4.1. Abstract

This chapter demonstrates the adaptation of the Generalized Watershed Loading Functions (GWLF) Model (Haith et al., 1992) for an Internet case study in the Spring Creek Watershed in central Pennsylvania. This adaptation was performed as part of the development of the Internet Watershed Educational Tool (InterWET) (Parson, 1999, and Parson et al., 1999). The GWLF model was used to predict changes in average annual loads of surface runoff, groundwater runoff, total runoff, detached sediment, delivered sediment, dissolved nitrogen and phosphorous, and sediment nitrogen and phosphorous for eight subwatersheds of the Spring Creek Watershed. Changes in load were calculated for different local government policy choices concerning land development, riparian zones, waste water management, and agricultural best management practices for the two regions within the watershed, either in or not in the Centre Region. The calculated changes reflected estimates based on 20 years of generated weather data. Computer programs were created which automated the creation of GWLF load change predictions for a particular policy scenario. Total runoff was found not to change significantly for any of the policy scenarios. Only a small number of policies affected surface runoff, groundwater runoff, and detached sediment, so they were modeled with look-up tables. Changes in delivered sediment, dissolved nitrogen and phosphorous, and sediment nitrogen and phosphorous loads were modeled with Forward-Backward Propagation Artificial Neural Networks (ANNs) for the eight subwatersheds for eight different load change categories. Training of each ANN consisted of 1000 policy scenario data sets and 100 data sets for testing. Training was performed until training error reach a relatively constant value. On average, the trained ANNs correctly classified 81.6 % of the tested data, and were within 1 change category for 95.7 % of the tested data. The trained ANNs and look-up tables were then implemented into InterWET using client-side JavaScript.

4.2. Introduction

Web pages are usually designed for content and speed. When accessing online items, things like commerce to the local weather, users demand that a web page have the desired content and can be loaded quickly. The time it takes a web page to load in a web browser depends on the speed of the Internet connection; the computer speed and platform of the user or client; the setup and speed of the sending or server computer; and the memory size of the web page. Web page developers must balance the memory needs of detailed content with the time it takes to deliver that content. Developers of hydrologic modeling applications for the Internet must also account for model run times.

Fortunately, there are several techniques that increase the speed of Internet-based models, while preserving content. First, the web page can make use of client-side techniques like JavaScripting, which places the computing load on the client's computer, not the overburdened server. Second, models can be made smaller or modular. Instead of having one model calculate all of the pertinent outputs, several sub-models can be developed which calculate only a select set of outputs. Third, output display can be simplified. For instance, if the client is only interested in whether a particular Best Management Practice (BMP) causes a decrease in some water constituent value, a model can be constructed which only determines increases or decreases of the value, not three-digit precision. Fourth, the models can utilize techniques that increase calculation speed like artificial neural networks.

This chapter will focus on adapting the Generalized Watershed Loading Functions (GWLF) Model (Haith et al., 1992) to the Internet for the case study of the Spring Creek Watershed in central Pennsylvania. This adaptation was performed as part of the development of the Internet Watershed Educational Tool (InterWET) (Parson et al., 1999). In addition to supporting the overall objectives for InterWET, this chapter has several specific objectives. The first objective is to show how GWLF can be applied to the Spring Creek watershed. A second objective is to provide, based on the literature, a foundation for the application of artificial neural networks to hydrology. The third specific objective of this chapter is to explain how the hydrologic modeling results from GWLF are portrayed on the Internet through the use of look-up tables and artificial neural networks.

To support these specific objectives, and InterWET’s overall objectives, this chapter provides: a brief overview of some techniques for Internet-based hydrologic modeling (additional details in Chapter 3), the development and structure of InterWET (detailed in Chapter 2), and artificial neural networks and their past use in hydrologic modeling. The methods section will detail how particular GWLF outputs were used to train artificial neural networks, which were then used in JavaScript web pages. The results section will detail the GWLF modeling, the success of the training and the testing of each artificial neural network, and their implementation into JavaScript. A summary of the results and some final thoughts are then given in the conclusions section.

4.3. Background

4.3.1. Techniques for Internet-Based Hydrologic Modeling

Parson (1999) gives an extensive review of techniques for Internet-based hydrologic modeling and several current web sites that utilize these techniques. Currently, there are a number of techniques that can be used for performing hydrologic modeling on the Internet. One of simplest techniques is the use of Hypertext Markup Language or HTML. HTML is not a programming language, but a formatting method. To change the size, font, or color of some text, one surrounds the text with HTML tags. Scripts and programs are also added to a web page with the appropriate tags. Used independently from other techniques, HTML provides ways to organize and display information and to access sites with downloadable models. HTML does not support mathematical calculations, but provides the foundation to build all other web site features. Embedded into the HTML coding can be server-side CGI script, client-side JavaScript, and Java object-oriented programming language (Vander Veer, 1997).

CGI scripts allow a web page to be dynamic by using server-side scripts (Wooldridge and Morgan, 1997). Scripting languages are interpreted line by line and executed immediately. This differs from programming languages, where the code is initially compiled into an executable file that is run later (Wooldridge and Morgan, 1997). By being server-side and dynamic, CGI scripts have capabilities that HTML does not. CGI can produce new web pages based on interactive user input, since it has the ability to write new files to a server (Cornell and Horstmann, 1996).

These features, though, come at a price. Being server-side, CGI scripting has limitations related to development, platform compatibility, data storage, and security. Another issue is server speed. Since the script must be interpreted each time it is run, a lengthy, complicated script may take a long time to run. Other time lags are added depending on the number of users on the server. As a result, the client may have to wait a long time to obtain results from a CGI script (Wooldridge and Morgan, 1997). For certain applications, like submitting form data, CGI scripts work well. However, other techniques have been developed which attempt to overcome some of the shortcomings of CGI scripts.

JavaScript is a client-side scripting language specifically developed to extend HTML (Vander Veer, 1997). Like CGI scripting, JavaScript is interpreted, not compiled. JavaScript works within the web browser to give advanced features to HTML and more control over browser characteristics. JavaScript can respond to user input from the keyboard or mouse to perform some action and allows more advanced interfaces like drop-down lists or radio buttons (Vander Veer, 1997). Instead of sending a brand new web page in response to client input like CGI scripts, JavaScript can maintain portions of the web page, while changing other portions to respond to input (Wooldridge and Morgan, 1997).

Being client-side, instead of server-side, gives JavaScript certain advantages over CGI scripting. Since JavaScript can not write files to the server and does all interpretation on the client's computer, data storage and security issues of CGI scripting are overcome. The speed for JavaScripts is also faster, since new files are not being created and all calculations are being done on the client's computer (Wooldridge and Morgan, 1997). However, JavaScript still has several shortcomings. JavaScript can not write files to the server, which can be a major drawback for hydrologic modeling. While JavaScript does allow files to be saved to the client's computer, these files are only copies of the HTML code, graphics, or text on the web page display. Because of the security restrictions posed by being client-side, JavaScript does not have the file input/output routines that are commonly used to save data (Wooldridge and Morgan, 1997). In addition, JavaScript is not completely standardized. Despite these shortcomings, JavaScript provides the best balance of learning difficulty with advanced features for beginning web programmers. In an attempt to go beyond the capabilities of JavaScript, the Java programming language was developed.

Java and JavaScript are two separate and different languages. Java is a platform independent object-oriented programming language. Being platform independent makes Java code the same for UNIX, Windows, and Macintoshes. Java code can be compiled into computer programs. However, Java code can also be converted into applets that can be run on a web page (Cornell and Horstmann, 1996). Java applets can be used in a web browser that is Java enabled. Using the proper HTML tag, an applet can be embedded into a web page, usually within a window or in a separate frame. The applet acts like an independent program within the web page. Therefore, Java applets can do many things, like dynamic graphics, that JavaScript cannot.

One big advantage of Java over JavaScript is the ability to access files. Java can read files from a server or a client computer. Java can also be used to write CGI scripts, so files can be written to the server. Still, Java applets currently can not write files to a client's computer. Cornell and Horstmann (1996) indicate this restriction is not an inherent part of Java and may be removed for future versions. The file input/output problem is indicative of the major shortcoming of Java; it is a very new programming language. Even more than JavaScript, Java is not standardized and its capabilities change significantly between different versions. As a result, Java is a powerful technique for modeling on the Internet, although difficult to use or trust.

In addition to CGI scripts, JavaScripts, and Java, third-party software in the form of helper applications, Plug-Ins and ActiveX components, and server-side software can be used to enhance web page features. When the browser is asked to load a file in a format it is not designed to handle, it can prompt the client to load the file using an existing software program or helper application on the client's computer. The helper application has no interaction with the web page. The program acts completely independent of the browser. To overcome this problem, web browser developers created components that could be embedded into a web page. Specifically, Netscape called these components Plug-Ins and Microsoft called theirs ActiveX. Both Plug-Ins and ActiveX components are platform specific programs that work with the web browser to add some new feature to a web page, like the ability to play a sound file or show a movie clip. Usually these components are very feature specific, making them smaller than helper applications, but also more limited (Wooldridge and Morgan, 1996). Both Plug-Ins and ActiveX are added to a web page using specific HTML tags and can communicate with JavaScript and Java components in the web page (Vander Veer, 1997).

Server-side software provides a server-based alternative to Plug-Ins and ActiveX. This very specialized software is designed to be run on a server and deliver specific features to a web site. Server-side software can run server programs and perform file input and output to the server.

All of these third-party software techniques have some serious drawbacks. First, all of these techniques are very computer platform dependent. As a result, developing new software using one of these techniques requires programming expertise for multiple browsers and for multiple platforms. Second, there are security issues; on the client-side for helper applications, Plug-Ins, and ActiveX and on the server-side for server-side software. Still, these techniques can provide features to a web page that CGI scripting, JavaScripting, and Java can not.

All of the techniques discussed can be used for hydrologic modeling on the Internet. The next section will highlight how HTML, JavaScript, and Java were used in InterWET.

4.3.2. Internet Watershed Educational Tool (InterWET)

The Internet Watershed Educational Tool (InterWET) was developed as part of a doctoral research fellowship in Environmental Engineering from the U.S. Department of Education. InterWET is available on the Internet at http://server.age.psu.edu/dept/grads/parson/research/home.htm (Parson et al., 1999). InterWET was created to help educate local officials about water resources, using as a case study the Spring Creek Watershed in central Pennsylvania (Parson, 1999).

InterWET consists of a set of web pages that present five water resource components from three different perspectives. The tool utilizes the "microworlds" concept, defined by Edwards (1995) as computer-based learning environments, "which embody mathematical and scientific concepts in a context which is engaging to the learner, and which allows for a certain degree of self-directed exploration or discovery of the implicit ideas and processes." InterWET was developed around fifteen different microworlds, where each microworld is a combination of one of five water resource components (runoff, groundwater, sediment, nutrient, fish) and one of three perspectives (researcher, conservationist, local official).

Each of the three perspectives is linked to a certain dimensionality. For a researcher like an agricultural engineer, the water resource components can often be seen as individual, two-dimensional processes. A process like runoff might be measured from a small field plot. This can be contrasted with the perspective of a conservationist from an organization like Trout Unlimited, who is more concerned with the geographic, three-dimensional aspects of water resources. From this perspective, the amount and quality of runoff in watershed streams on a particular day is important. A local official might expand this idea to a temporal or four-dimensional perspective. Not only is the amount of runoff on a particular day important, but also how would local policy choices made today affect average stream flow 10 and 20 years from now. While researchers, conservationists, and local officials are concerned with all dimensions in a watershed, they are more likely to encounter and work with information in the dimension they are linked to in InterWET (Parson, 1999).

To construct these microworlds, InterWET uses HTML, JavaScript, and Java. HTML is used to create the basic appearance of all web pages, text and images on each page, and navigational links between pages. JavaScript is used intensively for display and calculation. Since InterWET is an educational site, it does not need extensive file input and output. Each microworld is constructed to be a self-contained entity. All microworlds for the researcher or process perspective are JavaScript calculators. For instance, the erosion calculator uses the Universal Soil Loss Equation (Wischmeier and Smith, 1978). The input comes from text fields for numbers and radio buttons for different variables. Since InterWET's audience of local planners might not be familiar with things like slope length, some values are predefined and assigned to the radio buttons. Once all inputs are entered, the user clicks a Calculate Button and JavaScript calculates a value, which is displayed in a text field. Since input and output are both shown in a single interface, file storage is not needed. Parson (1999) gives details on the different methods used for each calculator.

Java is used for GIS display for the conservationist or geographic perspective. Specifically, ActiveMaps v2.0 was purchased from InternetGIS.com (InternetGIS.com, 1998). ActiveMaps is a Java, client-side GIS program, giving low cost GIS features to InterWET. ActiveMaps uses database files, so features like queried data look-ups are supported.

The microworlds for the local official, or temporal perspective, use JavaScript to represent output from the Generalized Watershed Loading Functions (GWLF) model (Haith et al., 1992). The GWLF Model utilizes simple loading functions to predict monthly runoff, sediment, soluble nitrogen and phosphorous, and sediment nitrogen and phosphorous estimates. InterWET uses artificial neural networks to replicate the predictions from GWLF. The next section will provide some background on artificial neural networks. The methods section will then describe how they were applied to GWLF.

4.3.3. Artificial Neural Networks and Hydrologic Applications

The fundamental concepts underlying Artificial Neural Networks (ANNs) are analogous to the functioning of neural networks in the human brain. ANNs consist of a number of interconnected processing elements called neurodes, analogs to neurons in the brain (Maier and Dandy, 1996). Each neurode combines a number of inputs and produces an output, which is transmitted to many different locations, including other neurodes. The difference between the various types of ANNs usually comes from the many different ways to arrange the neurodes (architecture) and the many ways to determine the weights and functions for inputs and neurodes (training) (Caudill and Butler, 1992a).

Figure 4.1 shows a typical artificial neural network architecture. Inputs to the ANN are usually binary, either "0 and 1" or "-1 and +1." The connections between the input layer and the middle or hidden layer contain weights, which are usually determined through training the system. The middle layer neurodes take the weighted inputs and sums them. To make a single value output from each neurode, the sum is used in an equation called a transfer function to create an output value. A typical transfer function is the sigmoid function, which varies from 0 to 1 for a range of inputs (Caudill and Butler, 1992a). The connections between the middle and output layer also have weights, and the output layer neurodes contain transfer functions. Finally, the output layer produces a set of values. These values might put a particular input into an output category or class, or the values may be combined or scaled to produce a numeric value (Caudill and Butler, 1992a). Maier and Dandy (1996) give a more detailed description of the architecture of ANNs.

ANNs are powerful tools, because architecture and training procedures allow a great deal of flexibility to model many types of systems. When there is a large data set of input and outputs available for some system, a supervised training procedure can be used to create an ANN. Some of the more popular supervised ANNs are the Forward-Backward Propagation network, Hopfield network, and the bidirectional associative memory network (Caudill and Butler, 1992a). For a system with highly complex input data, there are unsupervised ANNs which can help with pattern recognition and data compression. Some of the more popular unsupervised ANNs include Kohonen networks and resonance networks (Caudill and Butler, 1992b). In addition, there are many types of hybrid and higher-order networks that take portions of several networks and combine them into a single system.

ANNs are beginning to have an impact on water resource and hydrologic modeling. According to Maier and Dandy (1997), ANNs were first introduced to the water resource community by Daniell (1991), who used ANNs to predict monthly water consumption and to estimate flood occurrence. Since then, ANNs have been used for a variety of water resource applications. These include time-series prediction for rainfall forecasting (French et al., 1992), reservoir inflow time series (Raman and Sunilkumar, 1995), rainfall-runoff processes (Hsu, 1995 and Shamseldin, 1997), and river salinity (Maier and Dandy, 1996). ANNs also have been used for representing soil and water processes including soil moisture (Altenford, 1992), groundwater cleanup strategies (Ranjithan et al., 1993), water table fluctuations (Shukla et al., 1996; Yang et al., 1996a), pesticide movement in soils (Yang et al., 1996b), drainage pattern determination from a Digital Elevation Model (Kao, 1996), and water table management (Yang et al., 1998). In addition, they have been applied to irrigation systems (McClendon et al., 1996), a greenhouse management expert system (Seginer et al., 1996), and a water quality management system (ChingGung and ChihSheng, 1998).

As an emerging field of modeling, many of the methods and techniques used in ANNs are not standardized and vary between texts and software applications. Caudill and Butler (1992a) suggest the use of a trial and error approach for determining parameters and network architecture for a given problem. ANNs also tend to be very problem specific (Maier and Dandy, 1996). Still, some general guidelines emerge from the literature. Usually the larger the number of neurodes in the hidden layer or layers the better the predictive ability of the network (Caudill and Butler, 1992a). Training data sets will produce better networks when they are larger and contain the extremes of possible values. ANNs will almost never perfectly predict all values, so a reasonable error must be used for training and testing of networks (Caudill and Butler, 1992a). The key to a good network is the training data.

This research used Forward-Backward Propagation (FBP) ANNs for replicating GWLF output for InterWET. The specific design of the FBP networks is given in Caudill and Butler (1992a). The methods section will detail how GWLF was used to develop training data for these networks, how the FBP networks were constructed and tested for certain GWLF outputs, and how these trained networks were included in InterWET using JavaScript.

4.4. Methods

JavaScript ANNs of GWLF output were developed in three major steps. In step one, GWLF was applied to the Spring Creek Watershed in central Pennsylvania (shown in Figure 4.2.) Initially, GWLF input files were produced for the eight different subwatersheds of the main watershed (shown in Figure 4.3.) Then, software was developed to adjust these GWLF input files according to a set of four different local government policies. Other software was created to extract GWLF outputs for each subwatershed and policy scenario. Additional software was crafted to link these programs. Therefore, this system of linked software, available through Parson (1998), could produce a set of policy scenarios, run them through GWLF, extract outputs, and compare those outputs to a base case to predict the amount of change.

Step two used the results of step one to develop look-up tables and ANNs that could be used to replicate the GWLF predictions for the nine outputs. For three of the nine outputs, it was found that only 1 or 2 of the local government policies affected their predicted values, so they were modeled with look-up tables. Another output was found to have insignificant change to any policy scenario, so it did not require any modeling technique. The remaining five outputs were affected by most of the policies, and were modeled using FBP ANNs. Preliminary testing was performed for each of the five outputs to determine the proper network architecture, network parameters, number of training sets, and attainable minimum errors for training and testing. Using the setup determined from these preliminary tests, the GWLF estimates of change for the five outputs were used to train five FBP ANNs.

Finally, the results of step two were applied to the Internet via JavaScript calculators. These calculators were produced for eight outputs, in which policy choices were input and the changes of output for each of the 8 subwatersheds were predicted. Three outputs had calculators that used look-up tables and displayed either the numerical percent change or the percent change category (shown in Table 4.5.) The other five outputs used their trained ANNs to display output only in one of the percent change categories.

The final results of these three main steps were JavaScript calculators that predict how policy choices change eight different GWLF outputs. The following three sections will describe more fully the methods used in each of these main steps.

4.4.1. Step One: GWLF for Spring Creek Watershed

GWLF was used to model the Spring Creek Watershed in Centre County, Pennsylvania (shown in Figure 4.2.) Part of the Susquehanna River Basin, the watershed drains 381 km² in the Ridge and Valley physiographic province of central Pennsylvania. Spring Creek flows for a length of 35 km and is nationally known for its trout fishing (Schmalz, 1996). Figure 4.3 shows the eight subwatersheds modeled using GWLF. Four of the subwatersheds correspond to the four main tributaries of Spring Creek: Logan Branch, Buffalo Run, Slab Cabin Run, and Cedar Run. The other four subwatersheds represent the drainage areas into Spring Creek for which monitoring data are available. There are four years of flow and in-stream sediment data from Carline (1998) at the headwaters of Spring Creek, measured at the confluence of Spring Creek and Cedar Run at Oak Hall. The other three subwatersheds of Spring Creek (upstream from Houserville, Axeman, and Milesburg) all have a USGS gaging station. Figure 4.4 shows the major municipal boundaries in the watershed. Approximately half of the municipalities make up the Centre Region Council of Governments. For many local policy decisions that affect water resources, Centre Region communities coordinate their activities.

The Generalized Watershed Loading Functions (GWLF) Model (Haith et al., 1992) was applied to these eight subwatersheds of Spring Creek Watershed. GWLF requires input information for water transport, nutrients, and weather. Table 4.1 contains the inputs required and the data sources used for the water transport file. Likewise, Table 4.2 contains the inputs and data sources for the nutrient file. Weather data were generated with the CLIGEN program (Laflen et al., 1991), which will be described later.

Haith et al. (1992) provide the guidance for determining most of the inputs to the transport and nutrient files. ArcView GIS v.3.0 was used for map and database manipulation. The original base maps for the watershed came from ERRI (1998), Myers (1998), and Day (1998). The original map layers included detailed vector landuse, soils, watershed and subwatershed boundaries, streams, and municipal boundaries, and raster USGS DEMs for elevation. All vector layers were converted to the same raster format of the DEMs, approximately 30 m by 30 m. These raster maps were then used with the water and nutrient data (Table 4.1 and Table 4.2) to determine most of the inputs.

However, derivation of some of the input values requires further description. In the transport file, the USLE S Factor was derived from the DEMs using slope functions in ArcView based on Jenson and Domingue (1988) and the proper USLE equations from Haan et al. (1994). Derivation of the USLE L factor came from DEMs and stream maps using ArcView functions and USLE equations from Jenson and Domingue (1988), Haan et al. (1994), and Lehning (1998).

In the nutrient file, point sources were assumed from only two significant sources, the University Area Joint Authority (UAJA) sewage treatment plant (STP) near Houserville and another STP near Bellefonte, shown on Figure 4.4. Except for sewering on Penn State Campus, which is discharged through spray irrigation on State Game Lands, all sewering in Centre Region is assumed to go to the UAJA STP. All sewering in communities not in the Centre Region is assumed to go to the Bellefonte STP (Centre County Government, 1996, and U.S. EPA, 1998).

GWLF also requires daily average temperature and precipitation data for the time period of study. A 20 year time period was chosen, because the current high population growth is expected to continue for at least that length of time. The CLIGEN program in WEPP (Laflen et al., 1991) was used to generate 20 years of weather data for the watershed. The data were derived using parameters from the closest station (Lewistown) in the CLIGEN database, about 35 km from the watershed.

For this research, nine load predictions from GWLF were considered: surface runoff, groundwater runoff, total runoff, detached sediment, delivered sediment, sediment nitrogen and phosphorous, and dissolved nitrogen and phosphorous. These predicted values are average annual loads for the 20 years of simulation. The loads are related to different InterWET microworlds from the local official perspective. Surface runoff was used in the Runoff microworld. The Groundwater microworld contained the groundwater and total runoff. Detached and delivered sediment went into the Sediment microworld. The Nutrient microworld contained the sediment nitrogen and phosphorous and dissolved nitrogen and phosphorous.

GWLF was run with the transport, nutrient, and weather files described above with the nine loads being extracted from the GWLF output files. This created a base case to compare with GWLF outputs resulting from each scenario of policy choices.

The policy choices were based on areas of concern raised by the International Countryside Stewardship Exchange to the Spring Creek Watershed, during September 1996. The Exchange, sponsored by several national environmental organizations, including the Chesapeake Bay Program, brought a group of international watershed experts to the watershed area to help local organizations and governments develop a list of the most important water resource issues (International Countryside Stewardship Exchange Steering Committee, 1996a). The experts considered a large list of issues and concerns raised by local stakeholders and compiled a report that highlighted only the eight most important issues (International Countryside Stewardship Exchange Steering Committee, 1996b). Four local policy areas that relate to these issues, detailed in Table 4.3, are Land Development, Riparian Zone Protection, Waste Water Management, and Agricultural BMPs. Each of these policy areas have four different choices, where the first choice is a "no change" option.

Land development policies concern how existing land use would change due to an increasing population in the watershed. All of the change in land use is based on the premise of maintaining the current developed land population density. For policy options where agricultural land, forests, or all undeveloped land were developed, it was assumed that developed land population density would remain at current levels, but the population would increase. This required land area to change to developed for certain undeveloped areas. The policy choice of a growth boundary (or "no change") maintains the current land use, but increases the population density in developed areas.

Riparian Zone policies take a very general approach to show how differing amounts of riparian protection affect water resources. The policies are based on control of sediment delivery, since GWLF does not account for in-stream erosion, but does use a sediment delivery ratio (SDR) based on the size of the subwatershed (Haith et al., 1992). The policy choice of ‘less protection’ sets the SDR to 0.3, a maximum value suggested by Walling (1983). The policy choice of ‘total protection’ set the SDR to a minimum at 0.05 (Walling, 1983). The choice ‘no change’ kept the SDR at the value based on subwatershed area, and the ‘more protection’ option placed the SDR at half way between the area-based value and 0.05.

The Waste Water Management policies are based around whether homes use septic systems or sewer. Loads to either septic or sewer are based on projected population increases (Centre County Government, 1993). The different choices determine if the current percent of septic and sewer (U.S. Department of Commerce, 1990) levels are maintained during growth, if all new growth has sewer or septic, or if all wastewater becomes sewer.

The final set of policy choices concerned Agricultural BMPs. One BMP was to eliminate land application of manure that was assumed to occur during winter months (Evans and Zhu, 1998). Another BMP implemented contour tillage on all cropland, changing the USLE P factor for these areas. The other BMP adjusted the USLE C factor by implementing chisel tillage on all cropland (Jarrett, 1995).

These four policy areas cover only three of the four categories of concern for the Spring Creek Watershed: water quantity, water quality, and land development/protection. To incorporate the fourth category, cooperation and education, one additional facet is added to the policy choices. Policy choices can be made separately for the communities in the Centre Region versus the communities not in the Centre Region. Therefore, there are eight different policy areas; the four policy areas listed in Table 4.3 for the Centre Region and for policy areas for the rest of the watershed.

After deciding on the policy areas, a procedure was developed to change the original GWLF transport and nutrient files for each of the eight subwatersheds according to a scenario of policy choices. A computer program was written for this procedure, producing new GWLF input files. A second program was written to run GWLF, using these new input files, for each of the eight subwatersheds to produce GWLF output files. A third program was written to extract the nine outputs of concern for the eight subwatersheds and combine them into a file. A fourth program was written to compare the nine outputs with the base case and store the results. After initial testing of all four programs for one scenario of policy choices, a fifth program was written which produced a given number of policy scenarios, with randomly selected policy choices. A sixth and final program was written which used the inputs and outputs of the previous five programs to automate the entire data generation setup. The final result was a program that developed a given number of policy scenarios and produced files for each of the nine GWLF outputs reflecting the change from the base case caused by the policies. In addition, the program produced a file listing each of the policy scenarios.

The ten files produced by the final program (nine output files and the policy scenario file) were the training data for the ANN step of the research. Since the program automated the entire data generation step, it was easy to create new training data needed in the development of the ANNs.

4.4.2. Step Two: Artificial Neural Networks of GWLF Output

This step begins with the results from step one. Before proceeding to training the Forward-Backward Propagation ANNs, the relationships between the policy scenarios and the outputs will be explored. Table 4.4 shows what outputs affected each policy choice. The table shows that all three runoff outputs were only affected by land use changes, and detached sediment was only affected by land use changes and contour and chisel tillage options of the agricultural BMPs. The total runoff was found to change very little for any scenario of policies. Therefore, total runoff was eliminated from InterWET for this case study, because it did not show any significant change. Since the other two runoff outputs and detached sediment are only affected by a small number of policies, it was decided to use look-up tables instead of ANNs. These look-up tables were produced by running GWLF for all the possible combinations of policies that would produce different output levels for these three outputs. Appendix D contains the JavaScript calculators based on these look-up tables.

ANNs were developed for the five remaining outputs: delivered sediment, sediment nitrogen and phosphorous, and dissolved nitrogen and phosphorous. Following the trial and error procedures detailed by Caudill and Butler (1992a), preliminary testing was performed to determine the proper architecture, network parameters, number of training sets, and attainable minimum errors for training and testing. The first round of preliminary testing was used to determine the proper ANN architecture. The architecture refers to the number of inputs, hidden layer neurodes, outputs, and to their arrangement. The first round of preliminary testing began with an architecture having 16 bipolar (-1,1) inputs, 1 hidden layer with 16 neurodes, and 8 outputs, and with sigmoid transfer functions for both the hidden and output layer. The 16 inputs correspond to the 8 policy choice inputs (the four policy areas times the two regions in the watershed) represented in bipolar binary. The first choice for some policy would be -1 -1, second choice -1 +1, third choice +1 -1, and fourth choice +1 +1. The 8 outputs correspond to the percent change in delivered sediment, with precision to the tenths place, for the 8 subwatersheds being modeled. The preliminary testing also started with a learning factor = 0.9 and momentum = 0.0. The learning factor influences how fast a training network will learn the training patterns. The momentum causes the training to be influenced by previous training and sometimes cause a network to learn faster. The learning factor and momentum factors in the equations are used by the ANNs to train the weights in the network by reducing the output errors (Caudill and Butler, 1992a). The testing started with a 50 policy scenario training set and a different 250 policy scenario testing set.

Computer programs, based on FBP equations in Caudill and Butler (1992a), were written to train and test the ANN. The training programs consisted of four main parts. The first part converted the policy choice scenarios into 16 bipolar inputs. A policy choice was entered as a number from 1 to 4 and then was converted to a bipolar binary number, as detailed earlier. Therefore the 8 policy choices became 16 bipolar inputs.

The second part contained the training loop. Every policy choice scenario or training set was run through the FBP equations to train the model. One complete training cycle of all training patterns is called an epoch. The equations used to train a Forward Backward Propagation ANN, detailed in Caudill and Butler (1992a), are the forward activity equations (input to hidden to output layer) and the backward error propagation equations used to update the weights (output to hidden to input layer). Initially, weights between the input and hidden layers and between the hidden and output layers were set to some random value. Then, the weighted sums of the inputs were used with the hidden layer transfer functions, for this research sigmoid functions, to produce hidden layer outputs. Next, the weighted sums of these hidden layer outputs were combined in the output layer to predict the final output. Error values were determined by comparing the predicted output with the desired output for the training patterns. The backward error propagation equations then used these errors to adjust or train the weight between the output and hidden layers and between the hidden and input layers. After the weights had been trained for one pattern, the next pattern was introduced and the process repeated.

The third part of the FBP ANN computer programs consisted of controlling the number of epochs. This control came from setting a maximum number of epochs and setting a maximum error. The error used in this program was percent error of the testing patterns. The goal of ANN training is to produce a network that will predict a value within a certain amount of error. This amount of error is determined by testing a training ANN with some new set of testing data. In the preliminary testing, some initial trails showed that 75% testing error would be attainable for the five selected GWLF outputs, without excessive ANN memory requirement. At the end of each epoch, every testing pattern was tested and the testing error was determined from the number of incorrect output predictions. If this maximum error was not less than the maximum allowable, another epoch of training was performed, unless the total number of epochs exceeded the epochs limit. Both controls were needed, because the training might not be able to meet the maximum error limit in a reasonable number of epochs.

Once the training was stopped by the third part, the fourth part of the FBP ANN training program used the final weights and the forward activity equations to predict output for each training and testing pattern. The final weights and the training and testing errors were then written to a file. For cases when the training was stopped because the maximum error was smaller than the set limit, this would be the end of the training. For cases when the training stopped due to reaching the number of epochs limit, it could be decided to continue the training for more epochs or to increase the number of hidden layer neurodes, which gave the ANN more memory capacity.

Several rounds of preliminary testing were performed using these ANN programs. The purpose of the first round of preliminary testing was to evaluate the original representations of the input and output data. The first round of preliminary testing was run for about 30,000 epochs. The training error was low, but the testing error was too large. These results caused some rethinking of the representation of the output percent changes. Since many of the policy choices were very generalized, like "more riparian protection", the percent change values were generalized into 8 different categories, shown in Table 4.5. The table shows how 3-digit binary numbers were used to represent each of the eight categories. Since a 3-digit binary number was used for each of the eight subwatershed outputs, a 24-digit binary number represented all of the outputs from the ANN. The divisions between these categories were based on the ranges of values of the five outputs. In addition, the number of hidden layer neurodes was increased to 50 to help decrease the testing error. Training was adjusted so only those policy choices that affected sediment delivery, as shown in Table 4.4, were included. The values for the unaffecting inputs were changed to zero, insuring they would not add noise to the training.

Using this new setup, a second round of preliminary testing was done with 16 bipolar inputs, 50 hidden layer neurodes, and 24 binary outputs. In this round, different learning and momentum constants were used. For training with 50 training patterns and 250 testing patterns, a learning constant = 0.7 and a momentum constant of 0.0 were found to give the lowest testing errors. However, the testing errors were still excessive, so the number of training patterns was increased.

During the third round of preliminary testing, as the number of training patterns was increased, the training error began to increase and the testing error decreased. Testing found that 1000 training patterns and 100 testing patterns would produce networks that had testing error below 25%. However, each of the five GWLF outputs required different numbers of hidden layer neurodes (memory) to attain this goal. Dissolved phosphorous only required 100 hidden layer neurodes and easily had less than 25% testing error, while sediment-bound nitrogen required 400 hidden layer neurodes to barely have less than 25% testing error.

These findings determined the final setup of all five ANNs. Each Forward-Backward Propagation network, based on Caudill and Butler (1992a), used 1000 training patterns and 100 testing patterns, with 16 bipolar inputs (for the 8 policy choices) and 24 binary outputs (for the 8 change categories for the 8 subwatersheds of Spring Creek Watershed). Training used a learning constant of 0.7, momentum equal to 0.0, and all inputs that did not affect the outputs removed. The number of hidden layer neurodes was increased until testing error was below 25%.

4.4.3. Step Three: JavaScript Artificial Neural Networks

With the look-up tables for surface runoff, groundwater runoff, and detached sediment, and the trained ANNs for delivered sediment, sediment nitrogen and phosphorous and dissolved nitrogen and phosphorous, JavaScript calculators were produced for each output. Figure 4.5 shows the delivered sediment JavaScript calculator. Radio buttons were used for the policy choices. The user clicks one selection for each of the eight policy areas and then the calculate button. Text boxes for each of the eight subwatersheds display the percent change category for the particular policy scenario. The user can change their policy choice, recalculate, and see how the change in policy affected the delivered sediment.

Similar JavaScript calculator interfaces were used for eight of the GWLF outputs. As mentioned before, total runoff was not used because GWLF had not predicted any substantial changes in average annual values based on the policy choices. The calculators which employed look-up tables instead of ANNs had an additional set of radio buttons to display the output change either in the eight categories (Table 4.5) or as a percentage.

The results of these three steps were JavaScript calculators, which predict the impact on water resources resulting from the policy choices. The results section will detail the success of using these methods.

4.5. Results

4.5.1 Generation of GWLF training data

The methods section detailed the group of computer programs that were created to generate GWLF output for sets of policy scenarios. These programs were created, tested, and successfully used to generate the required data to produce either look-up tables or the training and testing data for the ANNs. A total of 1,100 out of the possible 65,536 policy scenarios were generated. 1,000 were used for training and 100 for testing. Certain scenarios included the most extreme change for each GWLF output.

As discussed earlier, Table 4.4 shows the policies that affect GWLF outputs. Table 4.6 shows the number of different possible policy scenarios for each GWLF output, the possible number of output values for all eight subwatersheds, and whether a look-up table, ANN, or neither was used for that output. For example, all three runoff outputs have 16 possible policy scenarios. This corresponds to four different land use policies for the Centre Region times four different land use policies for the rest of the watershed. Likewise, there are 128 possible values for each of the runoff output; 16 possible policy scenarios times 8 subwatersheds. However, inspection of the total runoff values reveal only two resulting values (0.0% and 0.2%), not 128 values. Therefore, modeling of the average annual change of total runoff was removed from InterWET.

Table 4.6 also shows that there were 1152 possible values for percent change of detached sediment. A look-up table was used instead of an ANN, because of memory considerations. An ANN using the architecture determined in the methods section would have 16*50 = 800 weights between the input and hidden layer and 50*24 = 1200 weights between the hidden and output layer, for a total of 2000 different values. Even if some of the weights between the input and hidden were eliminated, since they did not affect detached sediment, the total number of weights would still exceed the number of possible output values for detached sediment. Because the look-up tables and ANN weights are hard-coded in JavaScript, less memory was required by utilizing a look-up table for detached sediment.

A similar argument can be used to justify the use of ANNs with a relatively small number of hidden layer neurodes for the remaining GWLF outputs. A look-up for delivered sediment would take approximately as much memory as an ANN with 450 hidden layer neurodes. For the nutrient outputs, a look-up table would be equivalent to an ANN with approximately 800 hidden layer neurodes. Using ANNs with less than these numbers of hidden layer neurodes conserved memory. For Internet-based modeling, conserved memory translates into faster download and execution times.

4.5.2. Artificial Neural Network Training and Testing

Forward-Backward Propagation ANNs were produced for delivered sediment, dissolved nitrogen and phosphorous, and sediment nitrogen and phosphorous. These networks employed the setup detailed in the methods section as: 1000 training data sets, 100 testing data sets, 16 bipolar inputs, 24 binary outputs, learning constant of 0.7, no momentum, removal of all inputs which did not affect a particular output, and training until the average absolute testing error did not improve with additional training.

Table 4.7 shows the results from the training and testing of the ANNs for the five outputs. Because each output required different numbers of hidden layer neurodes, the number of training epochs also varied. All outputs had less than 25% testing error, the average at 18.4%. Dissolved phosphorous trained the best, with an overall testing error of only 9.0%. This error represents the accuracy for which the ANN will predict dissolved phosphorous for any given set of policy choices. The worst output was sediment nitrogen. Even though it had an average testing error of 24.6%, it had only 7.4% error for prediction within one change category. In other words, the change category for sediment nitrogen was usually correct or very close. Notice that all the outputs had less than 10% error for prediction within one change category. This level of accuracy is appropriate for the implementation of the ANN’s in the educational setting of InterWET.

4.5.3. JavaScript Implementation

JavaScript calculators were then developed for the outputs using look-up tables and trained ANNs. Look-up tables gave exact estimates of percent change of output, based on GWLF predictions of the outputs. Since the JavaScript ANN calculators used the same equations and weights as the trained ANNs, the error values in table 4.7 also apply to the JavaScript versions.

In addition to looking at the testing data errors, an "ad hoc" system verification or sensitivity analysis was performed on the JavaScript ANNs. While all other policies were kept constant, the values for one policy area were varied. These verification efforts responded as expected for the equations and calculations used in GWLF, although no numerical estimates of sensitivity were made. Since InterWET was designed as an educational tool, it was important to do this reality check to verify that the JavaScript ANNs would respond as expected to policy changes. As indicated earlier, the point behind using ANNs was to provide a fast, simple, but accurate, way of helping non-hydrology experts become familiar with policy impacts on water resources.

4.6. Conclusions

This chapter has shown the successful use of Artificial Neural Networks to assist hydrologic modeling on the Internet. GWLF predicted how different local government policies would change nine different average annual hydrologic loads in the Spring Creek Watershed in central Pennsylvania. Total runoff was found to not be impacted by the four policy choices used in InterWET. Surface runoff, groundwater runoff, and detached sediment loads were only affected by a small number of policies and were modeled in JavaScript using look-up tables. Since delivered sediment, dissolved nitrogen and phosphorous, and sediment nitrogen and phosphorous loads were affected by a large number of policies, forward-backward propagation Artificial Neural Networks were developed that accurately placed, on average, 81.6% of the load changes into one of eight change categories. JavaScript repeated the results of these neural networks.

The use of neural networks for Internet-based hydrologic modeling is just one example of balancing content, in the form of model accuracy, with speed. Web site designers are always balancing content and speed, often sacrificing one for the other. As hydrologists look to expand hydrologic models to the Internet, new tools will need to be employed which give the best accuracy possible while considering all on-line time lags. This research has shown that Artificial Neural Networks and JavaScript are two tools which can be used together for fast and relatively accurate hydrologic predictions in the educational context of the Internet Watershed Educational Tool.

4.7. References

Altenford, C.T. 1992. Using a neural network for soil moisture predictions. ASAE Microfiche No. 92-3557. St. Joseph, MI: ASAE.

Carline, R. 1998. Unpublished Data. University Park, PA: School of Forest Resources, The Pennsylvania State University.

Caudill, M. and C. Butler. 1992a. Understanding Neural Networks: Volume 1: Basic Networks. Cambridge, Massachusetts: The MIT Press.

Caudill, M. and C. Butler. 1992b. Understanding Neural Networks: Volume 2: Advanced Networks. Cambridge, Massachusetts: The MIT Press.

Centre County Government. 1993. Population trends and projections of the Spring Creek Watershed Community, 1980-2020. Bellefonte, PA: Office of Planning, Centre County Government.

Centre County Government. 1996. Centre County small water system regionalization study. Volume II: Regionalization Study. Bellefonte, PA: Office of Planning, Centre County Government.

ChingGung, W, and L. ChihSheng. 1998. A neural network approach to multiobjective optimization for water quality management in a river basin. Water Resources Research 34(3):427-436.

Cornell, G. and C.S. Horstmann. 1996. Core Java. Upper Saddle River, New Jersey: A Prentic Hall Title, SunSoft Press.

Daniell, T.M. 1991. Neural networks - Applications in hydrology and water resources engineering. Proceedings, International Hydrology and Water Symposium. Vol. 3:797-802. Nat. Conf. Publ. 91/22, Inst. of Eng., Australia, Barton, ACT, Australia.

Day, R. 1998. Spring Creek Watershed Landuse. University Park, PA: Land Analysis Laboratory, College of Agricultural Sciences, The Pennsylvania State University.

Edwards, L.D. 1995. The design and analysis of a mathematical microworld. Journal of Educational Computing Research 12(1):77-94.

Environmental Resources Research Institute (ERRI). 1998. Spring Creek Data Layers. University Park, PA: Environmental Resources Research Institute, The Pennsylvania State University.

Evans, B.M. and Y. Zhu. 1998. Simulation of hydrologic and nutrient transport processes for a portion of the Spring Creek Watershed in Centre County, PA. University Park, PA: Unpublished report from ABE 517, Agricultural and Biological Engineering Department, The Pennsylvania State University.

French, M., W.F. Krajewski, and R.R. Cuykendall. 1992. Rainfall forecasting in space and time using a neural network. Journal of Hydrology 137:1-31.

Haan, C.T., Barfield, B.J., and J.C. Hayes. 1994. Design hydrology and sedimentology for small catchments. New York: Academic Press.

Haith, D.A., R. Mandel, and R.S. Wu. 1992. GWLF: Generalized Watershed Loading Functions, Version 2.0, User’s Manual. Ithaca, New York: Department of Agricultural and Biological Engineering, Cornell University.

International Countryside Stewardship Exchange Steering Committee. 1996a. Spring Creek Watershed Community International Exchange Briefing Materials. State College, Pennsylvania: Clearwater Conservancy.

International Countryside Stewardship Exchange Steering Committee. 1996b. Spring Creek Watershed Community International Exchange Final Report. State College, Pennsylvania: Clearwater Conservancy.

InternetGIS.com. 1998. ActiveMaps Version 2.0. Oakton, Virginia: InternetGIS.com.

Jarrett, A.R. 1995. Water Management. Dubuque, Iowa: Kendall/Hunt Publishing Co.

Jenson, S.K. and J.O. Domingue. 1988. Extracting topographic structure from digital elevation data for geographic information system analysis. Photogrammetric Engineering and Remote Sensing 54(11):1593-1600.

Kao, J.-J. 1996. Neural net for determining DEM-based model drainage pattern. Journal of Irrigation and Drainage Engineering 122(2):112-121.

Laflen, J.M., Land, L.J., and G.R. Foster. 1991. WEPP-A new generation of erosion prediction technology. Journal of Soil and Water Conservation 46(1):34-38.

Lehning, D.W. 1998. Personal Communication. University Park, PA: Environmental Resources Research Institute, The Pennsylvania State University.

McClendon, R.W., Hoogenboon, G., and I. Seginer. 1996. Optimal control and neural networks applied to peanut irrigation management. Transactions of the ASAE 39(1):275-279.

Maier, H.R. and G.C. Dandy. 1996. The use of artificial neural networks for the prediction of water quality parameters. Water Resources Research 32(4):1013-1022.

Maier, H.R. and G.C. Dandy. 1997. Reply. Water Resources Research 33(10):2425-2427.

Myers, W.L. 1998. Pennsylvania GAP Analysis Terrabyte Data. CD-ROM. University Park, PA: Environmental Resources Research Institute. The Pennsylvania State University.

Novotny, V. and H. Olem. 1994. Water Quality. Prevention, Identification, and Management of Diffuse Pollution. New York: Van Nostrand Reinhold.

NRCS-MUIR. 1997. National MUIR Schema. Natural Resources Conservation Service, USDA. [On-line]. Available: http://www.statlab.iastate.edu/soils/muir/schema_nat.html.

Parson, S.C. 1998. Unpublished Software. University Park, PA: Agricultural and Biological Engineering, The Pennsylvania State University.

Parson, S.C. 1999. Development of an Internet watershed educational tool (INTERWET) for the Spring Creek Watershed of central Pennsylvania. Doctoral dissertation. University Park, Pennsylvania: The Pennsylvania State University.

Parson, S.C., J.M. Hamlett, P.D. Robillard, P.A. Johnson, and M. Urquidi-MacDonald. 1999. The Internet Watershed Educational Tool (InterWET). [On-line] Available: http://server.age.psu.edu/.

Raman, H. and N. Sunilkumar. 1995. Multivariate modeling of water resources time series using artificial neural networks. Hydrological Sciences Journal 40(2):145-163.

Ranjithan, S., Eheart, J.W., and J.H. Garrett Jr. 1993. Neural network-based screening for groundwater reclamation under uncertainty. Water Resources Research 29(3):563-574.

Schmalz, R. F. 1996. Water Resources and Hydrogeology of the Centre Region. Prepared for Board Members of the State College Water Authority with revisions and additions for Centre Region Council of Governments. Bellefonte, Pennsylvania: Centre County Government.

Seginer, I., Hwang, Y., Boulard, T., and J.W. Jones. 1996. Mimicking an expert greenhouse grower with a neural-net policy. Transactions of the ASAE 39(1):299-306.

Shamseldin, A.Y. 1997. Application of a neural network technique to rainfall-runoff modeling. Journal of Hydrology (Amsterdam) 199(3/4):272-294.

Shukla, M.B., Kok, R., Prasher, S.O., Clark, G., and R. Lacroix. 1996. Use of artificial neural networks in transient drainage design. Transactions of the ASAE 39(1):299-306.

Taylor, L.E. 1997. Water budget for the Spring Creek basin. Harrisburg, PA: Susquehanna River Basin Commission, Publication No. 184.

U.S. Department of Commerce. 1990. Washington, D.C.: Bureau of the Census, 1990 Census of population and housing, STF 3.

U.S. EPA. 1998. Water Discharge Permit Information. U.S. Environmental Protection Agency. [On-Line]. Available: http://www.epa.gov/enviro/html/pcs/pcs_overview.html.

Vander Veer, E.A. 1997. JavaScript for Dummies. 2nd Edition. Chicago, Illinois: IDG Books Worldwide, Inc.

Walling, D.E. 1983. The sediment delivery problem. In: I. Rodriguez-Iturbe and V.K. Gupta (Guest-Editors), Scale Problems in Hydrology. Journal of Hydrology (Amsterdam) 65:209-237.

Wischmeier, W.H. and D.D. Smith. 1978. Predicting Rainfall Erosion Losses - A Guide to Conservation Planning. USDA Handbook 537. Washington, D.C.: U.S. GPO.

Wooldridge, A. and M. Morgan. 1997. Special Edition: Using JavaScript, Second Edition. Indianapolis, Indiana: Que Corporation.

Yang, C.-C., Prasher, S.O., and R. Lacroix. 1996a. Application of artificial neural networks to land drainage engineering. Transactions of the ASAE 39(2):525-533.

Yang, C.-C., Prasher, S.O., and S. Sreekanth. 1996b. An artificial neural network model for pesticide fate and transport. ASAE Microfiche No.96-2025. St. Joseph, MI:ASAE.

Yang, C.-C., Prasher, S.O., and C.S. Tan. 1998. An artificial neural network model for water table management systems. In Drainage in the 21st century: food production and the environment. Proceedings of the Seventh International Drainage Symposium, Orlando, Florida, USA, 8-10 March 1998. St. Joseph, MI: ASAE.

TABLE 4.3. Policy Choices Modeled with GWLF for the Spring Creek Watershed.

Policy Area	Policy Choices
1. Land Development  (Changes in the areas of  different land use types.)	1.1. Growth Boundary
	1.2. Develop crop and pasture only
	1.3. Develop forest only
	1.4. Develop all undeveloped land
2. Riparians  (Changes in the sediment  delivery ratio of watershed.)	2.1. Same as current levels
	2.2. More developed
	2.3. More protected
	2.4. All protected
3. Waste Water System (Changes in the  sewage and septic  loads.)	3.1. Same as current ratios
	3.2. All new homes with sewer
	3.3. All new homes with septic
	3.4. All homes to sewer
4. Best Management  Practices or BMPs  (Changes in factors to  USLE or manure  application.)	4.1. Same as current levels
	4.2. No manure on crop
	4.3. Contour tillage on crop
	4.4. Chisel tillage on crop

TABLE 4.4. GWLF Outputs Affected by Policy Choices.

	Affected Outputs
Policy Codea	Surface Runoff	GW Runoff	Total Runoff	Detached Sediment	Deliver Sediment	Dissolved N	Dissolved P	Sediment N	Sediment P
1.1.	X	X	X	X	X	X	X	X	X
1.2.	X	X	X	X	X	X	X	X	X
1.3.	X	X	X	X	X	X	X	X	X
1.4.	X	X	X	X	X	X	X	X	X
2.1.					X			X	X
2.2.					X			X	X
2.3.					X			X	X
2.4.					X			X	X
3.1.						X	X
3.2.						X	X
3.3.						X	X
3.4.						X	X
4.1.				X	X	X	X	X	X
4.2.						X	X
4.3.				X	X	X	X	X	X
4.4.				X	X	X	X	X	X

^a Codes for each policy given in Table 4.3.

^aX represents the percent change of a GWLF prediction of a water resource load, between current conditions and future conditions based on a set of local government policies.

^aThese values represent the total possible number of combinations of the policy scenarios listed in Table 4.3 that will produce unique GWLF output values.
^bThese values represent the total number of unique GWLF output values.
^cArtificial Neural Networks.

^aEach epoch represents training performed on all 1000 training patterns.
^bA higher number of hidden layer neurodes, the greater the memory capacity of the artificial neural network.
^cPercent of artificial neural network output, in the 8 change categories in Table 4.5, from all 1000 training data sets from which did not match the GWLF output for those training data sets.
^dPercent of artificial neural network output, in the 8 change categories in Table 4.5, from all 100 testing data sets from which did not match the GWLF output for those testing data sets.
^ePercent of artificial neural network output, in the 8 change categories in Table 4.5, from all 1000 training data sets from which where not within ± 1 change category of the GWLF output for those training data sets.
^fPercent of artificial neural network output, in the 8 change categories in Table 4.5, from all 100 testing data sets from which where not within ± 1 change category of the GWLF output for those testing data sets.

Back to Table of Contents

Next to Chapter 5