulations 3
1 2 3 4
Observed runoff 2
Peak flow (m /s)
Time to peak (h)
R
22.48 5.9 34.76 17.6
7.50 3.75 19.75 11.50
0.95 0.94 0.94 0.95
ENS
Peak flow (m3/s)
Time to peak (h)
0.92 0.93 0.93 0.82
22.78 6.14 35.22 17.44
6.75 3.5 22.50 11.75
7
25
6 Event 1
Runoff (m3/s)
Runoff (m3/s)
20 15 10
5
Event 2
4 3 2
5 1 0
0 0
0
100
40
20
35
18
Runoff (m3/s)
Runoff (m3/s)
40
16
Event 3
30
20
60
15min/unit
15min/unit
25 20 15 10
Event 4
14 12 10 8 6 4
5
2
0
0 0
100
200
15min/unit Observed
0
100
15min/unit Simulation with missing precipitation
Fig. 5. STORE DHM simulations with synthetic missing precipitations.
estMP and KernelMP classes are created by an inheritance characteristic of object-oriented model and the two classes implement the estimation of missing precipitation values under the TSTable class. An object-oriented data class such as TSTable is a bridge between data handling, e.g., NearestMP and KernelMP, and calculating hydrologic processes (ProcessGrid). The first step executing NearestMP and KernelMP is to load time series rainfall data in the TSTable. Then, the TSTable will automatically scan for whether time series data is continuous or discontinuous. If the data has any missing rainfall data, it can be transferred to the NearestMP or KernelMP to reconstruct missing data depending on user selection. The NearestMP and KernelMP illustrates the ‘‘is-kind-of” relationships to the TSTable throughout inheritance characteristic in object-oriented framework. Each entity and attribute in NearestMP and KernelMP is associated with HydroTable which is an upper level class for all tabular data. For example, KF equations are associated with TSTable to generate time series precipitation data and it can also be associated with the HydroTable as an input table for the ProcessGrid. Likewise, KinematicWave and RadarBias classes (Fig. 3), updated classes in the STORE DHM, are implemented for runoff routing under the HydroLine and radar precipitation bias correction under the ProcessGrid, respectively (Kang
and Merwade, 2014). But, these classes are not presented their internal methodologies in this paper to avoid inconsistency of the theoretic topic. 6.2. STORE DHM simulation Instead of using theoretical calculation for getting geographical parameters, the STORE DHM used spatially distributed grid data in ArcGIS. Thus, calibration simulation only for manning’s n is executed for the first modeling step. After getting the calibrated manning’s n from the first event (Table 2), the model is applied to three additional three events at Nanticoke watershed. Fig. 4 presents hydrograph results from application of the model. Qualitative and quantitative comparison of each model hydrograph, peak flow and time to peak, with observed hydrograph show reasonable agreement. Coefficient of determination (R2) and Nash Sutcliffe efficiency coefficient (ENS) are used for model performance measures, and STORE DHM showed 0.94 average R2 and 0.90 average ENS. Model simulations with calibrated manning’s show that the peak flow and time-to-peak are underestimated by 6.4% and 5.7%, respectively. Table 3 presents a summary of model results and error statics for STORE DHM simulations. Overall, comparison
24
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
30
7
25
6 Event 2
5
Runoff (m3/s)
Runoff (m3/s)
Event 1 20 15 10
4 3 2
5
1
0
0 0
50
100
0
150
20
40
15min/unit 40
20
35
18
Runoff (m3/s)
Runoff (m3/s)
16
Event 3
30
60
15min/unit
25 20 15 10
Event 4
14 12 10 8 6 4
5
2
0
0 0
100
0
200
50
100
15min/unit Observed
150
15min/unit Simulation with NNR
Simulation with KN
Fig. 6. STORE DHM simulation with reconstructed missing precipitations. Table 4 Validation results using missing and reconstructed precipitation. Event
1 2 3 4
Simulation with missing prec.
Simulation with NNR
Simulation with KN
Peak flow (m3/s)
Time to peak (h)
R2
ENS
Peak flow (m3/s)
Time to peak (h)
R2
ENS
Peak flow (m3/s)
Time to peak (h)
R2
ENS
18.22 5.92 29.17 14.27
6.75 3.75 19.75 12.00
0.81 0.91 0.91 0.95
0.86 0.87 0.81 0.64
25.57 6.58 36.35 17.58
7.00 3.75 19.50 12.50
0.82 0.92 0.91 0.94
0.85 0.88 0.88 0.84
23.62 8.81 31.59 16.74
7.00 3.75 19.75 11.75
0.88 0.93 0.91 0.95
0.91 0.93 0.86 0.86
12
of the model output and observed data looks reasonable for the Nanticoke watershed (Fig. 4).
NNR
KN
10
Making a synthetic missing precipitation data from four selected event periods is the first step, and then STORE DHM simulations with missing precipitation are performed in Fig. 5 which are showed under estimated hydrographs comparing with observed hydrographs due to less quantity of precipitation. Fig. 6 and Table 4 presents results from application of the model with NearestMP and KernelMP classes for estimating missing precipitation in Nanticoke watershed. The results from STORE DHM with missing precipitation are compared with both NNR and KN reconstructed missing precipitation (Table 4). Results in NNR are showed 18% over estimation of the peak flow. The time-to-peak is underestimated in event 4. Fig. 7 presents scatter plots of reconstructed missing precipitation both using NNR and KF. Estimated missing precipitation with NNR showed over estimation than observation precipitation. On the other hand, estimated missing precipitation with KF showed well estimation to observation precipitation (Fig. 7).
Estimated precipitation
6.3. Additional precipitation options 8
6
4
2
0
0
2
4
6
8
10
12
Observed precipitation Fig. 7. Comparison of NNR and KN reconstructed precipitation data for Nanticoke watershed.
K. Kang et al. / Journal of Hydrology 540 (2016) 17–25
7. Conclusions This research is focused on developing a prototype objectoriented hydrologic modeling framework called the STORE DHM. This research presents an object-oriented approach using objectoriented design techniques and object-oriented language (Visual Basic) to the description and simulation of a watershed based hydrologic processes. The STORE DHM allows flexibility and extensibility to investigate future hydrologic issues without changing its main framework because it uses characteristics of inheritance and aggregation through object-oriented hydrologic approach. It would be a flexible and dynamic tool that is capable of simulating a wide array of management practices such as water contamination, supplying, flooding and drought from watershed scale to continental scale. The user necessaries have been incorporated in development of the STORE DHM to generate further investigation on multidisciplinary water resources researches. For example, nutrient contaminations from surface runoff in crop areas will be investigated by using of the Agricultural Policy Environmental Extender (APEX) model with linking class under the STORE DHM framework for making a sustainable water resources management. Model extensibility is an inescapable factor that builds up sustainable water resources management. Unlike a hierarchy model which is all handled by one core frame, the STORE DHM, composed of an objectoriented approach, can be easily updated and modified without changing the main frame. Each object can be worked individually by its attribute and method, and also it can be part of an entire model procedure through inheritance and aggregation characteristics. Furthermore, inheritance and aggregation of object-oriented approach in the STORE DHM is contributing to save the unnecessary computation demand, and also enhances modeling efficiency. The STORE DHM framework involves computing rainfall, volumetric flow rate, and travel time to the basin outlet by combining steady state uniform flow approximation with Manning’s equation (Kang and Merwade, 2011). While other travel time based distributed hydrologic models that route the flow from each cell to the watershed outlet along a common flow path at each time step (which can lead to improper water balance), the STORE DHM considers the flow contribution of neighboring cells by using continuity equation with storage release concept (which can lead to proper water balance). Expendable user precipitation options, NearestMP and KernelMP, are used for evaluating the STORE DHM flexibility with object-oriented approach. Lee and Kang (2015) used SWAT (Soil Water Assessment Tools) to evaluate better statistical scheme for reconstructing missing precipitation and they found that SWAT simulation with KF approach for reconstructing missing precipitation showed a better performance than simulation with NNR approach in Imha watershed. Likewise, STORE DHM simulation through the KernelMP class shows better simulation accuracy than simulation with the NearestMP class. Overall results from these comparisons show promising hydrograph outputs generated by the proposed two classes. These simulations are not always true for different events and areas. Therefore, additional object classes and study areas for the wide array of time series precipitation researches are necessary in the STORE DHM for further research. As the first part of the broader sustainable water resources modeling framework, this research develops the grid based hydrologic model for hydrologic modeling as modular development. The modeling framework presented in this research is operated within ArcGIS environment such that all the steps from extracting information from geospatial data to running model simulations are exe-
25
cuted in an ArcGIS environment. The approach from data to model output within a single environment is attractive from a practical point of view. Expanding of the STORE DHM will be essential for robustness of the hydrologic model. For example, linkage between STORE DHM and water quality model (e.g., SWAT) can predict effect of storm event on water quality assessment. Similarly, linkage between GCMs (General Circulation Models), which are the most advanced tools for estimating future climate change, and GHISMO will provide a foundation to assess the impact of climate change on rainfall-runoff prediction. The conceptualization and characterization of this coupling strategy can be extended to a sustainable water resources management and decision supporting tool. It the prototype of the object-oriented hydrologic framework, the STORE DHM, is successfully used to develop a coupled open source and platform for seamless hydrologic components, it may lead to change a hydrologic modeling paradigm to objectoriented approach through its flexible modeling schemes. Acknowledgements We would like to the editor (Dr. Geoff Syme) for constructive suggestion on previous STORE DHM paper. The first author would also like to thank proofreading from Darae Kang. References Band, L.E., Tague, C.L., Brun, S.E., Tenenbaum, D.E., Frernandez, R.A., 2000. Modeling watersheds as spatial object hierarchies: structure and dynamics. Trans. GIS 4 (3), 181–196. Bian, L., 2007. Object-oriented representation of environmental phenomena: is everything best represented as an object. Ann. Assoc. Am. Geogr. 97 (2), 267– 281. Boyer, J.F., Berkhoff, C., Servat, E., 1996. Object-oriented programming for a simulation of the rainfall-discharge relationship. In: Balkema, A.A. (Ed.), Proceedings of Hydroinformatics 96. Balkema, A.A. Zurich, Switzerland, pp. 299–305. Chen, H., Beschta, R., 1999. Dynamic hydrologic simulation of the Bear Crook Watershed in Maine (BBWM). Environ. Monit. Assess. 55, 53–59. Dumedah, G., Caulibaly, P., 2011. Evaluation of statistical methods for infillinf missing values in high-resolution soil moisture data. J. Hydrol. 400 (1–2), 95– 102. Garrote, L., Becchi, I., 1997. Object-oriented software for distributed rainfall-runoff models. J. Comput. Civ. Eng. 11, 190–194. Goodchild, M., Parks, B., Steyaert, L., 1993. Environmental Modeling with GIS. Oxford University Press, Oxford. Kang, K., Merwade, V., 2011. Development and application of a storage-release based distributed hydrologic modeling using GIS. J. Hydrol. 403, 1–13. Kang, K., Merwade, V., 2014. The effect of spatially uniform and non-uniform precipitation bias correction methods on improving NEXRAD rainfall accuracy for distributed hydrologic modeling. Hydrol. Res. 45 (1), 23–42. Kiker, G.A., Clark, D.J., Martinez, C.J., Schulz, R.E., 2006. A java based, object-oriented modeling system for southern African hydrology. Trans. ASABE 49 (5), 1419– 1433. Kralisch, S., Krause, P., David, O., 2005. Using the object modeling system for hydrological model development and application. Adv. Geosci. 4, 75–81. Lee, H., Kang, K., 2015. Interpolation of missing precipitation data using kernel estimations for hydrologic modeling. Adv. Meteorol. 935868, 1–13. Maidment, D.R., 1993. GIS and hydrologic modeling. In: Goodchild, M., Parks, B., Steyaert, L. (Eds.), Environmental Modeling with GIS. Oxford University Press, New York, USA. Ogden, F.L., Julien, P.Y., 1993. Runoff sensitivity to temporal and spatial rainfall variability at runoff plane and small basin scales. Water Resour. Res. 29 (8), 2589–2597. Richardson, M.C., Branfireun, B.A., Robinson, V.B., Graniero, P.A., 2007. Towards simulating biogeochemical hot spots in the landscape: a geographic objectbased approach. J. Hydrol. 342 (1–2), 97–109. Sharif, M., Burn, D.H., 2006. Simulating climate change scenarios using an improved K-nearest neighbor model. J. Hydrol. 325, 179–196. Wang, J., Hassett, J.M., Endreny, T.A., 2005. An object oriented approach to the description and simulation of watershed scale hydrologic processes. Comput. Geosci. 31 (4), 425–435. Whittaker, A.D., Wolfe, M.L., Godbole, R., Van Alem, G.J., 1991. Object-oriented modeling of hydrologic processes. Appl. Nat. Resour. Manage. 5 (4), 49–58.