This paper proposes a fuzzy expert system to calculate the potential environmental impact of the application of a pesticide in a field crop. It has been submitted for publication to a peer-reviewed journal for possible publication. Comments on this paper can be addressed to Hayo van der Werf.

Version 23 May 1997

Hayo M.G. van der Werf & Christophe Zimmer

INRA, Station d'Agronomie, 28 rue de Herrlisheim, BP 507, 68021 Colmar, France

Email: vdwerf@colmar.inra.fr

Pesticide use options available to farmers differ strongly with respect to the risks they pose to the environment. This paper proposes a fuzzy expert system to calculate an indicator "Ipest" which reflects an expert perception of the potential environmental impact of the application of a pesticide in a field crop. We defined four modules, one reflecting the presence (rate of application) of the pesticide, the other three reflecting the risk for three major environmental compartments (groundwater, surface water, air). The input variables for these modules are pesticide properties, site-specific conditions and characteristics of the pesticide application. For each input variable two functions describing membership to the fuzzy subsets Favourable (F) and Unfavourable (U) have been defined. These functions are based on criteria drawn from the literature or on the authors’ expert judgement. The expert system calculates the value of modules according to the degree of membership of the input variables to the fuzzy subsets F and U and according to sets of decision rules. The four modules can be considered individually or can be aggregated (again according to membership to fuzzy subsets F and U and a set of decision rules) into the indicator Ipest. The system is flexible and can be tuned to expert perception, it can be used as a decision aid tool to rank or choose between alternative pesticide application options with respect to potential environmental impact. Results of a sensitivity analysis and module and Ipest scores for some pesticide application cases are presented. An agro-ecological indicator IPEST, based on the expert system, is proposed as a tool to assess the environmental impact of all pesticide applications related to a crop within a year. The practical implementation of the expert system and its validation are discussed.

Key words: decision support, environmental impact, expert system, fuzzy logic, leaching, pesticides, runoff, toxicity, volatilisation

Pesticides are xenobiotic substances which are used in crop production for the control of pests, diseases and weeds. As crops generally are grown outdoors, the application of pesticides to field crops by definition implies emission to the environmental compartments air and soil. However, there are immense differences in the degree to which pesticides are mobile and biologically active in the environment. Consequently, the pesticide use options available to farmers differ strongly with respect to the risks they pose to the environment.

Commercial products for the control of pests, diseases or weeds generally contain active ingredients, adjuvants and inert ingredients. The term "pesticide" is often used as a synonym for "active ingredient" (e.g. Tomlin, 1994), but it is also used as a synonym for "commercial product for the control of pests, diseases or weeds" (e.g. Hayes, 1991). This is confusing, in this paper we use "pesticide" as a synonym for "active ingredient" only. Adjuvants used in the formulation of the commercial product can change its agronomic effects (effectiveness, phytotoxicity) as well as its environmental impact, as dispersion patterns may be altered and the functional activity period of the active ingredient may be lengthened or its degradation may be delayed (Levitan et al., 1995). Unfortunately, very little information on the effects and fate of adjuvants is available in the scientific literature and therefore the role of adjuvants will not be taken into account in this paper.

In general, the anticipated effectiveness against the pest, the risk of phytotoxicity to the crop and the cost of the application are the main factors considered by a farmer in choosing a particular pesticide. An increasing number of farmers and other decision-makers would like to be able to take the potential environmental impact of the application of a pesticide into account. Within the last two decades a number of methods have been proposed for estimating this impact (reviewed in Shahane and Inman 1987, Levitan et al. 1995, van der Werf 1996). These methods assess potential pesticide environmental impact, as perceived by the "experts" that created them.

Girardin et al. (1997a) propose a set of "agro-ecological indicators" to evaluate the environmental impact of arable farming systems. The term "indicator" has been defined as: a variable which supplies information on other variables which are difficult to access (Gras et al., 1989). Indicators are valuable tools for evaluation and decision making as they synthesise information and can thus help to understand a complex system (Mitchell et al. 1995). This paper describes an expert system which is at the core of the agro-ecological indicator reflecting pesticide environmental impact.

The environmental impact of the application of a pesticide in a field crop will depend on characteristics of: a) the pesticide (e.g. its toxicity to water organisms), b) the local environment (e.g. soil type), c) the application (e.g. application in the soil, on the soil or on the crop) (van der Werf 1996). If one accepts the premise that the "environmental impact" of a pesticide results from a combination of exposure and toxicity (Severn and Ballard 1990, Emans et al. 1992), obviously the use of a simulation model to estimate the exposure component is tempting (Reus and Pak 1993, Russel and Layton 1992). Using pesticide, environmental and application characteristics as input variables a simulation model can yield predicted environmental concentrations (PEC), which can be related to predicted no-effect concentrations or maximum admissible concentrations. This approach to estimating pesticide environmental impact is attractive: its results (at least regarding PEC) can be validated and it provides an elegant way to combine the three types of input variables which should be considered. However, there are disadvantages to the use of simulation models for impact estimation. In the first place, none of the currently available models describing pesticide environmental fate simultaneously represents the major relevant processes: behaviour in soil, volatilisation, drift, leaching, runoff and degradation (Wagenet and Rao 1990, EPPO 1993). Secondly, the validation status of pesticide fate models generally is low (Calvet 1995, FOCUS 1995, Gustafson 1995). Finally, these models generally require a large amount of data for their parameters and input variables (pesticide properties, soil characteristics, meteorological data). For most pesticide application situations a major part of these data are unavailable. When available, values often are imperfect: they may be imprecise (e.g. pesticide field half-life), or uncertain (e.g. risk of the occurrence of surface water runoff on a particular field).

Considering the inadequacy of the available simulation models and the shortcomings of current indexing approaches for the assessment of pesticide environmental impact, we decided to follow a different avenue by setting up a fuzzy expert system (Hall and Kandel 1991). This technique is robust when uncertain and imprecise data is used and allows the aggregation of dissimilar input variables in a consistent and reproducible way (Bouchon-Meunier 1993). In this paper we describe an expert system to calculate the agro-ecological indicator "Ipest", which reflects an expert perception of the potential environmental impact of the application of a pesticide in a field crop.

We consider that pesticide impact on the environment depends on: a) the presence of a certain amount of pesticide, b) the extent to which the pesticide leaves the field on which it is applied by drift, volatilisation, runoff or leaching, c) the toxicity of the pesticide. This implies that we do not take into account the soil, or the organisms in the soil or on the crop. We do consider these to be part of the natural environment, but there is insufficient data regarding pesticide impact on soil and terrestrial non-target organisms.

We defined four indicator modules. The module Presence reflects the rate of application of the pesticide, the modules Risk of surface water contamination (Rsur), Risk of groundwater contamination (Rgro) and Risk of air contamination (Rair) reflect the risk for three major environmental compartments. For each module a value on a dimensionless scale between 0 (no risk of environmental impact) and 1 (maximum risk of environmental impact) is calculated. The value of the module Presence depends on a single input variable (Rate of application). The value of the other three modules depends on four or five input variables (Table 1) and a set of decision rules. Values are calculated according to a procedure which will be explained in detail in the next section of this paper. We distinguish three types of input variables: a) pesticide properties, b) site-specific conditions, c) characteristics of the pesticide application (Table 1). In selecting these input variables we first considered their relevance for impact assessment and secondly data availability. Obviously there is not much point in taking into account input variables for which data are unavailable or hard to come by for many pesticide application situations.

The four indicator modules can be considered individually or can be aggregated into an overall indicator estimating the total potential environmental impact of a pesticide application, again on a 0 to 1 scale. This modular structure presents several advantages. In the first place users have access both to an indicator reflecting overall impact, and to each of the modules. Secondly, the mode of aggregation of modules can be changed and new modules (e.g. risk for soil, impact on beneficial arthropods) can be added, as availability of data and understanding of pesticide impact evolves, or depending on the demands of the users.

A fuzzy expert system was used to aggregate input variables into indicator modules, and to subsequently aggregate the modules. Fuzzy logic is a superset of conventional (Boolean) logic that has been extended to handle the concept of partial truth: truth values between "completely true" and "completely false". It was proposed by Zadeh (1965) to deal with the uncertainty of natural language. Fuzzy set theory can be used to cope with vaguely defined classes or categories.

In classical set theory, an element either is or is not in a set. For example, if subset A consists of the pesticides with a maximum field half-life of 20 days, a particular pesticide can be classified as a member or not a member of the subset. If, however, A is defined to be the subset of "non-persistent" pesticides, then it is more difficult to determine if a specific pesticide is in the subset. If one decides that only pesticides with a maximum field half-life of 20 days are in the subset, then a pesticide with a 21 day half-life cannot be classified as non-persistent even though it is "almost" non-persistent. The use of fuzzy set theory is particularly compelling because available values for field half-life and several other relevant variables are imprecise and/or uncertain. Thus classification based on the conventional approach, where the transition between classes is abrupt, is doubtful.

Fuzzy set theory addresses this type of problem by allowing one to define the "degree of membership" of an element in a set by means of a membership function. For classical or "crisp" sets, the membership function only takes two values: 0 (non-membership) and 1 (membership). In fuzzy sets the membership function can take any value from the interval [0,1]. The value 0 represents complete non-membership, the value 1 represents complete membership, and values in between are used to represent partial membership.

For all input variables given in Table 1 we defined two fuzzy subsets F (Favourable) and U (Unfavourable). We based the membership functions on data available in the literature or on our own "expert" knowledge. For example, Gustafson's (1989) Groundwater Ubiquity Score (GUS) has three classes for pesticide leachability (Fig. 1a). We gave pesticides classified by Gustafson as "leacher" (i.e. GUS > 2.8) a membership value of 1 for the fuzzy subset U and a membership value of 0 for the fuzzy subset F. Pesticides classified as "nonleacher" (GUS < 1.8) are given a membership value of 0 for the fuzzy subset U and a membership value of 1 for the fuzzy subset F. The class of "borderline compounds" (1.8 < GUS < 2.8) falls within a "transition interval" in which the membership value for F decreases from 1 (at GUS = 1.8) to 0 (at GUS = 2.8), and the membership value for U increases from 0 to 1 (thus the functions characterising F and U are complementary, Fig. 1b). According to this approach we can characterise the shape of the membership function of each input variable by the two limits of the "transition interval". We used membership functions that are sinus shaped in the transition interval, as they provide smoother variations of the values of the modules and the indicator than membership functions that are linear in the transition interval.

For each module we formulated a set of decision rules attributing values between 0 and 1 to an output variable according to the membership of its input variables to the fuzzy subsets F and U. Sugeno's inference method (Sugeno 1985) was used to compute the modules as well as the indicator. Its principle is briefly described in the following. Sugeno's inference method uses decision rules of the form :

If x₁ is A₁₁ and x₂ is A₁₂ Then y is B₁

If x₁ is A₂₁ and x₂ is A₂₂ Then y is B₂

where x_j (j = 1,2) is an input variable (e.g. field half-life, or volatilisation risk), y is an output variable (e.g. the value of the module), A_ij a fuzzy subset (Favourable or Unfavourable), and B_i a number called conclusion of the rule. "x_j is A_ij" is called a premise of the i-th rule.

Let x₁⁰ and x₂⁰ be the values taken by x₁ and x₂, and A_ij(x_j⁰) the membership value of x_j⁰ to the fuzzy set A_ij (given by the membership function that defines A_ij). Then, one can define w₁ and w₂, the truth values of the rules, as follows:

where min means "minimum value of". The first rule infers w₁B₁, the second one w₂B₂, and the global output y⁰ is inferred by:

y⁰ = (w₁B₁ + w₂B₂)/(w₁ + w₂)

To illustrate Sugeno's inference method an example of a calculation will be given. Let's assume (only to illustrate the approach and for simplicity's sake) that the output variable Environmental impact of the application of a pesticide depends on two input variables only: Rate of application and pesticide Field half-life. For both input variables membership to fuzzy subsets F (Favourable) and U (Unfavourable) has to be defined. Let's assume that experts tell us that a low Rate of application and a short Field half-life are favourable, whereas a high Rate of application and a long Field half-life are unfavourable. To define the shape of the membership functions we have to be specific: for Rate of application we assign complete membership to the fuzzy subset F if Rate of application < 0.001 kg ha^-1 and complete membership to the fuzzy subset U if Rate of application > 2 kg ha^-1; for Field half-life we assign complete membership to F if Field half-life < 1 day and complete membership to U if Field half-life > 120 days (Fig. 2). Consequently, rates between 0.001 and 2 kg ha^-1 and field half lives between 1 and 120 days fall within a "transition interval" as explained above.

In this example, which has two input variables and two fuzzy subsets for each input variable, four situations may occur, as reflected by the decision rules which are summarised in Table 2. These decision rules reflect expert knowledge and/or expert judgement; they read as (e.g. first line of the table): "If Rate of application is favourable and if Field half-life is favourable then Environmental impact is 0.". The decision rules consist of two premises (if....) linked by and, followed by a conclusion (then....). As can be seen from Table 2, when both input variables are F, the value of Environmental impact is 0.0 (no risk of environmental impact), when both input variables are U the value of Environmental impact is 1.0 (maximum risk of environmental impact) and when one input variable is F and the other is U the value of Environmental impact is 0.5.

Having thus defined the membership functions and formulated the decision rules, we proceed to calculate the value of the output variable Environmental impact for a pesticide application. Let's assume that the herbicide atrazine is applied at 1.5 kg ha^-1 of active ingredient and that its field half-life is 60 days. The membership functions defined above allow the calculation of the truth value of the premises, i. e. the degree of membership to the fuzzy subset concerned (Favourable or Unfavourable) for each input variable (Fig. 2). According to Sugeno's (1985) inference method, when the premises are linked by and, the truth value of a decision rule can be defined as the smallest of the truth values of its premises (Table 3). The value of Environmental impact is calculated as the average of the conclusions of the decision rules, weighted by their truth value:

Environmental impact = (0.0*0.147 + 0.5*0.147 + 0.5*0.506 + 1.0*0.494) = 0.634

This method fits our requirements, since it allows to build a modular structure for our indicator, providing usable values for the indicator modules. Moreover, the rules are easy to read, and the numerical scores used for their conclusions are easy to tune to be in accord with the opinion of the experts.

The indicator module Presence reflects quantity of active ingredient applied, it depends on a single input variable: Rate of application. A single application of a pesticide may range from a few g ha^-1 of active ingredient for low-dose herbicides to several hundreds of kg ha^-1 for a soil treatment against nematodes. A major part of this enormous range of application rates is due to differences in the biological activity of the active ingredients used, and thus a lower rate of application does not necessarily imply less risk for the environment. On the other hand, improved application techniques (better targeting) allow lower application rates and many farmers are successful in using application rates below those recommended on the label by matching timing of application with vulnerable periods of the targeted organism or by combining pesticides with other means of control. Many experts judge that, generally speaking, low rates of application are desirable from an environmental point of view (Kovach et al. 1992, Reus and Pak 1993, Jouany and Dabène 1994, Vereijken et al. 1995, Halfon et al. 1996, Shukla et al., 1996).

For the input variable Rate of application we did not find literature data to define the values delimiting the transition interval within which the fuzzy subsets F (Favourable) and U (Unfavourable) are complementary. As the transition interval should encompass a wide range of values and as obviously environmental impact will not increase linearly with application rate we defined the input variable Rate of application as the log₁₀ of the application rate in g ha^-1. We arbitrarily decided to assign complete membership to F if Rate of application < 1 (10 g ha^-1) and complete membership to U if Rate of application > 4 (10,000 g ha^-1).

The value of the module Presence depends on the input variable Rate of application according to only two decision rules: 1) If the Rate of application is favourable then Presence is 0.0 (minimal risk of environmental impact); 2) If the Rate of application is unfavourable then Presence is 1.0 (maximum risk of environmental impact).

The indicator module Rgro (Risk of ground water contamination) reflects the potential of a pesticide to reach groundwater through leaching and to affect its potential use as a source of drinking water for humans. The value of Rgro depends on four input variables: 1) pesticide leaching potential, 2) the position of application of the pesticide (on the crop, on the soil, in the soil), 3) soil leaching risk and 4) the toxicity of the pesticide to humans.

A number of authors have proposed indices based on pesticide properties to classify pesticides according to their groundwater contamination or leaching potential (Laskowski et al. 1982, Cohen et al. 1984, Jury et al. 1987, Gustafson 1989, Hornsby 1992). We use the groundwater ubiquity score (GUS) as proposed by Gustafson (1989) to quantify the leaching potential of a pesticide. This index is simple and can effectively discriminate between pesticides that leach and pesticides that do not. GUS is a function of the pesticide characteristics Field half-life (DT50) and Organic Carbon sorption constant (Koc):

GUS = log₁₀(DT50) x (4 - log₁₀(Koc)).

Pesticides detected in groundwater generally have GUS values exceeding 2.8, whereas compounds with GUS values below 1.8 were not detected in groundwater. We therefore define the limits of the transition interval within which the fuzzy subsets F (Favourable) and U (Unfavourable) are complementary by assigning complete membership to F if GUS < 1.8 and complete membership to U if GUS > 2.8.

The position of application of the pesticide strongly affects its leaching potential. Depending on crop cover, pesticides may be distributed between the crop and the soil surface. Although some of the pesticide present on the crop may be washed off and reach the soil (Leonard 1990), we consider application on the crop to be favourable. Apart from reducing the amount of pesticide on the soil, crop cover in itself reduces leaching risk, as transpiration by the crop reduces drainage. We therefore define the input variable Position of application_gro as a value on scale between 1 (100% of the pesticide on the crop) and 0 (0% on the crop, pesticide applied on or in the soil). Its value is calculated as: Position of application_gro = % soil covered by the crop/100. The % soil covered by the crop is to be estimated by the user of the system. We define the limits of the transition interval by assigning complete membership to F if Position of application_gro = 1 and complete membership to U if Position of application_gro = 0.

The risk of pesticide loss from a field to groundwater (Leaching risk) depends on characteristics of the soil, of the unsaturated zone above the water table and of the saturated aquifer. Several methods have been proposed to assess this risk as a function of soil characteristics only (e.g. Goss and Wauchope 1990), or of overall hydrogeologic settings including soil characteristics (e.g. Aller et al. 1985, Hollis 1991). Which approach is most appropriate will depend on local hydrogeologic settings and on the availability of data. We express Leaching risk on a scale between 0 (minor leaching risk) and 1 (major leaching risk). The value for Leaching risk can be obtained by transforming a score obtained from one of the methods cited above or from another appropriate method to a 0 to 1 scale. If this is not possible, a preliminary estimation of leaching risk can be based on soil organic matter content, which can be considered as the single most important soil characteristic affecting pesticide leaching (e.g. van der Zee and Boesten 1991). We define the transition interval by assigning complete membership to F if Leaching risk = 0 and complete membership to U if Leaching risk = 1.

As we consider the groundwater to be a potential source of human drinking water, we selected the Acceptable Daily Intake (ADI), which reflects chronic toxicity to humans, as our criterion for toxicity. ADI is defined as (Edelman 1991): "the daily dosage of a chemical which, during an entire lifetime, appears to be without appreciable risk on the basis of all the facts known at the time". ADI is expressed as mg of chemical residue in food, per kg of body weight (mg kg^-1day^-1). Jouany and Dabène (1994) classify pesticides into 5 classes according to their ADI: A) <0.0001 mg kg^-1d^-1, B) 0.0001 - 0.001 mg kg^-1d^-1, C) 0.001 - 0.01 mg kg^-1d^-1, D) 0.01 - 0.1 mg kg^-1d^-1, E) > 0.1 mg kg^-1d^-1. Drawing from their classification we use 0.01 mg kg^-1d^-1 as the median value for the transition interval. We define the input variable Human toxicity as the log₁₀ of the ADI. The limits of the transition interval are defined by assigning complete membership to F if Human toxicity > 0 (1 mg kg^-1d^-1) and complete membership to U if Human toxicity < -4 (0.0001 mg kg^-1d^-1).

The value of the module Rgro depends on the input variables GUS, Position of application, Leaching risk and Human toxicity according to a set of 16 decision rules summarised in Table 4. These decision rules reflect our "expert" perception of the risk of groundwater contamination as a result of a pesticide application. The expert reasoning is summarised in Fig. 3, which reads as follows: if GUS is F (non-shaded box), the value of Rgro is 0.0 (no risk). If GUS is U (shaded box) and Position of application is F, Rgro is 0.1. If GUS is U and Position of application is U, the value of Rgro depends on Leaching risk and, subsequently, on Human toxicity. If Leaching risk is F, the value of Rgro is 0.2 or 0.6 (depending on Human toxicity), if Leaching risk is U, the value of Rgro is 0.6 or 1.0.

The indicator module Rsur (Risk of surface water contamination) reflects the potential of a pesticide to reach surface water through runoff or drift and to harm aquatic organisms. Its value depends on five input variables: 1) the runoff risk of the field site, 2) the drift percentage of the application, 3) the position of application of the pesticide (on the crop, on the soil, in the soil), 4) the field half-life of the pesticide and 5) the toxicity of the pesticide to aquatic organisms. We did not take into account other pesticide characteristics such as solubility or sorption properties as these have been shown not to consistently affect the pesticide's runoff potential (Leonard 1990, Larson et al. 1995).

The risk of pesticide transport from a field to surface water by runoff (Runoff risk) depends on many factors, e.g. slope steepness, slope length, soil texture, surface condition, soil particle aggregation and stability, crop cover, distance to surface water (Leonard 1990, Simon 1995). Several methods have been proposed to assess this risk as a function of soil, slope and/or watershed characteristics (e.g. Papy and Boiffin 1988, Goss and Wauchope 1990, Hollis 1991, Aurousseau 1996). Which approach is most appropriate will, in the same manner as for Leaching risk, depend on the local situation and on availability of data. We express Runoff risk on a scale between 0 (no runoff risk) and 1 (major runoff risk). The value for Runoff risk can be obtained by transforming a score obtained from one of the methods cited above or from another appropriate method to a 0 to 1 scale. If this is not possible, Runoff risk can be estimated by the user. We define the transition interval by assigning complete membership to F if Runoff risk = 0 and complete membership to U if Runoff risk = 1.

The risk of pesticide transport from a field to surface water by drift depends on many factors, e.g. distance to surface water, application technique, crop structure and crop cover and wind speed (Emans et al. 1992, CLM-IKC 1994). We quantify the drift potential of a field site by its Drift percentage, defined as the pesticide rate reaching surface water expressed as a percentage of the intended rate of field application (CLM-IKC 1994). The value of Drift percentage depends on application technique and distance to surface water (CLM-IKC 1994, Table 5). We did not find literature data to define the values delimiting the transition interval for Drift percentage. We therefore arbitrarily decided to assign complete membership to F if Drift percentage = 0% and complete membership to U if Drift percentage > 1%.

The position of application of the pesticide strongly affects its runoff potential. Application on the soil surface is unfavourable as pesticide concentrations in surface runoff have been shown to be strongly correlated with concentrations in the surface 10 mm of soil (Leonard et al. 1979). Consequently, application in the soil (that is, either on the seed or incorporated in the soil) is favourable. Depending on crop cover, pesticides may be distributed between the crop and the soil surface. Although some of the pesticide present on the crop may be washed off and reach the soil (Leonard 1990), we consider application on the crop to be favourable. Apart from reducing the amount of pesticide on the soil surface, crop cover itself reduces runoff risk. We therefore defined the input variable Position of application_sur as a value on scale between 1 (100% of the pesticide on the soil) and 0 (0% on the soil, pesticide applied on the crop or in the soil). Its value is calculated as: Position of application_sur = (100 - % soil covered by the crop)/100, unless the pesticide is applied in the soil, in that case Position of application_sur = 0. The % soil covered by the crop is to be estimated by the user of the system. We define the limits of the transition interval by assigning complete membership to F if Position of application_sur = 0 and complete membership to U if Position of application_sur = 1.

Field half-life (DT50) affects a pesticide's runoff potential, persistant pesticides that remain at the soil surface for longer periods of time have a higher probability of runoff than non-persistent pesticides (Leonard 1990). With respect to the risk of surface water contamination Jouany and Dabène (1994) consider a pesticide having a DT50 < 8 days as "not unfavourable", one having a DT50 between 8 and 30 days as "moderately unfavourable" and one having a DT50 > 30 days as "unfavourable". Their proposition seems reasonable, as it has been shown that pesticide removal occurs mainly during the first or the first few runoff events after application (Leonard, 1990). Thus, drawing from Jouany and Dabène (1994), we defined the limits of the transition interval by assigning complete membership to F if Field half-life < 1 day and complete membership to U if Field half-life > 30 days.

The input variable Aquatic toxicity is based on biological effects on three aquatic species forming a food chain: algae (EC50), crustaceans (EC50) and fish (LC50), all expressed in mg l^-1. EC50 and LC50 being the concentration of the pesticide that would produce a specific effect (EC50) or death (LC50) in 50% of a large population of a test species. In each case the concentration for the most sensitive species tested was retained, thus relating the assessment of the risk to the ecosystem to the most sensitive organism (Canton et al. 1991). Linders et al. (1994) consider a pesticide which has an EC50 or LC50 < 1 mg l^-1 as highly toxic, one with an EC50 or LC50 between 1 and 10 mg l^-1 as moderately toxic, one with an EC50 or LC50 between 10 and 100 mg l^-1 as slightly toxic and one having an EC50 or LC50 above 100 mg l^-1 as very slightly toxic. Jouany and Dabène (1994) classify pesticides into 5 classes according to their EC50 or LC50 for aquatic organisms: a) <0.001 mg l^-1, b) 0.001 - 0.01 mg l^-1, c) 0.01 - 0.1 mg l^-1, d) 0.1 - 1 mg l^-1, e) > 1 mg l^-1. Drawing from these references we define the input variable Aquatic toxicity as the log₁₀ of the EC50 or LC50 of the most sensitive species of the three considered. We use 0 (log₁₀ of 1 mg l^-1) as the median value for the transition interval, and define the limits of the transition interval by assigning complete membership to F if Aquatic toxicity > 2 (100 mg kg^-1) and complete membership to U if Aquatic toxicity < -2 (0.01 mg kg^-1).

The value of the module Rsur depends on the input variables Runoff risk, Drift percentage, Position of application, Field half-life and Aquatic toxicity according to a set of 32 decision rules (not shown). These decision rules (summarised in Fig. 4) reflect our "expert" perception of the risk of surface water contamination as a result of a pesticide application.

The indicator module Rair (Risk of air contamination) reflects the potential of a pesticide to volatilise and to contaminate air. The value of Rair depends on four input variables: 1) pesticide volatility, 2) the position of application of the pesticide (on the crop, on the soil, in the soil), 3) the field half-life of the pesticide and 4) the toxicity of the pesticide to humans.

Henry's law constant (KH, dimensionless), the ratio of vapour pressure to water solubility, is considered a more appropriate indicator of the volatilisation rate of a pesticide than its vapour pressure alone (Jury et al. 1984, Spencer and Cliath 1990). Compounds with KH much greater than 2.65 x 10-5 are volatile; compounds with KH much smaller than 2.65 x 10-5 are much less volatile (Jury et al. 1984, Clendening et al. 1990). We use KH = 2.65 x 10-5 as the middle of the transition interval, and define the input variable Volatility as log₁₀ KH rather than as KH. We define the limits of the transition interval by assigning complete membership to F if Volatility < Log₁₀ 2.65 x 10-6 and complete membership to U if Volatility > Log₁₀ 2.65 x 10-4.

The position of application of the pesticide strongly affects its volatilisation potential. Application in the soil is favourable as volatilisation is greatly reduced by incorporation into the soil (Taylor and Spencer 1990). Consequently, application on the soil or on the crop is unfavourable. We therefore defined the input variable Position of application_air as a value on scale between 1 (100% of the pesticide in the soil) and 0 (0% of the pesticide in the soil, pesticide applied on the soil/crop). Its value is calculated as: Position of application_air = % of pesticide applied in the soil/100. The % of pesticide applied in the soil is to be estimated by the user of the system. We define the limits of the transition interval by assigning complete membership to F if Position of application_air = 1 and complete membership to U if Position of application_air = 0.

Field half-life (DT50) affects a pesticide's volatilisation potential, persistent pesticides that remain in the soil will volatilise over longer periods of time than non-persistent pesticides (e.g. Spencer et al. 1996). The limits of the transition interval for Field half-life have been defined in section 4.3.

We selected the Acceptable Daily Intake (ADI), which reflects chronic toxicity to humans, as our criterion for toxicity. The definition and the limits of the transition interval for Human toxicity were given in section 4.2.

The value of the module Rair depends on the input variables Volatility, Position of application, Field half-life and Human toxicity according to a set of 16 decision rules (not shown). These decision rules (summarised in Fig. 5) reflect our "expert" perception of the risk of air water contamination as a result of a pesticide application.

The modules described above can be used to compare several pesticide application options with respect to their potential environmental impact. The available options might be ranked for instance by means of a multicriteria analysis technique, using the modules as evaluation criteria (Girardin et al., 1997a). An alternative approach would be to aggregate the four modules in some way into an overall indicator (Ipest). This can be done by summation, multiplication or a combination of both. For example: Ipest = Presence (Rsur + Rgro + Rair), or Ipest = Presence max(Rsur, Rgro, Rair).

We present here another mode of aggregation of the four modules using decision rules, similar to that used for the aggregation of input variables into indicator modules, as described above. In setting up the set of decision rules we had to decide on the relative importance of the environmental compartments surface water, groundwater and air. In most indicators of environmental impact surface water and groundwater are given much emphasis, whereas air is rarely taken into consideration (van der Werf 1996). It is hard to decide on the relative importance of air pollution by pesticides, because whereas data concerning pesticide concentrations in groundwater and surface water are readily available, those for air are scarce. However, calculations by Wortham (pers. comm. 1996) based on data by Millet (1994) show that pesticide concentrations in air in Alsace (NE France) are such (0.068 µg m^-3) that pesticide uptake by humans from air is six times larger than uptake from drinking water (assuming a 0.1 µg l^-1 concentration in drinking water). Considering this evidence we decided to give the same weight to the air compartment as to the surface water and the groundwater compartments in the aggregation of the modules.

The value of the indicator Ipest depends on the modules Presence, Rsur, Rgro and Rair (used as input variables) according to a set of 16 decision rules (not shown). The definition of the limits of the transition interval is the same for the four input variables: we assign complete membership to F if the value of the module is 0 and complete membership to U if the value of the module is 1. The expert reasoning is summarised in Fig. 6: if all modules are F, the value of Ipest is 0.0 (no risk). If Presence is F and one or more of the other three modules is U, the value of Ipest is 0.3, (one module U, two F) 0.4 (two modules U, one F) or 0.5 (three modules U). If Presence is U and the other three modules are F, the value of Ipest is 0.1; if one or more of the other three modules is U, the value of Ipest is 0.8, (one module U, two F) 0.9 (two modules U, one F)or 1.0 (three modules U). These rules obviously reflect value judgements, and can be modified to suit the perception of a user of the system.

In order to illustrate the functioning of the system, we present an analysis of the sensitivity of the indicator Ipest to variation in the values of its input variables. Each input variable was varied over its transition interval, while the other input variables were kept fixed at the median value of their transition interval, or at the extremes of the transition interval, i.e. at "Favourable" or at "Unfavourable" (Fig. 7). We assumed the pesticide was applied by field spraying on the crop and/or the soil. Consequently, the value of the input variable Position of application depended only on soil cover by the crop, which was varied between 0% and 100%. The sensitivity analysis reflects the functioning of the system and supplies in particular some indication of the relative weight the input variables may have in the value of Ipest. However, one should be aware that the effect of the variation of an input variable over its transition interval on the value of Ipest depends very much on the value of the other input variables. Therefore, results presented here should be considered as illustrations of the functioning of the system.

The most influent input variable is Rate of application (Fig. 7d), its effect is very large when all other input variables are unfavourable and much smaller when the other input variables are favourable. This obviously is the result of the mode of aggregation (Fig. 6) we adopted, giving a lot of "weight" to the module Presence (which depends Rate of application only) when the other three modules are unfavourable. The extent to which other input variables affect the value of Ipest in this sensitivity analysis can similarly be deduced from the decision rules involved. For instance: the influence of Field half-life is large (Fig. 7a, 7b), because it is an input variable to three modules (Rsur, Rgro and Rair). The effect of Runoff risk is smaller than that of Drift percentage (Fig. 7b, 7c), because the effect of Runoff risk is modulated by Position of application, Field half-life and Aquatic toxicity, whereas Drift percentage is modulated by Aquatic toxicity only (Fig. 4).

Another illustration of the functioning of the system is given in Table 6, which shows the value of the four modules and of Ipest for some pesticide application cases. Results are shown for different pesticides, applied at their recommended rate of application in a field with Leaching risk = 0.5, Drift percentage = 0 and with either Runoff risk = 0 (no runoff risk) or Runoff risk = 1 (major runoff risk). This type of output can be used to compare or rank different pesticide applications with respect to the value of one or several of the modules or of the Ipest indicator. The results obtained here again evidently are the consequence of the choices we made regarding the selection of input variables, the definition of their transition intervals and the values given to the conclusions of the decision rules.

The expert system described above assesses the environmental impact of a single application of a pesticide. The system can be used as a decision support system to rank or choose between alternative pesticide application options. In this section we will briefly discuss some major issues concerning the practical implementation of the expert system. These issues are: the construction of a so-called "agro-ecological indicator", mixtures of pesticides, pesticide metabolites and variability of the values of input variables.

This work was carried out within an ongoing project concerning "agro-ecological indicators for the evaluation of farming systems" as presented by Bockstaller et al. (1997). Agro-ecological indicators assess the effects of cultural practices and farming system (e.g. nitrogen management, crop diversity) on the environment and are destined to be used by farmers, farmer-advisors or other decision makers concerned with the environmental impact of agriculture. Agro-ecological indicators take a value between 0 and 10, the value 7 representing the achievement of the minimum requirements of Integrated Arable Farming Systems (IAFS). Values below 7 indicate that the cultural practices do not satisfy minimum IAFS requirements, whereas an indicator value above 7 indicates that the farmer does better than these minimum requirements. The time scale of calculation is generally about one year: the period between the harvest of the preceding crop and the harvest of the crop in the current year. The value of Ipest as described in section 5 concerns a single pesticide application. The agro-ecological indicator IPEST is calculated for all applications related to a crop within a year as:

IPEST = 10 - k S Ipest_i

where k is a constant depending on the crop and the region and Ipest_i is a value of Ipest (between 0 and 1) for a single application of a pesticide. The value of k is chosen such that a value of 7 for IPEST is obtained for a crop protection programme satisfying the minimum requirements for IAFS according to a group of experts.

Often a commercial product contains two or more pesticides. For the calculation of Ipest such an association can either be considered as several separate applications or as a single application. In the latter case Ipest can be calculated as the weighted mean of the values obtained for each of the component pesticides at the rate of application of the mixture. For example: 2.88 kg ha^-1 of a mixture consisting of 70% of alachlore and 30% of atrazine is applied on the soil in a field with Runoff risk =0, Leaching risk = 0.5 and Drift percentage = 0. Under these circumstances the application of 2.88 kg ha^-1 of alachlore would result in a value for Ipest of 0.17, and the application of 2.88 kg ha^-1 of atrazine would result in a value for Ipest of 0.68. The value of Ipest for the mixture is calculated as: Ipest = 0.7*0.17 + 0.3*0.68 = 0.323.

As pointed out in the introduction of this paper, a major problem in the assessment of pesticide environmental impact is the uncertainty and/or imprecision of some of the major input variables involved (e.g. DT50, Koc). Measured values for these variables show a huge variability related to soil, climatic or other environmental conditions. The use of some consensus "average" value for this type of data is a pragmatic but far from satisfactory solution. Provided the variability of input variables can be adequately characterised (which is a problem in itself) the approach involving fuzzy sets described in this paper is very well suited to take into account uncertainty of input data (Bouchon-Meunier 1993).

In the design of a system to assess pesticide environmental impact two major questions have to be answered: a) which input variables should be taken into account?, b) how should the input variables be aggregated? The method presented in this paper proposes an answer to both questions, however, its originality lies in the answer it provides to the second question. Compared to other methods to assess pesticide environmental impact (reviewed in Shahane and Inman 1987, Levitan et al. 1995, van der Werf 1996) our approach contains two new elements: a) the use of fuzzy sets, b) the use of decision rules. The use of fuzzy sets provides an elegant solution to the problem of deciding on the cutoff values for input variables: e.g. the limit between "non-persistent" and "moderately persistent" pesticides. The use of decision rules allows an "intelligent" aggregation of input variables: e.g. Position of application, Leaching risk and Human toxicity affect Rgro only when GUS is unfavourable (Fig. 3). The combinations of these two concepts in sets of fuzzy rules (see Sections 3 and 4) is attractive, because although the combinations of values of input variables (e.g. Rate of application and Field half-life, Section 3) are infinite, a single set of fuzzy rules connects them all.

The system to assess pesticide environmental impact proposed here requires validation, which means that we should test whether its original objective has been achieved. The agro-ecological indicator IPEST is based on an expert system. The objective of an agro-ecological indicator is to render reality intelligible, and thus its validation consists of determining its value to potential users (Girardin et al. 1997b). The objective of an expert system is the simulation of a human expert, thus the expert system is validated if it displays under a variety of conditions the responses to input that the human expert would display (Plant and Stone 1991).

IPEST and the other agro-ecological indicators are currently being tested on 17 commercial arable farms of the Rhine valley in France and Germany (Bockstaller et al. 1997) and on the farms of three secondary schools for agricultural training in France. These users will be surveyed to obtain their evaluation and proposed amendments. The improvement and validation of IPEST and the other indicators is thus carried out in interaction with the users.

The validation of the expert system similarly is ongoing. A draft version of this paper with a description of a first prototype of the system was sent to over 20 experts in the field of pesticide environmental impact and its assessment. Most of these experts supplied comments (see Acknowledgements). Several modifications and additions were made, leading to the version presented here. The system is currently used to compare the environmental impact of single pesticide applications and of sequences of pesticide applications in different crops and cropping systems. The results thus obtained will be presented along with those obtained using other pesticide impact assessment systems such as the Environmental Impact Quotient (Kovach et al. 1992) and the CLM Environmental Yardstick for Pesticides (Reus and Pak 1993). Experts will be invited to comment on the results. If there is disagreement between expert perception of pesticide environmental impact and the output of the expert system, we will examine the cause of this divergence, which may be one or several of the following: choice of input variables, choice of the limits of the transition interval, formulation of the decision rules, values given to the conclusions of the decision rules, or the mode of aggregation of the modules. All of these may be modified according to expert consensus.

We propose a fuzzy expert system reflecting an expert perception of potential pesticide environmental impact. The system takes into account three types of input variables: pesticide properties, site-specific conditions and pesticide application factors. It can be used as a decision support to rank or choose between alternative pesticide application options. The system has a modular structure and thus provides both a synthetic indicator reflecting overall impact as well as more detailed information through its four modules. New modules can be added if necessary. The system is flexible and can be tuned to expert perception.

We used the database Microsoft Access for Windows 95, version 7.0 to implement the Ipest expert system and a data set of relevant pesticide properties. The application may be supplied on request on two conditions: 1) in any use for commercial or paid consultancy purposes a suitable royalty agreement must be negotiated with INRA-Colmar. 2) In any publication arising from use for research purposes the source of the program should be properly acknowledged and a copy of the publication sent to Philippe Girardin at INRA-Colmar.

This research was funded by a research training fellowship (Contract number AIR3-BM-0022) of the Commission of the European Communities Research and Technological Development programme of DG XII. The graphical representation of the decision rules is due to Olivier Roussel. A draft of this paper was submitted to a number of experts in the field of pesticide environmental impact and its assessment. The authors would like to thank G. Assouline, E. Barriuso, R. Belamie, C. Bockstaller, A. Cavelier, J. Delphin, P. Girardin, D. Gustafson, C. Kempenaar, J. Kroes, L. Levitan, R. Martin-Clouaire, R. Neumann, B. Réal, J. Reus, C. Walter, F. Wijnands and H. Wortham for providing valuable comments on the manuscript. We thank F. Bouneb (INRA-Phytopharmacie) and R.L. Glenn (USDA) for supplying pesticide data.

Table 1. The indicator modules Presence, Risk of surface water contamination (Rsur), Risk of groundwater contamination (Rgro) and Risk of air contamination (Rair), and their input variables. For details, see text.

Table 2. Summary of decision rules describing the effect of the input variables Rate and Field half-life on the hypothetical module Environmental impact. F = favourable, U = unfavourable. For details, see text.

Table 3. Summary of decision rules describing the effect of the input variables Rate and Field half-life on the hypothetical module Environmental impact. Truth values of premises and conclusions for an application of 1.5 kg ha^-1 of atrazine (Field half-life 60 days) are shown in brackets. F = favourable, U = unfavourable. For details, see text.

Rate	Field half-life	Environmental impact
F (0.147)	F (0.506)	0.0 (0.147)
F (0.147)	U (0.494)	0.5 (0.147)
U (0.853)	F (0.506)	0.5 (0.506)
U (0.853)	U (0.494)	1.0 (0.494)

Table 4. Summary of decision rules describing the effect of the input variables GUS, Position of application, Leaching potential and Human toxicity on the indicator module Rgro (Risk of groundwater contamination). F = favourable, U = unfavourable. For details, see text.

Table 5. Estimation of drift percentage according to application technique and distance to surface water (adapted from CLM-IKC 1994).

Application method	% drift at distance to surface water : 0 - 2 m 2-10 m
Row spraying, knapsack spraying	0.5	0.2
Full field spraying, crop height £ 25 cm	1.0	0.3
Full field spraying, crop height > 25 cm	2.0	0.5

Table 6. The values of the modules Presence (Pres.), Risk of surface water contamination (Rsur), Risk of groundwater contamination (Rgro) and Risk of air contamination (Rair), and of the Indicator of Pesticide Environmental Impact (Ipest) for a number of pesticides applied at their recommended rate in a field with either Runoff risk = 0 (no runoff risk) or Runoff risk = 1 (major runoff risk), Leaching risk = 0.5 and Drift percentage = 0. Rate of application and position of application are given, pesticide properties used for the calculations are in Table 7.

1 on c/s : applied on the crop and/or on the soil, the percentage indicates the fraction of soil covered by the crop at the time of application; in soil: applied in the soil.

 

 

Table 7. Pesticide properties used for the calculation of module values and Ipest values in Table 6. Data are taken from the INRA database AGRITOX (F. Bouneb, pers. com. 1996), except for K_H, which was taken from Linders et al. (1994). Data taken from other sources are indicated by index letters, C: data from the Comité de liaison (1996), E: estimate by the authors, P: Pesticide Manual (Tomlin 1994), R: RIVM data (Linders et al. 1994), S: the SCS/ARS/CES pesticide properties database (R.L. Glenn, pers. com. 1996).

 

Pesticide name	DT50 (days)	Koc (cm3g-1)	GUS	KH	ADI (mg kg-1d-1)	Aquatox (mg l-1)
2,4-D	10	20	2.70	1.4 10-9	0.3000	0.9
alachlore	15	170	2.08	9.5 10-7	0.0005	0.05
atrazine	60	100	3.56	1.2 10-7	0.0005	4.3
carbofuran	50	22	4.52	9.6 10-7	0.0100	0.015P
cyfluthrin	30S	100000	-1.47	1.8 10-4	0.0200P	0.00014R
EPTC	6S	200	1.32	9.6 10-4	0.3160C	1.4R
glyphosate	47S	11275R	-0.09	<2.5 10-6E	0.0500	15R
isoproturon	17P	155	2.23	4.7 10-9	0.0060	0.004R
lindane	400	1100	2.49	6.1 10-5	0.0080	0.034
parathion	49	3010R	0.88	2.5 10-5	0.0050P	0.0018R
pendimethalin	90	5000	0.59	1.5 10-3	0.05	0.138
rimsulfuron	3P	60R	1.06	2.5 10-5	0.0160	1.6R

 

 

 

Figure 1. Graphical presentation of crisp (a) and fuzzy (b) sets.

 

Figure 2. Membership to the fuzzy sets Favourable and Unfavourable for atrazine (field half-life 60 days) applied at 1.5 kg ha^-1 as a function of Rate of application (top graph) and Field half-life (bottom graph).

 

 

 

Figure 3. The effect of the input variables GUS, Position of application, Leaching risk and Human toxicity on the value of the conclusions of the decision rules for the indicator module Rgro (Risk of groundwater contamination) according to their membership to the fuzzy sets Favourable (non-shaded boxes) and Unfavourable (shaded boxes). For details see text.

 

Figure 4. Summary of decision rules. The effect of the input variables Runoff risk, Drift percentage, Position of application, Field half-life and Aquatic toxicity on the value of the conclusions of the decision rules for the indicator module Rsur (Risk of surface water contamination) according to their membership to the fuzzy sets Favourable (non-shaded boxes) and Unfavourable (shaded boxes). For details see text.

 

Figure 5. The effect of the input variables Volatility, Position of application, Field half-life and Human toxicity on the value of the conclusions of the decision rules for the indicator module Rair (Risk of air contamination) according to their membership to the fuzzy sets Favourable (non-shaded boxes) and Unfavourable (shaded boxes). For details see text.

 

 

 

 

Figure 6. The effect of the modules Presence, Rsur, Rgro and Rair on the value of the conclusions of the decision rules for Ipest (indicator of pesticide environmental impact) according to their membership to the fuzzy sets Favourable (non-shaded boxes) and Unfavourable (shaded boxes). For details see text.

 

 

 

Figure 7a. Analysis of sensitivity of Ipest to variation of all input variables except Rate. Each input variable is varied over its transition interval from 0.0 (completely favourable) to 1.0 (completely unfavourable) while the other input variables are kept at unfavourable. Some curves coincide: GUS and Volatility, Aquatic toxicity and Leaching risk, Runoff risk and Drift percentage.

 

Figure 7b. Analysis of sensitivity of Ipest to variation of all input variables except Rate. Each input variable is varied over its transition interval from 0.0 (completely favourable) to 1.0 (completely unfavourable) while the other input variables are kept at the median value of their transition interval.

 

 

Figure 7c. Analysis of sensitivity of Ipest to variation of all input variables except Rate. Each input variable is varied over its transition interval from 0.0 (completely favourable) to 1.0 (completely unfavourable) while the other input variables are kept at favourable. Most curves coincide: Human toxicity, Aquatic toxicity, Runoff risk, Leaching risk and Crop cover; DT50, GUS and Volatility.

 

 

Figure 7d. Analysis of sensitivity of Ipest to variation of the input variable Rate. Rate is varied over its transition interval from 0.0 (completely favourable) to 1.0 (completely unfavourable) while the other input variables are kept at favourable, the median value of their transition interval or unfavourable.

 

---