This file was created by scanning the printed publication. Errors identified by the software have been corrected; however, some errors may remain. Knowledge-Based Approach to Watershed-Scale TMDL Assessment1 Keith Reynolds2 Mark Jensen 3 James Andreasen 4 Iris Goodman 5 Abstract-The Ecosystem Management Decision Support (EMDS) system is an application framework for knowledge-based decision support of ecological landscape analysis at any geographic scale. The system integrates geographic information system and knowledge base system technologies to provide an analytical tool for environmental assessment and monitoring. The basic objective ofEMDS is to improve the quality and completeness of environmental assessments and the efficiency with which they are performed. The USDA Forest Service and Environmental Protection Agency have cooperatively developed an EMDS knowledge base for assessment and monitoring of ecological states and processes in 6th code watersheds. The knowledge base evaluates watershed processes, patterns, general effects of human influence, and specific effects on salmon habitat. The Total Maximum Daily Load (TMDL) program, Section 303(d) of the Clean Water Act, identifies sources of pollution remaining after end-of-pipe discharges are regulated and a pplying the best available technology. Remaining sources of pollutant are termed non-point sources (NPS). Under requirements ofthe Act, States develop lists of waters that do not meet State water quality standards, even after point sources of pollution have installed required levels of pollution control technology, States must establish priority rankings based on severity of pollution and beneficial uses of water bodies, such as recreation or fishing, and must develop TMDLs for waters on the lists. TMDLs specify amounts of pollutants that need to be reduced to meet State water quality standards and allocate pollution control responsibilities among pollution sources in a watershed. The U.S. Environmental Protection Agency (EPA) has established a five-step approach to setting TMDLs: (1) identify waters requiring TMDLs; (2) priority ranking and targeting; (3) develop TMDLs; (4) implement control actions; and (5) assess control actions. Ipaper presented at the North American Science Symposium: Toward a Unified Framework for Inventorying and Monitoring Forest Ecosystem Resources, Guadalajara, Mexico, November 1-6,1998. 2Keith Reynolds is a Research Forester, USDA Forest Service, Pacific Northwest Research Station, 3200 SW Jefferson Way, Corvallis, OR 97331, Phone: (541) 750-7434; Fax: (541) 750-7329; e-mail: reynoldsk@fsl.orst.edu 3Mark Jensen is aHydrologist, USDA Forest Service, Northern Region Headquarters, located at Missoula, MT. 4James Andreasen is a Research Fisheries Biologist, U.S. Environmental Protection Agency, National Center for Environmental Assessment, located at Washington, DC, Headquarters. sIris Goodman is a Research Ecologist, U.S. Environmental Protection Agency Landscape Ecology Branch, located at Las Vegas, NV. USDA Forest Service Proceedings RMRS-P-12. 1999 Conventional TMDL development is carried out on individual stream reaches by analyzing impaired conditions. Conventional methods for TMDL analysis cannot address the spatial and temporal scales required to: (1) establish adequate reference conditions for NPS parameters; (2) estimate the predictive capabilities of scale relations for spatially continuous ecoregions; (3) project likely scenarios of water quality change due to changes in land use, cover, or climate; (4) relate monitoring technologies and standards to defined ecoregional scales; and (5) establish schedules for TMDL development that are ecologically meaningful and compatible with Federal Agency responsibilities under the Endangered Species Act. The EPA Office of Research and Development and the Forest Service (U.S. Department of Agriculture) are cooperatively developing new analytical techniques for landscape-scale TMDL assessment, using knowledge-based processing of landscape databases that enable environmental managers to make better decisions. The objectives of this study were to design a knowledge base as a logical framework for assessment of 6th code watershed condition and illustrate its application in landscape analysis with the Ecosystem Management Decision Support (EMDS) system (Reynolds 1999a; Reynolds and others 1997a, 1997b). Materials and Methods _ _ _ __ NetWeaver Knowledge Bases This section summarizes key concepts and constructs related to design and use of NetWeaver knowledge bases (Stone and others 1986). Reynolds (1999b) gives a more detailed description of the technology as implemented in EMDS. Formally, a knowledge base is a meta database that provides a specification for interpreting information. Knowledge bases in this sense effectively are cognitive maps of the elements in a problem domain and the logical relations among those elements. In the context of watershed assessment, for example, the elements of the problem are typically ecosystem states and processes related to vegetation structure and composition, water quality, stream flow properties, etc. The primary structural element of a knowledge base as implemented in N etWeaver is the network whose function is to evaluate a proposition. The key attribute of a network is its truth value, which is a measure of the degree to which the proposition is true, based upon the state of logically antecedent conditions. NetWeaver networks are recursive 81 insofar as a network may be evaluated in terms of other networks. For example, the network for watershed processes (Figure 1) is evaluated in terms of its logically antecedent networks hydrologic processes, erosion processes, and fire processes (hereafter, NetWeaver objects are identified in bold type). Thus, the proposition that watershed processes are within a suitable range of reference conditions is true to the degree that the propositions associated with its logically antecedent networks are true. The network architectures under hydrologic processes, erosion processes, and fire processes define the manner in which these networks are evaluated in turn and so on. Logical operators in NetWeaver such as AND and OR are fuzzy logic operators. That is, they perform fuzzy math operations that propagate truth values, derived at the level of data links, upward through the logical structure of a knowledge base. Zadeh (1965, 1968) presented basic concepts of approximate reasoning with fuzzy logic. Subsequent concept papers (Zadeh 1975a, 1975b, 1975c) elaborated on the syntax and semantics oflinguistic variables, laying the foundation for what has become a significant new branch of mathematics. Fuzzy logic is concerned with quantification of set membership and associated set operations. Data links (graphically illustrated as rectangles in figures) essentially are elementary networks. Like networks, data links may evaluate a proposition, yielding a truth value (although data links do not necessarily evaluate anything). The primary distinction between networks and data links that yield truth values is that data links only evaluate data rather than a logical expression of antecedent conditions. Data links evaluate a proposition by comparing the value of a data item, or the result of a mathematical expression involving one or more data items, to an argument that defines the conditions under which the proposition is considered true. An argument may test for a simple true/false condition as in classical rule-based systems based on bivalent logic, or an argument may be a fuzzy membership function that tests an observed value's degree of membership in a fuzzy subset (Kaufmann 1975, Zadeh 1992). A fuzzy membership function provides an explicit mathematical expression for testing an observation's degree of affinity for the concept represented by the fuzzy subset. Problem Domain Given the objectives of the EPA water quality assessment program discussed in the introduction, the primary knowledge base topics included in design were watershed processes, watershed patterns, general effects of human influence, and specific effects of human influences on aquatic species (Table 1). A key decision, made early in the design process, was that the method of assessment be sufficiently general for application in any geographic region. This design criterion was implemented by constructing all fuzzy membership functions as dynamically-defined functions of data representing standards. That is, all fuzzy membership functions in the knowledge base are defined by standards input during analysis. All data are evaluated by comparison to standards for which we conceptually distinguish three basic types: reference conditions representing attributes of unmanaged watersheds, management standards set by resource management agencies such as the USDA Forest 82 Figure 1.-Network for watershed processes. The truth of the proposition that watershed processes are within a suitable range of conditions depends on the degree to which its three premises, represented by the networks, hydrologic processes, erosion processes, and fire processes, are true. Service, and regulatory standards set by regulatory agencies such as the EPA. Knowledge Base Application in EMDS Major components of the EMDS system (Reynolds 1999a) include the NetWeaver knowledge base system, the EMDS Arcview application extension, and the Assessment system (Figure 2). This section briefly summarizes system structure and function in terms of system level objects, their methods, and relations. More detailed descriptions of the system are provided in Reynolds (1999a) and Reynolds and others (1997a, 1997b). The NetWeaver knowledge base system (Reynolds 1999b) is composed of an engine and a graphic user interface for knowledge base developers that provides controls for designing, editing and interactively evaluating knowledge bases (Figure 2). Primary components ofthe EMDS Arcview application extension are the DataEngine and MapDisplay objects that customize the Arcview environment with methods and data structures required to integrate NetWeaver's knowledge-based reasoning schema into Arcview (Figure 2). The Assessment system is a graphic user interface to the NetWeaver engine for EMDS application end-users that Table 1.-Primary networks in the knowledge base for assessing watershed condition. Network name watershed processes watershed patterns human influence aquatic species Proposition evaluated by network Watershed processes are within acceptable ranges. Watershed patterns are within acceptable ranges Aggregate effects of human influence are within acceptable ranges. Likelihood of longterm viability of aquatic species is good. USDA Forest Service Proceedings RMRS-P-12 . 1999 ,/ - i / NetVVeaver -', : , ( -_ /"'--'--\--~-) \' Arcview, ~ - oelectNetwolkO EM~~ , ( system, requeotThemeO " ' __ interface:' ~~~;;sme~t - - - " " _, reqUestsMap"-- -- uses ,_ - / /uses / __ , \ Ne~ea~er eng~~e- - " ,/ "c:':~=~)O evaluateNetwolkO /' " "":~r()~ readstw rites 1 -~O~I~nk :r:~;~rdioplayState() / ~ ,_ - - _' - 'DAM \./ ,,,d, readSIW~rites data~nk - - ': / ,d::t~~:~:~O ··"'·~"f'.a(L - - ' I __ -, /~__' ___ . - \!","~m" netwolk {pe rs io l MapDlsplay /' -:' ',' ,--- ____ ,'--, __ cddThemeO asseosArea \ \ __ ' \ 1 / - - - - -' - - - - _ _ " Knowledge base " --~~, requestS~Data use\ { ..:-uses""-:>-- - -' / ,/ - _, c~::~~)~: ] -\ - -' -' I, ' J 1 .. n /"--/'--/~h;~\ /readSlWrites 1.. n _ '-/ - - - ''/1 .. n ,,' AssessTables ' ~ , datalinkState: - - netwolkState I " {persis} ___ ) , __ , ;sData DataEngine- , ___ , } _ ,( GeoData - " ' , _ {pe 1$ is} ~ 1 n)- _,' __ " ,, _ TabData - " {pe ... is} ~ Figure 2.-System level object diagram of EMDS system. Lines indicate object relations and annotations on lines indicate primary nature of the relation ("uses" indicates a general relation in which several to many methods of the used object are relevant). Text items within objects of the form "xxx 0" indicate object methods. Only key methods are shown for each object. controls setup and running of analyses, runtime editing of knowledge bases, and display of maps, tables, graphs, and evaluated knowledge base state related to analyses. broad in conceptual scope, requiring evaluation of possibly numerous and diverse data. Consequently, several to many data elements needed for complete evaluation of a knowledge base or any of its components may be missing at the Results --------------------------------------The primary networks for assessing watershed condition are watershed processes, watershed patterns, human influence, and aquatic species. Each network evaluates a specific proposition about the state of watershed condition (Table 1). An example analysis of erosion processes in a portion of the Columbia River Basin was performed to illustrate landscape application of the knowledge base for watershed copdition in EMDS (Figure 3). The Assessment system (Figure 2) was used to specifically select the erosion processes network (Figure 1) for evaluation in our example. In general, the Assessment system can be used to select any combination of networks for analysis. Map output shows the computed truth value for the proposition that erosion processes are within a suitable range of conditions th for each 6 code watershed in the assessment area selected for this example. Partial evaluations, based on currently available data, can be performed in EMDS. Truth values for erosion processes in the map output (Figure 3) only reflect a partial evaluation of the network because data values for volumes of mass wasting and debris avalanche are missing in our example (Table 2). Ecological assessments frequently are USDA Forest Service Proceedings RMRS-P-12. 1999 Legend Figure 3.-Truth value map for the proposition that erosion processes in 6th code watersheds are within a suitable range of reference conditions. 83 Table 2.-Propositions associated with networks antecedent to the erosion processes network. Network name erosion processes surface erosion mass wasting debris avalanche sediment delivery Proposition Erosion processes are within suitable ranges. Amount of surface erosion is within a suitable range. Amount of mass wasting is within a suitable range. Amount of debris avalanche is within a suitable range. Amount of sediment delivery is within a suitable range. start of an assessment. In our example, complete evaluation of erosion processes requires data values for volumes of surface erosion, sediment delivery to streams, mass wasting, and debris avalanche, but only the first two data elements were available at the time of analysis. However, given the set of knowledge base objects and their logical organization within the knowledge base, the NetWeaver engine computes the relative influence of missing data (Figure 4). Finally, the Hotlink browser (Figure 2) provides a means to examine details underlying an evaluation, by allowing the user to view the evaluated state of the knowledge base for any landscape feature selected on a truth value map (Figure 5). Discussion --------------------------------Application of fuzzy logic to natural resource science and management is still relatively new. General areas of application include classification in remote sensing (Blonda 1996), environmental risk assessment (Holland 1994), phytosociology (Moraczewski 1993a, 1993b), geography (Openshaw 1996), ecosystem research (Salski and Sperlbaum 1991), and environmental assessment (Smith 1995, 1997). More specific applications include catchment modeling (Anonymous 1994), cloud classification (Baum et a1. 1997), evaluation of plant nutrient supply (Hahn et a1. 1995), soil interpretation (Mays et a1. 1997, McBratney and Odeh 1997), and land suitability for crop production (Ranst et a1. 1996). The knowledge base for evaluation of watershed condition was designed for general application. The architecture is such that the knowledge base should be applicable in any geographic region with no more than minor adaptation. Specification of standards as data to be read from a database is an important ingredient of this general applicability. Clearly, our approach to a general solution begs the question, ''Where do specifications for reference conditions come from?" We suggest the following approach. For any geographic region or subregion, the vegetation potential of unmanaged watersheds is conditioned by geographic and climatic factors (Whittaker 1975). Widely available synecological analysis tools such as detrended correspondence analysis (Hill and Gauch 1980) provide a basis for arranging watersheds along geographic and climatic gradients, and identifying groupings indicative of reasonably separable vegetation potentials in the absence of management. Most resource management agencies have sufficiently detailed GIS coverages to identify watersheds that have experienced little or no management, and the attributes of such watersheds can be used as reference conditions. The knowledge-based reasoning schema of NetWeaver uses an object- and fuzzy logic-based propositional network architecture for knowledge representation (Reynolds 1999b). The system facilitates evaluation of complex, abstract topics such as water quality that depend on numerous, diverse subordinate conditions because NetWeaver is fundamentally logic based. The object-based architecture ofN etWeaver knowledge bases is conducive to incremental, evolutionary design of complex knowledge representations (Booch 1994) which has been recognized as crucial to successive design of complex systems (Gall 1986). The propositional network architecture of NetWeaver knowledge bases allows both the ability to evaluate the influence of missing information and the ability to reason with incomplete information (Reynolds and others 1997a, 1997b). Use of fuzzy logic in NetWeaver affords significant practical advantages over Bayesian belief networks (Ellison Relative influence • debrisAvMean II massWasteMean • sedDelivMean • debrisAvSD o surfErosMean Iilll massWasteSD Q] sedDelivSD • surfErosSD II massWasteQ4 • massWasteQ3 o massWasteQ2 o 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Figure 4.-Relative influence of missing data with respect to completing an analysis of erosion processes. 84 USDA Forest Service Proceedings RMRS-P-12 .1999 ..................... ......... 1 ....... .........•......... ~::::::::~::::::: Figure 5.-The EMDS NetWeaver browser displays the evaluated state of a knowledge base for selected landscape features in EMDS map outputs. 1996, Howard and Matheson 1981) and classical rule-based knowledge representations that depend on bivalent (e.g., yes/no or true/false) logic (Waterman 1986, Jackson 1990) in the context of knowledge bases that are conceptually broad and that include a wide variety of topics. Bayesian belief networks work well on narrow, well-defined problems, and may be preferable to fuzzy logic networks when conditional probabilities of outcomes are known. However, Bayesian belief networks are difficult to apply to large, general problems because the number of conditional probabilities that must be specified can quickly become extremely large as the conceptual scope of a problem increases. In such situations, model design not only becomes difficult to manage, but many probabilities will not be well characterized and will therefore need to supplied by expert judgment, thus negating much ofthevalue to be gained by a more statistically-based approach to knowledge representation. Similarly, the number of rules required in a bivalent logic knowledge base increase to unmanageable levels as soon as the model designer attempts to account for shades of outcomes such as poor, fair, good, excellent, etc. These arguments should not be taken to infer that fuzzy logic networks are inherently superior to other forms of knowledge representation. On the contrary, the various methods just discussed may be highly complementary to one another. In particular, we believe that fuzzy logic networks are ideally suited as logical frameworks for integrating model results from a variety of analytical systems such as simulators, linear programs, Bayesian belief networks, and rule bases. USDA Forest Service Proceedings RMRS-P-12. 1999 Conclusions __________ A knowledge-based approach to landscape analysis for TMDL assessment was shown to be quite feasible with application of the EMDS system despite the broad conceptual scope of the problem domain. The complete knowledge base has large data requirements, but any combination of networks, representing subsets of the full knowledge base, may be selected for analysis. Key advantages of a landscape analysis based on fuzzy logic networks as implemented in NetWeaver and used in EMDS include the ability to reason with incomplete information, and the ability to evaluate the influence of missing information. Fuzzy-logic based landscape analysis may be most useful for construction oflogical frameworks wi thin which a wide variety of analytical resul ts can be effectively integrated into a single, coherent analysis. Literature Cited Anonymous. 1994. Fuzzy logic applied to catchment modelling. Water & Wastewater International 9:40-45. Baum, Bryan A.; Tovinkere, Vasanth; Titlow, Jay; Welch, Ronald M. 1997. Automated cloud classification of global AVHRR data using a fuzzy logic approach. Journal of Applied Meteorology 6: 1519-1526. Blonda, P.; Bennardo, A.; Satalino, G.; Pasquariello, G. 1996. Fuzzy logic and neural techniques integration: an application to remotely sensed data. Pattern Recognition Letters 17:1343-1347. Booch, G. 1994. Object-oriented analysis and design. New York: Benjamin/Cummings Publishing Company. 578 p. 85 Ellison, A M. 1996. An introduction to Bayesian inference for ecological research and environmental decision-making. Ecological Applications 6:1036-1046. Gall, J. 1986. Systematics: how systems really work and how they fail. Ann Arbor, MI: The General Systematics Press. 342 p. Hahn, A; Pfeiffenberger, P.; Wirsam, B.; Leitzmann, C. 1995. Evaluation and optimization of nutrient supply by fuzzy logic. Ernahrungs-Umschau 42:367-375. Hill, M. 0.; Gauch, H. G. Detrended correspondence analysis: an improved ordination technique. Vegetatio 42:47-58. Holland, J. M. 1994. Using fuzzy logic to evaluate environmental threats. Sensors 11:57-62. Howard, R; Matheson, J. 1981. Influence diagrams. Pages 721-762 in Howard, R, and Matheson, J" eds., Readings on the principles and applications of decision analysis, Volume II. Menlo Park, CA: Strategic Decisions Group. Jackson, P. 1990. Introduction to expert systems. Addison-Wesley Publishers. Reading, MA 526 p. Kaufmann, A 1975. Introduction to the theory of fuzzy subsets. Volume 1. Fundamental theoretical elements. New York: Academic Press. 416 p. Mays, M. D.; Bogardi, 1.; Bardossy, A 1997. Fuzzy logic and riskbased soil interpretations. Geoderma 77:299-309. McBratney, A B.; Odeh, 1. O. A. 1997. Application of fuzzy sets in soil science: Fuzzy logic, fuzzy measurements and fuzzy decisions. Geoderma 77:85-91. Moraczewski, 1. R 1993a. Fuzzy logic for phytosociology 1. Syntaxa as vague concepts. Vegetatio 106:1-12. Moraczewski, 1. R 1993b. Fuzzy logic for phytosociology 2. Generalizations and prediction. Vegetatio 106:13-27. Openshaw, S. 1996. Fuzzy logic as a new scientific paradigm for doing geography. Environment & Planning A 28:761-767. Ranst, E. Van; Tang, H.; Groenemans, R; Sinthurahat, S. 1996. Application offuzzy logic to land suitability for rubber production in peninsular Thailand. Geoderma 70:1-12. Reynolds, K M. 1999a. EMDS users guide (version 2.0): knowledgebased decision support for ecological assessment. Gen. Tech. Rep. PNW-GTR Portland, OR: U.S. Department of Agriculture, Forest Service, Pacific Northwest Research Station. (in press). Reynolds,KM.1999b.NetWeaverforEMDSversion2.0userguide: a knowledge base development system. Gen. Tech. Rep. PNWGTR Portland, OR: U.S. Department of Agriculture, Forest Service, Pacific Northwest Research Station. (in press). 86 Reynolds, K; Saunders, M.; Miller, B.; Murray, S.; Slade, J. 1997a. An application framework for decision support in environmental assessment. Pages 333-337 in: GIS 97 Conference Proceedings of the Eleventh Annual Symposium on Geographic Information Systems. Vancouver, BC. February 17-20, 1997. Washington, DC: GIS World. Reynolds, K; Saunders, M.; Miller, B.; Slade, J.; Murray, S. 1997b. Knowledge-based decision support in environmental assessment. Pages 344-352 in: Resource Technology 97 Conference Proceedings Volume 4. ACSM 57 th Annual Convention. ASPRS 63 rd Annual Convention. Seattle, WA April 7-10, 1997. Bethesda, MD: American Society for Photogrammetry and Remote Sensing and American Congress on Surveying and Mapping. Salski, A; Sperlbaum, C. 1991. Fuzzy logic approach to modelling in ecosystem research. Lecture notes in computer science XX:520-527 Smith, P. N. 1995. A fuzzy logic evaluation method for environmental assessment. Journal of Environmental Systems 24:275-279. Smith, P. N. 1997. Environmental project evaluation: a fuzzy logic based method. International J oumal of Systems Science 28:467-471. Stone, N. D.; Coulson, R N.; Frisbie, R E.; Loh, D. K 1986. Expert systems in entomology: three approaches to problem solving. Bulletin of the Entomological Society of America 32, 161-66. Waterman, D. A. 1986. A guide to expert systems. Addison-Wesley Publishers, Reading, MA 419 p. Whittaker, R H. 1975. Communities and ecosystems. New York: MacMillan Publishing. 385 p. Zadeh, L. A 1965. Fuzzy sets. Information and Control 8:338-353. Zadeh, L. A 1968. Probability measures of fuzzy events. J. Math. Anal. and Appl. 23:421-427. Zadeh, L. A 1975a. The concept of a linguistic variable and its application to approximate reasoning. Part 1. Information Science 8:199-249. Zadeh, L. A. 1975b. The concept of a linguistic variable and its application to approximate reasoning. Part II. Information Science 8:301-357. Zadeh, L. A 1975c. The concept of a linguistic variable and its application to approximate reasoning. Part III. Information Science 9:43-80. Zadeh, L. A 1992. Knowledge representation in fuzzy logic. Pages 1-25 in: An introduction to fuzzy logic applications in intelligent systems (RR Yager and L.A Zadeh, eds.). Boston, : Kluwer Academic Publishers. 356 p. USDA Forest Service Proceedings RMRS-P-12 . 1999