AI and vegetation management in forestry

[fore057 at csc.canterbury.ac.nz]

Quite a number of people asked me to send copies of this when it was complete, 
so I'm posting it on the board.  Any comments would be welcome.



       APPLICATION OF KNOWLEDGE-BASED PROGRAMMING TECHNIQUES TO
         COST-EFFECTIVE SELECTION OF HERBICIDES IN FORESTRY


       E.G. Mason, D.J. Geddes, B. Richardson & N.A.  Davenhill School
    of Forestry, University of Canterbury; Tasman Forestry Limited;
              and New Zealand Forest Research Institute

      This paper will be presented to the Joint Australian and New Zealand 
    Institutes of Foresters Conference on "The costs and benefits of change",
                Christchurch, New Zealand, September 1991


                            ABSTRACT


    Knowledge-based programming techniques were employed to build a
    PC-based system for selecting herbicides, to improve the
    cost-effectiveness of vegetation management regimes and increase
    users' awareness of environmental hazards.  The strategy used to
    develop the tool is explained, and the structure of the system
    described.  It is written in PDC Prolog, and incorporates both
    knowledge-based and traditional procedural programming structures.
    The program is very user-friendly, with capabilities specifically
    relevant to forest supervisors.  It may be adapted to different
    localities and herbicide/weed regimes without further programming.
    The benefits of having this kind of decision-support tool in
    managing plantations with the intensity and sensitivity needed
    today are discussed.


                          INTRODUCTION


    A vegetation management adviser is one essential component of any
    forest management decision-support system, and knowledge-based
    programming techniques provide an excellent way to accomodate such
    a tool.  Jeffers (1989) and Mason (1989) outlined the form which
    future computerised decision-support systems may take in forestry.
    User-friendly, comprehensive and malleable environments are
    possible, within which managers can select the types of analyses
    they desire.  These environments may comprise geographic
    information systems, growth and yield models, other stand models,
    forest-level models (simulators, linear programming, and dynamic
    programming combinations), and other such useful tools.  It is
    knowledge-based programming, however, which enables full
    integration of the tools, and which fills the gaps hitherto
    occupied by handbooks, rules of thumb and/or experts.

    Knowledge-based programming is a name given to a set of computer
    programming techniques which enable machines to represent and
    process qualitative, symbolic information in a logical way.
    Saarenmaa (1989) comprehensively outlined these techniques within
    a forestry context.  Several varieties of knowledge-based
    programming are available, but the two most commonly employed in
    everyday applications are rule-based systems and object-oriented
    programming.  Expert systems are a subset of knowledge-based
    programming applications.

    The vegetation management components of forestry decision-support
    systems are best implemented in a knowledge-based structure.
    Design of vegetation management strategies or "regimes" involves
    many non-numerical analyses.  Experienced managers acquire a
    qualitative understanding of the components of the problem: for
    example, susceptibility of weeds to different herbicides; times of
    year weeds are physiologically active; behaviour of weeds
    following alternative treatments; effects of different weeds on
    tree crops; and so on.  This type of knowledge currently defies
    numerical analysis.


           WHY BUILD A VEGETATION MANAGEMENT ADVISER?


    Criteria for expert system domain selection

    Stock (1987) proposed the following seven criteria for a suitable
    expert system domain, which term in this context means "knowledge
    area represented".  Designing a vegetation management regime meets
    these criteria.

        1) Expertise should be scarce and time consuming to learn, but
        the task should take only a few hours or days.

        Tasman Forestry Ltd., for example, employs an expert (D.  J.
        Geddes) in vegetation management, who acquired his knowledge
        from many years of field experience.  Field supervisors vary
        in their abilities to design cost-effective vegetation
        management strategies, and often rely on the recommendations
        of a single expert within the organisation.  During a test at
        Tasman Forestry Ltd., supervisors prescribed treatments in
        response to the same weed problems: their solutions varied in
        cost by a factor of three (Geddes pers. comm.).  In some cases
        the treatments would have been unnecessarily expensive, whilst
        in others they would have had a low level of control.

        2) The problem domain should be narrow, but deep (highly
        specialised), and there should be a large number of possible
        solutions.

        Forest managers proficient in the design of vegetation
        management regimes are specialists with an in-depth
        understanding of the biology of local weed species and the
        effects of many treatment alternatives.  Different chemicals,
        and/or different physical treatments are available, as set out
        by, for example, Preest (1985), Davenhill (1985), and Preest &
        Davenhill (1986) for the New Zealand scene.  When these are
        considered over a range of weed species, environments, and
        seasons, the number of possibilities is large.

        3) The problem solution should require heuristics (rules of
        thumb), ie: a set of equations could not be used to arrive at
        a satisfactory solution.

        Given the range of qualitative rules required for effective
        design of vegetation managment regimes, it is unlikely that a
        set of equations would be adequate.  In part this is because
        models of weed behaviour are almost entirely qualitative, and
        strategies for their control have often arisen from field
        experience rather than from quantitative research.

        4) Competent experts must be available and willing to help
        with development.

        In the system described here, one company expert (D.  J.
        Geddes) and one research expert (N. A.  Davenhill) were much
        involved in pooling their knowledge and interpreting knowledge
        stored in data-bases.

        5) The problem should be financially important enough to
        warrant building the system.

        Based on responses to a questionnaire which asked for areas to
        be treated, it was estimated that New Zealand's forest
        industry planned to spend approximately $7,000,000 annually on
        vegetation management between 1987 and 1992 (Trewin & Mason
        1989). The direct costs of effective vegetation management can
        vary from just a few tens of dollars to several hundred
        dollars per hectare, while the opportunity costs of
        misapplication of techniques can be very high, in the form of
        poor subsequent crop performance, or as unnecessary
        expenditure.

        6) Experts in the area should agree.

        In New Zealand there is general agreement among experts about
        the broad principles of vegetation management regime design.
        Davenhill and Geddes occasionally differed in opinions but
        only on points of detail.

        7) Ample data, test cases, and potential users should be
        available for testing the system.

        Data, test cases and users were all available.  Geddes (1987)
        had compiled a very complete vegetation management manual for
        Tasman Forestry Ltd., and the company's forest supervisors
        were keen to help with the project.

    Environmental hazards

    Herbicides vary in their impact on the environment (Adams, 1988),
    and managers should avoid using particular products in
    circumstances where their use may pose a risk to adjacent crops,
    wildlife, fisheries or people.

    A computerised vegetation management adviser could accurately and
    quickly alert supervisors when use of a herbicide may be risky.
    In the system described here, warnings of potential hazards are
    brought to the user's attention when a herbicide is selected, and
    toxicity information is available at the touch of a key.

    Training

    For inexperienced supervisors, a decision-support system could be
    used to assist with training.  It is difficult for them to cope
    with the wide range of substances, application rates and methods,
    non-chemical control methods, responses of weeds, and costs
    involved in vegetation management.  A computerised system can make
    the problem more manageable, without removing them from the
    decision process.

    Identification of gaps in knowledge

    When knowledge is collated within a decision-support package, it
    is common that important gaps in knowledge are highlighted.  It
    was therefore expected that the project would identify research
    opportunities.

    Component of a comprehensive decision support system

    Future forestry decision-support systems are likely to comprise a
    range of models, tools, and databases, and a vegetation management
    adviser was deemed to be an important component of such a system.
    As this was the first knowledge-based application implemented by
    Tasman Forestry Ltd., it would serve as an indication of the
    potential for such systems.


                    CONSTRUCTION METHODOLOGY


    Construction of the system proceeded in four distinct stages: an
    initial protoype; knowledge acquisition; coding; and a
    testing/adjustment cycle.

    Initial prototype

    A small prototype system was devised as a result of a brief
    meeting with Tasman Forestry Ltd. staff, based on some information
    contained in the firm's Weed Control manual.  This was a crude
    program, written in BASIC, which contained knowledge of three
    weeds and ten herbicides.  It served to illustrate the potential
    for a knowledge-based system, and it elicited specific suggestions
    for im