Ecological Modelling 130 (2000) 111 119 www.elsevier.

com/locate/ecolmodel

Fuzzy indices of environmental conditions

William Silvert *
Department of Fisheries and Oceans, Habitat Ecology Section, Bedford Institute of Oceanography, P.O. Box 1006, Dartmouth, NS, Canada B2Y 4A2

Abstract Fuzzy logic provides a powerful and convenient formalism for classifying environmental conditions and for describing both natural and anthropogenic changes. Whereas traditional indices are based either on crisp sets with discontinuous boundaries between them (e.g. pristine vs. polluted), or on continuous variables whose values are only meaningful to experts (such as so many ppm of a toxin), fuzzy sets make it possible to combine these approaches. Conceptually the use of fuzzy logic is simple (for example, one can describe a site as 20% pristine and 80% polluted), but the real power of the methodology comes from the ability to integrate different kinds of observations in a way that permits a good balance between favourable and unfavourable observations, and between incommensurable effects such as social, economic, and biological impacts. In addition, fuzzy logic can be used to classify and quantify environmental effects of a subjective nature, such as bad odours, and it even provides a formalism for dealing with missing data. The fuzzy memberships can be used as environmental indices, but it is also possibly to defuzzify them and obtain a more traditional type of index. The fuzzy methodology is illustrated by examples based on research to evaluate of the effects of nsh mariculture on coastal zone water quality. 2000 Elsevier Science B.V. All rights reserved.
Keywords: Fuzzy logic; Fuzzy sets; Classication; Environmental effects

1. Fuzzy memberships as environmental indices Environmental indices rarely have much signicance by themselves. Knowing that some variable like pollutant concentration or soil acidity has a specic value is usually meaningful only in the context of knowledge of natural background levels, regulatory policy, and the vulnerability of key
* Present address: Instituto de Investigac a o das Pescas e do Mar, Avenida de Bras lia, 1449-006 Lisbon, Portugal. Tel.: + 351-21-3027194; fax: + 351-21-3015948. E -mail address: wsilvert@ipimar.pt (W. Silvert).

environmental components. For this reason it is useful, and probably more practical, to relate these indices to some sort of acceptability measure, which can be interpreted as the membership in a fuzzy set of acceptable environmental conditions. The concept of acceptability is itself seen by some as fuzzy, in the colloquial rather than mathematical sense, but this reects the reality that we can measure environmental effects far more accurately than we can evaluate their signicance. It also reects a lack of consensus about what different levels mean, and a basic problem within our

0304-3800/00/$ - see front matter 2000 Elsevier Science B.V. All rights reserved. PII: S 0 3 0 4 - 3 8 0 0 ( 0 0 ) 0 0 2 0 4 - 0

112

W. Sil6ert / Ecological Modelling 130 (2000) 111119

social structures in dening environmental values and objectives. It may therefore be seen as more honest, although less objective, to transform quantitative measurements of environmental variables into a fuzzy membership, which represents the degree of acceptability of those variable values in the set of acceptable conditions. Because different segments of society often cannot agree on what is acceptable and what is not, it may be desirable to identify several sets of acceptable conditions, representing the view of these different social groupings. This is discussed in Section 5.

2. Mathematical formalism Traditionally the symbol v has been used to represent fuzzy memberships. If x represents the value of an environmental variable, then v (x ) is the corresponding membership in the set of acceptable conditions, and takes a value between zero and one. For example, if a lake becomes hypoxic and all the sh die, then v would presumably be zero, indicating that this situation is totally unacceptable. In most situations more than one environmental variable is important, and we can dene vi(xi ), which is called the partial membership, to represent the acceptability of the i th environmental variable. One then needs to develop ways of combining these partial memberships to obtain a general measure of acceptability. The way in which partial memberships are combined depends on application. In some cases the appropriate combination is a standard fuzzy relationship; for example, if our objective is to have a healthy sh population in a lake, and the environmental variables are oxygen level, water temperature, and nutrients, then since adverse levels of any of these can be lethal to sh, the fuzzy intersection, v = min(v1, v2, v3), is appropriate. This denition states that the combined acceptability is as low as the lowest of the three partial acceptabilities. More commonly, however, there is some compensatory effect and low acceptability of one variable value can be compensated by good values of another. For this reason the choice of combination rule has to

be made in the context of a specic application, as described in Section 3.1. The relationship between an environmental variable x and its acceptability v (x ) can be quite complicated. Often values that are too low or too high are unacceptable, and the acceptability is high only for an intermediate range. More important, and more signicant in terms of the practical use of fuzzy logic in appraising environmental conditions, one can estimate the acceptability of environmental conditions that are themselves difcult or even impossible to quantify. Odours are a good example of this there are no commonly accepted techniques for measuring how bad something smells, but it is possible to determine that people nd certain smells unacceptable to varying degrees, and so we can assign at least an approximate value to v without having a quantitative measure of x. This, in fact, is one of the great strengths of fuzzy logic it lets us deal with subjective and nonquantitative data. Actually, some scientists see this as a great weakness in the use of fuzzy logic, since they feel that only hard data should be used in projects where scientists are involved. However, one has to be realistic; environmental effects are of concern not just to scientists, but also to businessmen, politicians, and ordinary citizens, and once one has ventured into the social and political realm where the real decisions are made, it simply is not possible to tell people that their concerns about foul smells and noise-induced headaches will be left out of the equation because they cannot be adequately measured!

3. Examples Several applications of this approach have been implemented, or are under consideration. Two are described below, one of which is very specic and has been worked out in detail, the other of which is more general and addresses issues which have not yet been fully resolved.

3.1. Benthic impacts under sh farms

The condition of the seabed under sh farms can be surveyed by divers, and various types of

W. Sil6ert / Ecological Modelling 130 (2000) 111119

113

observations can be made these include visual records of the abundance and diversity of fauna and ora, as well as geochemical cores obtained with hand-held samplers.

Fig. 1. Partial membership scores for the four fuzzy classications, based on observation of eight different benthic variables from 100 dive logs under sh pens. Taken from Angel et al. (1998).

Fig. 2. Weighted (defuzzied) impact score over the period of the study, dened as SCORE = 0vNil + 1vMod + 2vSev + 4vXtr. Vertical lines indicate signicant environmental periods during the study period indicated. Taken from Angel et al. (1998).

Although in principle it is possible to deal only with scientic quantitative data, this is not always practical, especially when a large number of farms are to be monitored. It is quite fast and easy for a diver to record that there is, for example, very little seaweed, a few crabs, and thick patchy bacterial mats, but to actually quantify this in terms of the biomass of seaweed, number of crabs per square meter, and thickness and percentage cover of mats requires special equipment and additional dive time. Other useful data, such as a strong sulphide smell to the cores, are virtually impossible to quantify. We therefore carried out a pilot project to explore the use of fuzzy logic for developing indices of benthic condition under a sh farm in Eilat, Israel, that has been studied by divers for several years (Angel et al., 1998). Four fuzzy sets were dened, representing nil, moderate, severe and extreme impacts. By using fuzzy sets we immediately resolved one of the major conceptual problems that had made the data difcult to interpret, which is that different types of observations can lead to incompatible results. For example, there may be a healthy benthic community which indicates that the impact is nil or at worst moderate, but extensive patches of thick bacterial mats may point to a severe or even extreme impact. Fuzzy sets allow for non-zero memberships in more than one of these categories, which offers a mechanism for resolving these apparent inconsistencies. Fig. 1 shows the partial memberships in the four fuzzy sets over time. There is a lot of useful information in this graph, but it is difcult to interpret, and for this reason the four memberships have been combined to produce a single comprehensive score which represents an overall measure of environmental quality, shown in Fig. 2. This process of combining several partial memberships to produce a single numerical value is called defuzzication and is commonly used in applications of fuzzy logic to produce a result which is compatible with non-fuzzy approaches and can be easily understood without reference to fuzzy set theory. The partial memberships were obtained by a procedure of rst assigning membership values to

114

W. Sil6ert / Ecological Modelling 130 (2000) 111119

each observation, and then combining them. For example, the presence of seagrass is generally considered to be a strong indicator of a healthy seabed (the species dominant in this area, Halophila stipulacea, is extremely sensitive to pollution), so if the divers identied the seagrass cover as normal, the partial memberships for nil, moderate, severe and extreme impacts were assigned as vNIL=0.8, vMOD=0.2, and vSEV= vXTR=0. These assignments, known as association rules, were applied to eight types of observations, namely the extent, thickness and colour of bacterial mats (Beggiatoa sp.), quantity of seagrass and epi-macrofauna, degree of bioturbation by macrofauna and by sh, and visibility. The observations were mainly qualitative, of the sort that could be made by divers without use of measuring instruments; for example, seagrass coverage was absent, few, or normal, and the thickness of the bacterial mats was thin, thick, or massive. Each variable was also assigned a weight reecting its importance in the benthic assessment process. The presence of seagrass and the extent of bacterial mats were considered strong indicators of benthic condition, so these were assigned high weights, while the colour of the mats and the presence of visible epi-macrofauna (which are not always easy to see) were assigned relatively low weights. As pointed out above, these different types of observations did not generally lead to consistent results, so a single set of observations usually had non-zero membership in all four categories. The important point to recognise about the assignment of partial memberships was that the entire procedure was developed to t the observational data, rather than requiring data that were considered excessively costly or impractical to collect. In addition to facilitating a simple and practical sampling scheme (with the additional benet that a high sampling frequency could be maintained), this also gave us access to an historical database that we might have had to discard if we were only prepared to accept rigorously quantitative measurements. There are however some important trade-offs, since simplifying the experimental aspect of the program is accomplished at the expense of some

fairly sophisticated mathematical analysis. Given a large number of benthic observations, we needed to combine the partial memberships to obtain overall membership values for the four fuzzy sets of nil, moderate, severe and extreme impact. For this we chose to use the method of symmetric summation (Silvert, 1979, 1997). Aside from certain mathematical advantages to using this rather than more traditional fuzzy logic relationships such as intersection (dened as the minimum of several partial memberships), there is a political reason for using symmetric summation, which is that it makes it easier to avoid value judgements. This may seem surprising in the context of scientic analysis, but it has relevance to the evaluation of environmental impacts. For example, consider two possible impacts of sh farming, nutrient enrichment and removal of certain benthic species. From the conservationist (environmental) point of view, both of these would normally be seen as negative effects, but from the point of view of a sherman they could be considered desirable, since nutrication could lead to increased productivity, and if the benthic species have no commercial value it might seem good to get rid of them (like weeds). One therefore needs to describe effects in such a way that there is no implicit assumption that certain levels are good and others are bad. Symmetric summation is a type of operator that addresses this by generating the same results whether one uses sets or their complements. In mathematical terms, if represents the symmetric summation operator and A is the complement of A (i.e. the membership in A is 1 vA), then A B= (A B)

In linguistic terms, symmetric summation offers a compromise between the logical operators AND and OR. Instead of saying that the combined effects are acceptable if A is acceptable AND B is acceptable, or that they are unacceptable if A is unacceptable OR B is unacceptable, one has a more neutral dividing line between what is and what is not acceptable. Another advantage of using symmetric summation rather than the intersection operator or other more traditional fuzzy relations is that it permits

W. Sil6ert / Ecological Modelling 130 (2000) 111119

115

weighting different effects according to their importance and signicance, as previously described. The general equation for the weighted symmetric sum of an arbitrary number of fuzzy memberships is v /(1 v ) = {[v1/(1 v1)]A[v2/(1 v2)]B [v3/(1 v3)]C}1/(A + B + C + ) (1) where v is the combined membership, the vi are partial memberships for different observations, and the exponents A, B, C, are the weighting factors. The meaning of this complex looking expression for the symmetric sum is simply that the value of v /(1 v ) is the weighted geometric mean of the ratios for partial memberships, vi/ (1 vi ), and it is clear that one gets the same result by replacing v by 1 v and vi by 1 vi all the way through. One of the advantages of this formalism is that it offers a way of dealing with missing data. Suppose for simplicity that one has just three observations, so that the memberships are given by v /(1 v ) = {[v1/(1 v1)]A[v2/(1 v2)]B [v3/(1 v3)]C}1/(A + B + C) and suppose that for one set of data the second observation is missing (this frequently happens, as a result of equipment failure, bad observing conditions, or many other situations beyond the control of the scientists organising the monitoring program). One can then use v /(1 v ) = {[v1/(1 v1)]A[v3/(1 v3)]C}1/(A + C) to calculate the memberships for this set of data, and although the gures are not as reliable as they would be if one had complete data, this is better than having to discard the complete set of observations because they are not complete. Of course, the real test of these ideas is whether they lead to meaningful results. Figs. 1 and 2 both show a great deal of structure, some of which is probably just random noise, but some patterns are evident. In Fig. 2 we have divided the time axis into six periods, and the external factors during these periods are as follows: period A during summer/autumn 1991 there were no outstanding environmental

events and impact scores were generally high. Membership in the nil category was low. period B in winter/spring 1992, there were strong storms and heavy sh bioturbation, which presumably removed some of the detritus from under the cages and led to increased membership in the nil impact category (some indicators showed little evidence of pollution). A sharp decline in the levels of severe and extreme impact was also observed, conrming that some of the worst effects were moderated by these events. period C during summer/autumn 1992, impact scores were generally high again, as in period A. period D during winter 1992/1993, sh were gradually removed from the cages and the underlying sediments were colonised by massive numbers of bioturbating decapods. The decline in extreme scores during this period shows that the decapods alleviated the worst impacts of the farm. period E during spring/summer/autumn 1993, all cages were removed, the sh farm was not active, and impact scores were low. period F in winter 1993/1994, sh cages were reinstated and stocked and impact scores increased rapidly, soon reaching levels as high as had been seen previously.

The most interesting results are those seen during periods E and F. During period E, when the cages were removed, there was dramatic improvement the severe and extreme scores fell immediately to a very low level, and the nil and moderate impact levels shot up. One can also see how the immediate effect of the cage removal is to increase the level of moderate impact, which then turns into a nil impact. However, the fuzzy scores reveal something that had not been detected by other means, namely a sudden worsening of conditions in late summer of 1993. The nil conditions that had been improving over the summer fell to a low level, and the major change was an increase in membership in the moderate impact category, accompanied by a small jump in severe and extreme impacts. The cause of this deterioration is not known, but it is remarkable that the use of

116

W. Sil6ert / Ecological Modelling 130 (2000) 111119

fuzzy logic to quantify a set of subjective and qualitative observations could so dramatically reveal an effect that would have otherwise remained buried in masses of data. The end of the time series, which reects the restocking of the farm in 1994, reveals a rapid return to conditions as heavily impacted as any seen earlier. It appears that leaving the farm fallow for a year had little long-term effect, and the rapid return to high impact status in periods C and F, which is evident both in the partial memberships of Fig. 1 and the defuzzied score of Fig. 2, does not offer any evidence of long-term recovery during periods when the impacts are lessened.

may improve the harvest of other organisms (one often nds lobster and crab shermen in the vicinity of sh farms), but it may also lead to undesirable algal blooms and has been blamed for the appearance of decaying seaweed along the shoreline. It is probably a general truth that environmental indices are meaningful only within a restricted context, and that the degree of impact of any project depends on the viewpoint of the observer. In the case of aquaculture this is certainly true, and I raise the issue without hope that it will be easily resolved.

4. Combining fuzzy indices

3.2. Different types of en6ironmental effect

A second example is really a reection on how far one can go in quantifying environmental effects. The preceding example deals with one specic class of impacts, but most activities, whether human or natural, have numerous different effects, and any environmental index must allow for these and provide a means of synthesising a comprehensive measure of all these effects. For example, aquaculture has many potential environmental effects, including (Silvert 1992): carbon accumulation under the site, and associated degradation effects oxygen consumption by dense populations of farmed organisms release of nutrients into the water column As shown in the preceding section, one can quantify the rst of these by use of fuzzy logic, and the other two are relatively easy to quantify. However, how does one combine the results of such different types of effects, which not only affect different components of the ecosystem, but operate on very different space and time scales? Is it possible to incorporate the fact that some of these effects, such as the release of nutrients into the water column, may actually be benecial? These questions reach well beyond the realm of the present study and can involve issues from a wide range of elds. For example, increased productivity due to enrichment of the water column One aspect of the problem of applying fuzzy indices to complex situations is the need to combine different indices representing different impacts. Perhaps the strongest positive feature of fuzzy logic in developing environmental indices is the ability to combine such indices much more exibly than one can combine discrete measures, which are often simply binary indices corresponding to ordinary (crisp) sets, such as acceptable versus unacceptable. For this reason it is important to discuss how to combine different fuzzy indices. The issues involved are both mathematical and strategic. The different fuzzy sets used in classifying environmental effects can be classied as complementary or independent. Complementary sets are ones which describe different ranges of the same properties; examples are pristine versus polluted, or the sets nil/moderate/serious/extreme impact used in the preceding example from Angel et al. (1998). Independent sets are ones that address different properties, such as very little seaweed, a few crabs, and thick patchy bacterial mats. A common example is that in describing humans, the sets of short people and tall people are complementary, but the sets of short people and fat people are independent. As a special case, any set and its complement are complementary. Of course with fuzzy sets it is possible to have membership in two or more complementary sets; for example it is possible to belong both to a set

W. Sil6ert / Ecological Modelling 130 (2000) 111119

117

and to its complement, but the dening characteristic is that the sum of partial memberships in complementary sets must be one. For example, if all men are divided into the four fuzzy sets short, tall, thin and fat, then someone of medium build might be 40% short, 60% tall, 30% thin and 70% fat; the memberships in short and tall sum to one, as do the memberships in thin and fat, but there is no fundamental relationship between memberships in short and thin (although the memberships in independent sets may be statistically correlated). The manner in which one combines the memberships in two or more sets depends on whether they are complementary or independent. Complementary sets are the easiest to deal with, because they must of necessity be comparable. Independent sets do not have to be comparable in any sense, and they can represent different space and time scales, different areas of effect, and even different sets of objectives. However, the rules of combination cannot be determined on purely mathematical grounds, which is what I mean by strategic considerations. Two examples have already been given of situations which call for different acceptability criteria. One is the case of environmental conditions which affect the survival of a valued species all environmental variables must be acceptable, or the species will die out, so the rule of combination is v = min(v1, v2, v3,). Another case is one where we wish to avoid this kind of value judgement and use the symmetric sum, dened in Eq. (1). It should also be noted that some operators, like min (the minimum value of its arguments), treat all of the partial memberships equally, while the symmetric sum and some other operators permit different observations to have different weights. The complexity of deciding how to combine partial memberships is dramatically represented by the interplay between scientic, conservation, social, economic, and political issues in assessing environmental effects. It is difcult to reconcile the scientic and quality-of-life issues associated with cutting old-growth forests or actions that might reduce whale populations with concerns about jobs and revenue. It is possible that fuzzy

logic could play a role in addressing these conicts, as it does in the more narrow range of factors relevant to benthic impacts on sh farms. However, I acknowledge that fuzzy logic has never been applied to such problems (at least so far as I am aware), and there is no solid evidence that it would provide a common framework for resolving conicts, rather than being seen as a particularly suspect form of mathematical obfuscation. One can envision fuzzy logic as a valuable tool in quantifying environmental conditions that could have applicability in conict resolution going well beyond its purely scientic utility, but that remains unproved.

5. Multi-objective decision making As pointed out earlier, society is not always able to reach consensus on the value of certain components of the environment, so that effects which are acceptable to some segments may be far less acceptable to others. Examples include the abundances of certain birds and marine mammals, which are highly prized for their beauty and entertainment value by recreational users of the environment, but are seen as predators and competitors by shers and farmers. Many complex issues deal with the marginal (i.e. incremental) value of natural lands, such as the question of how much old-growth forest should remain protected and how much can be exploited. An obvious generalisation of the concept of acceptability, and of the mathematical concept of a fuzzy set of acceptable conditions, is to dene several acceptability sets, each representing the viewpoint of a different segment of society. For each environmental variable i, let vij be the membership of the measured value of i in acceptability set j (i.e. the measure of how acceptable i is from viewpoint j, on a scale of 01). Then the overall degree of satisfaction of objective j is the combined fuzzy memberships over all variables i. We can also dene a set of weights wij which represent the weight of variable i in satisfying objective j, namely its importance in terms of meeting that objective. Although the mathematics tends to become cumbersome, for example we

118

W. Sil6ert / Ecological Modelling 130 (2000) 111119

could transform Eq. (1) into a system of even more complicated equations indexed by the different objectives j, the underlying idea is fairly straightforward. As an example of these considerations, consider indices for marine systems. Possible objectives could be good recreational use, sustainable wild sheries, and protable shellsh aquaculture. Relevant variables could include the abundance of birds, and the occurrence of toxic marine algae. From a recreational point of view birds are usually considered very desirable (except for very common species like seagulls), and the more abundant they are, the higher the degree of satisfaction of the recreational objective. Fishermen do not always share this love for birds; shermen often see cormorants as dangerous competitors, while some of the most attractive birds like oystercatchers and eider ducks are efcient predators on shellsh. Toxic algae are merely a nuisance to swimmers and other recreational users, but can cause the closure and economic ruin of shellsh farms. Thus the memberships are different (the higher the number of eider ducks the greater the membership, i.e. the acceptability, for bird watchers and the lower the acceptability to mussel farmers) and the weights can also vary greatly (toxic algae are a minor irritant to swimmers but a nancial disaster for both wild and farmed sheries).

factors. For example, point source air pollution is usually of more concern to nearby residents than to the producers, but everyone agrees that air pollution is undesirable, so it should be possible to arrive at an agreement regarding acceptability levels. 2. This enables the participants to focus on areas where there is real disagreement, such as marine mammals and birds in the situation described above, without being distracted by issues on which consensus is readily achievable. 3. Once the basically different objectives of various groups have been clearly delineated, sets of acceptability scores for the different objectives can be calculated for different scenarios, and used to provide a focus for further discussion. It is of course unrealistic to assume that mathematical calculations will be accepted as a decisive means to solve complex social and political issues, but they offer a quantitative expression of the differing values and needs expressed during negotiations, and as such can help clarify the basic underlying issues.

7. Conclusion Fuzzy logic can be applied to the development of environmental indices in a way that resolves many common problems, such as incompatible observations and implicit value judgements. It bridges the gap between scientic measurement and the fullment of social objectives and provides a way to translate a wide variety of information objective data, qualitative information, subjective opinions, and social needs into a common language for characterising environmental effects. Because it offers a means for assigning different weights to different types of observation, it can deal with differing perceptions of environmental risk. It can also deal with missing data, a common problem in assessing environmental impacts. Although it has yet to be extensively used in developing environmental indices, it has shown its

6. Consensus building with fuzzy logic Multi-objective decision making is at the heart of the political process, which involves trying to build a consensus among groups with different values and goals. The formalism described above can be used to identify key areas of disagreement and may possibly contribute to the resolution of conict in complex situations by providing a language for quantifying these disagreements. I propose a three-step procedure to deal with these kinds of disagreements: 1. Identify environmental variables on which agreement can be reached, and reach consensus on the partial memberships and weighting

W. Sil6ert / Ecological Modelling 130 (2000) 111119

119

value in a pilot study, and I believe that it merits considerable further investigation.

References
Angel, D., Krost, P., Silvert, W., 1998. Describing benthic impacts of sh farming with fuzzy sets: theoretical background and analytical methods. J. Appl. Ichthyol. 14, 18.

Silvert, W., 1979. Symmetric summation: a class of operations on fuzzy sets. IEEE Trans. Syst. Man, Cyber. SMC-9, 657 659. Reprinted in: Dubois, D., Prade, H., Yager, R.R. (Eds.), Readings in Fuzzy Sets for Intelligent Systems. Morgan Kaufman, San Mateo, CA, 1993, pp. 77 79. Silvert, W., 1992. Assessing environmental impacts of nsh aquaculture in marine waters. Aquaculture 107, 67 71. Silvert, W., 1997. Ecological impact classication with fuzzy sets. Ecol. Model. 96, 1 10.