Content uploaded by Gavin Leslie Simpson
Author content
All content in this area was uploaded by Gavin Leslie Simpson on Oct 08, 2021
Content may be subject to copyright.
Final Report Project WFD60
Macroinvertebrate classification diagnostic tool development
August 2007
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
© SNIFFER 2007
All rights reserved. No part of this document may be reproduced, stored in a retrieval system or
transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise
without the prior permission of SNIFFER.
The views expressed in this document are not necessarily those of SNIFFER. Its members, servants or
agents accept no liability whatsoever for any loss or damage arising from the interpretation or use of the
information, or reliance upon views contained herein.
Dissemination status
Unrestricted
Use of this report
The development of UK-wide classification methods and environmental standards that aim to meet the
requirements of the Water Framework Directive (WFD) is being sponsored by UK Technical Advisory
Group (UKTAG) for WFD on behalf its member and partners.
This technical document has been developed through a collaborative project, managed and facilitated by
SNIFFER and has involved the members and partners of UKTAG. It provides background information to
support the ongoing development of the standards and classification methods.
Whilst this document is considered to represent the best available scientific information and expert
opinion available at the stage of completion of the report, it does not necessarily represent the final or
policy positions of UKTAG or any of its partner agencies.
Project funders
SNIFFER, Scottish Environment Protection Agency (SEPA), Environment Agency
Research contractor
This document was produced by:
Don Monteith and Gavin L. Simpson
ECRC-ENSIS
Pearson Building, Gower Street
WC1E 6BT
SNIFFER’s project manager
SNIFFER’s project manager for this contract is: Ian Fozzard, SEPA
SNIFFER’s project steering group members are:
Ben McFarland, Environment Agency
Geofff Philips, Environment Agency
Deirdre Tierney,Environmental Protection Agency of Ireland
David Rendall, SEPA
Robin Guthrie, SEPA
Mary Gallagher, Environment and Heritage Service Northern Ireland
Mary Hennessy, Scottish Natural Heritage
Tristan Hatton Ellis, Countryside Council for Wales
Stewart Clarke, Natural England
SNIFFER
First Floor, Greenside House
25 Greenside Place
EDINBURGH EH1 3AA
Scotland
UK
Company No: SC149513
Scottish Charity: SCO22375
www.sniffer.org.uk
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
EXECUTIVE SUMMARY
WFD60: Macroinvertebrate diagnostic tool development (August, 2007)
Project funders/partners: SNIFFER
Background to research
This project (WFD60) forms part of the UK Strategy for the implementation of the EC Water
Framework Directive (WFD: European Union, 2000). Within its broad remit the WFD requires
the development of ecological classification tools for the purpose of determining ecological
status, with reference to specific environmental pressures. The WFD requires that these tools
should assign lakes to one of five categories, (High, Good, Moderate, Poor, Bad) to indicate
conditions relative to what is considered to be “good status”. This report focuses on the
development of a tool with which to determine the extent of the pressure of acidification on lake
macroinvertebrate communities.
Objectives of research
The primary objective is the development of a method and tool with which to assess the
pressure of acidification (a major threat to the ecology of acid-sensitive fresh waters, particularly
in the UK uplands) on the benthic macroinvertebrate assemblage of lakes.
Key findings and recommendations
Tool development under WFD60 was severely delayed due to problems obtaining sufficient high
quality biological and chemical data. The dataset used to support this phase is still less than
satisfactory, comprising data for only 105 sites and representing a subset only of the chemical
variables that would have been useful for explanatory data analysis. Due to the paucity of acid
anion data from one source and dual endpoint (or Gran) alkalinity from another, the final
physico-chemical dataset was built using one of two commonly used expressions of acid
neutralising capacity (ANC) and a few associated determinands.
Our assessment of the literature regarding macroinvertebrate-acidification inference techniques
concluded that none were appropriate for this assignment. In most cases macroinvertebrate
communities have been used to infer pH, but pH per se carries little information on acid
sensitivity or the likelihood that a site has acidified.
We show, through an investigation of the output of the Steady State Water Chemistry (SSWC)
Model and palaeoecological diatom-pH reconstructions, how ANC can be used as an indicator
of damage, in terms of modelled ANC change, diatom-inferred pH change and the mobilisation
of labile inorganic aluminium (Allab) concentration.
Furthermore, we show that prediction of the likelihood and level of acidification can be refined
by using ANC in conjunction with calcium concentration.
Assessment of chemical data from the UK Acid Waters Monitoring Network demonstrates that
Allab concentration, possibly the most important agent of damage associated with acidification,
will rarely if ever reach biologically toxic concentrations in sites with an ANC above 40 µeq l-1.
Conversely, sites which currently have a negative ANC are highly likely to exhibit biologically
toxic Allab concentrations.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
We show that ANC and Allab explain as much variance in a small high quality macroinvertebrate
dataset as pH and propose that macroinvertebrate community structure may carry sufficient
information for the level of physico-chemical damage to be inferred through its relationship with
ANC and calcium concentration.
In the expanded dataset, representing 105 lakes, we again show that ANC is strongly related to
the principal axis of macroinvertebrate species variation between sites.
We show that certain attributes of macroinvertebrate community structure pertinent to normative
definitions also vary along an ANC gradient. In particular, a crude measure of macroinvertebrate
species richness, as inferred by the total number of species identifiable to species level, is
tightly related to ANC. This is consistent with observations in the literature that
macroinvertebrate diversity may be reduced by anthropogenic acidification but not by natural
acidity (i.e. at sites where pH is depressed by organic acids only). Several individual species
show sharply truncated distributions on Allab gradients and species often ceased to be present in
waters with mean annual Allab concentrations over 10 µg l-1.
We created a “damage matrix” to provide an a priori physico-chemical classification of all sites
in the WFD60 database by ANC and calcium concentration into WFD compliant classes, i.e.
HIGH, GOOD, MODERATE, POOR, BAD. Owing to the sparsity of the data we then
condensed these classes into three representing HIGH-GOOD, MODERATE and POOR-BAD.
We used a classification tree approach to predict the a priori defined class of each site using its
macroinvertebrate assemblage. Classification trees are a powerful yet simple way of predicting
classes from a set of predictor variables (in this case, macroinvertebrate species and broader
macroinvertebrate groups).
After using a large range of biological input variables, including data at species level (i.e. the
proportions of individual taxa) we found that summary data only, in the form of minimum species
richness (MSR) of the full assemblage, the minimum number of species in certain biological
groups, and the proportion of individuals represented by certain groups, was necessary to
maximise the successful classification rate. The final tree classification used these variables
only.
We found that a simple rule, i.e. MSR >or<12.5, provided the most powerful criterion for
distinguishing between damage classes at the primary level. Further splits were based on the
number of non-leptophebid (i.e. mostly acid-sensitive) Mayfly taxa, the presence/absence of
bivalves, the proportion of Ephemeropteran, Plecopteran and Trichopteran individuals in the
entire assemblage, and the minimum number of stonefly taxa. The apparent misclassification
rate of this tree was 18.3%. We determined that the tree should be able to correctly assign class
status to random independent samples between 77 – 78% of the time.
This simple approach was able to distinguish between acidified and naturally acid (i.e. high
DOC, low sulphur deposition) lakes that tend to support relatively large numbers of taxa.
Apparently more complex, species-based, models such as the Acid Water Indicator Community
model (AWIC) are perhaps better tuned to predict pH but have limited value in this sense.
While the divisions on this tree form our current “best” model, we have major reservations with
respect to the total number of sites in the dataset and the distribution of sites at the acidified end
of the gradient. The model as it stands is clearly not fit for purpose but would benefit greatly
from the addition of 30-40 more sites in an acidified condition.
While this is a categoric approach to classification, class predictions could be converted to
EQR-compatible site scores to meet WFD requirements. There are a number of methods to
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
achieve this, but the most robust would use a method known as “bagging” to determine the
probability of membership of each site in the most likely class and neighbouring classes, to
provide a sliding score. The proposed increased number of sites would be essential for this
technique to be used effectively.
We tested the tool qualitatively on 51 sites for which chemical data were not adequate to be
included in the original training set. Generally the classification of sites was highly consistent
with geographical location although a few sites were clearly misclassified.
Current model weaknesses are likely to be principally due to the paucity of sites for which data
are available at the acid and acidified end of the physico-chemical gradient. The imbalance of
sites in the training set also prevents us from deriving predictions of the probability of correct
classification using a “tree bagging” technique.
We recommend that biological and physico-chemical data are gathered for a further 30-40
acidified sites before any attempt is made to refine the existing model.
Before implementation, we recommend the tool is tested on 1) time series data, to allow an
assessment of temporal variability of output, and 2) sites for which detailed multi-proxy
biological records are available, so that the macro-invertebrate inferred damage class can be
related to wider-ecosystem indications of damage by acidification.
Key words: Water Framework Directive, Lakes, Acidification, Littoral Macroinvertebrates,
Classification, Classification Trees, Acid Neutralising Capacity, Aluminium, pH.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
TABLE OF CONTENTS
1. INTRODUCTION 1
1.1 The Water Framework Directive and Macroinvertebrates 1
1.2 Water Framework Directive and Acidification 2
2. ACIDIFICATION 2
2.1 Freshwater acidification in the UK 2
2.2 Acidification and macroinvertebrates 3
3. PHYSICO-CHEMICAL INDICATORS OF THE PRESSURE OF ACIDIFICATION 4
3.1 Distinguishing between the effects of Acidity and Acidification 4
3.2 Acid Neutralising Capacity (ANC) 4
3.2.1 Estimating change in ANC using the Steady State Water Chemistry model 6
3.2.2 ANC as a direct indicator of acidification pressure 7
3.3 Labile inorganic aluminium 9
3.4 Palaeoecological support for the use of ANC as an indicator of acidification 11
3.5 Summary of physico-chemical indicators of acidification 12
4. MACROINVERTEBRATES AS INDICATORS OF ACIDIFICATION 12
4.1 Raddum Indices 12
4.2 The Henriksson and Medin Index 12
4.3 AWIC – Acid Water Indicator Community 13
4.4 Weighted Averaging based approaches 14
4.5 Diversity based indices 14
5. PRELIMINARY DATA ASSESSMENT 16
5.1 Data sources 16
5.2 Quality and screening of the macroinvertebrate – water chemistry dataset 16
5.3 The WFD60 database 17
5.4 The interim dataset 17
5.4.1 Indirect ordination 18
5.4.2 Direct ordination with chemical variables 19
5.4.3 Species distributions and labile inorganic aluminium 23
5.5 Exploratory Analysis of the full WFD60 dataset 28
5.5.1 Site ordination 28
5.5.2 Macroinvertebrate distributions on ANC and Ca2+ gradients 29
5.5.3 Summary of exploratory analysis 46
6 WFD TOOL DESIGN PRINCIPLES 47
6.1 Other WFD Schemes under development 47
6.2 Classification under WFD60 48
6.3.1 Generation of a “damage matrix” 49
6.3.2 Macroinvertebrate input data 51
6.3.3 Decision trees 51
7 WFD60 RESULTS 54
7.1 WFD60 Classification 54
8. A GEOGRAPHIC TEST OF THE CURRENT TOOL 68
9. RECOMMENDATIONS 71
10 CONCLUSIONS 72
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
List of Tables
Page
Table 1.1 WFD Normative definitions for biological elements
(macroinvertebrates) relating to High Good and Moderate status.
1
Table 1.2 WFD Normative definitions for acidification-related physical-chemical
quality elements relating to High Good and Moderate status.
2
Table 4.1 Original classification framework defined by Raddum (1988). 12
Table 5.1 Variance of the 35 lake macroinvertebrate dataset explained by
chemical variables applied individually in Canonical Correspondence Analysis
(CCA).
19
Table 5.2 Summary statistics for CCA of macroinvertebrate assemblages for
35 acid-sensitive UK lakes.
19
Table 5.3 Biplot scores for chemical variables and ordination axes for the CCA
of macroinvertebrate assemblages for 35 acid-sensitive UK lakes
20
Table 5.4 Macroinvertebrate species scores for the CCA of macroinvertebrate
assemblages for 35 acid-sensitive UK lakes.
21
Table 6.1 Damage matrix, based on understanding of relationships between
ANC and calcium concentrations and evidence from palaeoecological and
hydrochemical models of acidification, and contemporary relationships with Allab
and macroinvertebrate assemblage characteristics.
49
Table 7.1 WFD60 Tree Cross-classification 67
Table 7.2 Tree Posterior Class Probabilities 68
Table 8.1 Sites used to test the WFD60 tool, resulting classification and
posterior probabilities derived from the classification tree.
70
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
List of Figures
Page
Figure 3.1 Relationship between contemporary ANC and SSWC inferred pre-
industrial ANC based on 830 water samples from UK acid sensitive waters.
8
Figure 3.2 The amount of ANC reduction as a result of anthropogenic
sulphur deposition (inferred by the SSWC model) related to current
(measured) ANC.
9
Figure 3.3 Relationship between labile inorganic aluminium concentration
and ion balance ANC for all water samples in the UK Acid Waters Monitoring
Network database.
10
Figure 3.4 Change in pH (as inferred from the difference in diatom inferred
pH between samples from the top and bottom of sediment cores) in the
context of contemporary ANC.
11
Figure 5.1 Correspondence Analysis (CA) of macroinvertebrate
assemblages for 35 acid-sensitive UK lakes
18
Figure 5.2 CCA Ordination plot for Axes 1 and 2 of the macroinvertebrate –
water chemistry dataset for 35 acid-sensitive UK lakes
20
Figure 5.3 The relationship between labile inorganic aluminium concentration
and the number of species identified to species level. Lines represent a fitted
GAM model
24
Figure 5.4 Presence/absence of the more common taxa in the interim
dataset in relation to labile inorganic aluminium concentration.
25
Figure 5.5 Non-metric multidimensional scaling (NMDS) ordination plots of
the 105 sites in the WFD60 database, based on macroinvertebrate species
chord distances.
28
Figure 5.6 Non-metric multidimensional scaling (NMDS) ordination plots of
the 105 sites in the WFD60 database, based on macroinvertebrate species
chord distances.
30
Figure 5.7 The presence/absence of taxa which occur in 10 or more sites
(full dataset), across the ANC gradient. Black line represents a GAM function
(Poisson error distribution and logit function).
31
Figure 5.8 The presence/absence of taxa which occur in 10 or more sites
(full dataset), across the calcium gradient.
38
Figure 5.9 The number of taxa identified to species level related to mean
ANC for the 105 sites in the WFD60 training set. Lines represent a GAM
function (Poisson error distribution and log-link function)
45
Figure 5.10 The number of taxa identified to species level related to mean
Ca2+ for the 105 sites in the WFD60 training set.
46
Figure 6.1 Distribution of sites according to ANC and Ca2+ concentration. 50
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
Diagram shaded according to the damage matrix (Table 6.1)
Figure 7.1 Classification tree based on the 105 lakes in the WFD60 training
set and acidification damage matrix provided in Table 6.1
56
Figure 7.2 Minimum species richness (MSR) on a gradient of ANC for the
105 lakes in the WFD60 training set.
57
Figure 7.3 Minimum species richness (MSR) on a gradient of calcium
concentration for the 105 lakes in the WFD60 training set.
58
Figure 7.4 Minimum number of Mayfly taxa for each site that are not in the
family Leptophlebiidae (nonLeptoMayfly.MSR) on a gradient of ANC for the
105 lakes in the WFD60 training set.
59
Figure 7.5 Minimum number of Mayfly taxa for each site that are not in the
family Leptophlebidiidae (nonLeptoMayfly.MSR) on a gradient of calcium
concentration for the 105 lakes in the WFD60 training set.
60
Figure 7.6 The proportion of all individuals for each site that are either
Ephemeroptera, Plecoptera or Trichoptera on a gradient of ANC for the 105
lakes in the WFD60 training set.
61
Figure 7.7 The proportion of all individuals for each site that are either
Ephemeroptera, Plecoptera or Trichoptera on a gradient of Ca2+ for the 105
lakes in the WFD60 training set.
62
Figure 7.8 The minimum number of stonefly species for each site on a
gradient of ANC for the 105 lakes in the WFD60 training set.
63
Figure 7.9 The minimum number of stonefly species for each site on a
gradient of Ca2+ for the 105 lakes in the WFD60 training set.
64
Figure 7.10 The minimum number of bivalve taxa for each site on a gradient
of ANC for the 105 lakes in the WFD60 training set.
65
Figure 7.11 The minimum number of bivalve taxa for each site on a gradient
of Ca2+ for the 105 lakes in the WFD60 training set.
66
Figure 8.1 UK Map of the WFD60 classification tree classification of 51 lakes
69
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
APPENDICES
Appendix I: Sites and macroinvertebrate sample dates included in the WFD60 training set. 76
Appendix II: FURSE species codes and names included in the WFD60 training set 79
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
1
1. INTRODUCTION
This project (WFD60) forms part of the UK Strategy for the implementation of the EC Water
Framework Directive (WFD: European Union, 2000). Within its broad remit the WFD requires
the development of ecological classification tools for the purpose of determining ecological
status, with reference to specific environmental pressures. Our primary objective is the
development of a method and tool with which to assess the pressure of acidification (a major
threat to the ecology of acid-sensitive fresh waters - particularly in the UK uplands) on the
benthic macroinvertebrate assemblage of lakes.
The WFD requires that the tool should assign lakes to one of five categories, (High, Good,
Moderate, Poor, Bad) to indicate conditions relative to what is considered to be “good status”.
Inherent errors are to be defined and quality assurance provided. Ultimately, the tool will be
provided in the form of simple software that will allow the funding agencies to test further
datasets.
In this report we review WFD lake classification requirements and our understanding of how
these may apply to the issue of macroinvertebrates and acidification. We describe the physico-
chemical and biological database compiled under the project and demonstrate links between
acidification pressure metrics and macroinvertebrate normative definition compatible
characteristics of lakes. We briefly consider the philosophy behind the ongoing development of
other classification tools within the UK before presenting an alternative approach, based
primarily on the concept of classification trees, which is used in WFD60 to underpin the tool.
1.1 The Water Framework Directive and Macroinvertebrates
The Water Framework Directive (WFD) requires EU Member States to monitor the ecological
status of water bodies with the aim of achieving ‘good ecological status’ for all water bodies by
2015. The Directive provides normative definitions for biological status classification (Annex V)
and these are summarised for macroinvertebrates in Table 1.1. Waters deemed less than
moderate are classed either poor (major alterations to ecological quality) or bad (severe
alterations to ecological quality).
Table 1.1 WFD Normative definitions for biological elements (macroinvertebrates)
relating to High Good and Moderate status.
Feature High status Good status Moderate status
Taxonomic
composition and
abundance.
Ratio of disturbance
sensitive to
insensitive taxa.
Level of diversity
Corresponds totally or
nearly totally to undisturbed
conditions.
No sign of alteration from
undisturbed levels.
No sign of alteration from
undisturbed levels.
Slight changes from the
type-specific communities.
Slight alteration from type-
specific levels.
Slight signs of alteration
from type-specific levels.
Differ moderately from the
type-specific communities.
Major taxonomic groups of
the type-specific community
are absent.
Substantially lower than the
type-specific level and
significantly lower than for
good status.
Substantially lower than the
type-specific level and
significantly lower than for
good status.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
2
1.2 Water Framework Directive and Acidification
Acidification is one of the key pressures covered by the WFD. Physico-chemical normative
definitions (Table 1.2) are defined directly with regard to High status, but are otherwise
dependent on the biological classification (Table 1.1).
Table 1.2 WFD Normative definitions for acidification-related physical-chemical quality
elements relating to High Good and Moderate status.
Feature High status Good status Moderate status
Level of pH, acid
neutralising capacity
etc.
Corresponds totally or
nearly totally to undisturbed
conditions.
No sign of anthropogenic
disturbance alteration from
undisturbed levels.
Does not reach outside the
range established so as to
ensure the functioning of the
ecosystem and the
achievement of the values
specified for biological
quality elements.
Conditions consistent with
the achievement of the
values specified for
biological quality elements.
2. ACIDIFICATION
2.1 Freshwater acidification in the UK
Acidification results from the atmospheric deposition of sulphurous and nitrogenous
compounds, derived from industrial, vehicular and agricultural sources, and represents one of
the most detrimental of anthropogenic impacts on upland freshwater ecosystems. Its effects are
most pronounced in regions where geochemical weathering rates are relatively poor, including
many upland areas in the west of Britain, southwest Northern Ireland and the Republic of
Ireland where acidic peaty soils overlie poorly weatherable lithologies such as granites,
sandstones and shales. Palaeoecological work has shown that many lakes (and by inference
connecting streams) in these regions have acidified by up to 2 pH units since the onset of the
industrial revolution as a direct result of acid deposition (Battarbee et al., 1990). Other effects on
water chemistry include the mobilisation of biologically toxic labile inorganic aluminium (Allab),
reduced availability of dissolved inorganic carbon (DIC) in the form of bicarbonate or dissolved
carbon, the chronic depletion of concentration of base cations, such as calcium and magnesium
and, possibly, the depleted availability of phosphorus. Recently it has also been shown that
dissolved organic carbon (DOC) concentration is influenced by acid deposition (e.g. Evans et
al., 2006) and it is likely that acidified lakes would often have exhibited substantially higher DOC
levels prior to the onset of acidification. The UK Acid Waters Monitoring Network (AWMN)
demonstrates that acidified waters have benefited from a substantial drop in the rate of sulphur
deposition over the last two decades (Monteith and Evans 2005); severely acidified sites have
shown reductions in Allab, while pH and alkalinity show increases in less acidic waters.
Acidification influences aquatic biota at all levels of the food chain, from primary producers,
such as aquatic algae and macrophytes, to macroinvertebrates, fish and even water birds.
Primary producers may be affected by the reduced availability of dissolved inorganic carbon
(DIC – required for photosynthesis), macronutrients such as phosphorus, and changes in inter-
specific competition. Aquatic animals are vulnerable to increased aluminium, hydrogen ion and
heavy metal toxicity, and changes in food availability and quality. The AWMN has found
evidence of recent changes in epilithic diatom, aquatic macrophyte, macroinvertebrate and
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
3
salmonid populations which are indicative of improved chemical conditions (Monteith et al.,
2005).
2.2 Acidification and macroinvertebrates
Macroinvertebrates are a particularly valuable biological group for bio-monitoring of aquatic
systems due to their sensitivity to various physico-chemical stressors, ubiquity, local
abundance, functional diversity and the relative ease with which species can be sampled and
identified. Their use as indicators of water quality has been widely documented (see for
example Rosenberg and Resh, 1993).
The relationship between macroinvertebrate community structure and the acidity of their aquatic
habitat has been thoroughly investigated over the past three decades, although the
predominant focus has been on running waters. Routine monitoring of macroinvertebrate
communities in lakes has only recently become commonplace so it is rarely possible to
demonstrate the nature of the biological response to acidification on a site specific basis. The
primary source of information has come, therefore, from spatial studies of the relationship
between species composition and acidity. On a broad acidity scale of, for example, pH 4.5 –
7.0, approximating to the full pH range over which sensitive sites in the UK may have acidified,
relationships between macroinvertebrate community structure and acidity are very clear and
may be summarised as follows:
a) Where species distributions are assessed with respect to a range of water quality
parameters in acid sensitive systems, water pH (the inverse of the logarithm of hydrogen
ion concentration (H+)) is often identified as the chemical variable which explains the
greatest amount of variance in the species data (Larsen et al., 1996; Davy-Bowker et al.
2003; Johnson et al., 2004).
b) Increased hydrogen ion concentration is thought to adversely affect osmoregulation in a
number of macroinvertebrate species (Herrman et al., 1993). However, it is not clear to
what extent the physiological effect of declining pH compares with co-varying factors
including:
• toxic effects of increasing aluminium solubility (which also interferes with
osmoregulation);
• the effect of iron and aluminium precipitates on feeding activity and oxygen
uptake;
• increasing toxic heavy metal solubility (e.g. cadmium, iron, lead and zinc);
• indirect nutrient controls on food availability; etc.
Indeed aluminium concentration, either represented by “total aluminium” or, more
appropriately, Allab is also often found to be a strong predictor of the assemblage in
spatial studies (e.g. Johnson et al., 2004) in addition to assessments of temporal
variation in monitoring studies (e.g. Monteith et al., 2005).
c) Species show varying distributions across pH gradients, as illustrated, for example by
Hämäläinen and Huttenen (1996) and Larsen et al., (1996). Certain species, including
several molluscs, amphipods and mayflies, are confined to the less acidic end of the
spectrum, whereas more tolerant species, including several stoneflies and chironomids
are often present throughout much of the range. Few acid tolerant species are solely
restricted to acidic sites.
d) Patterns are evident in ecological functioning across acidity gradients. Many acid
tolerant species feed predominantly on detritus, although certain carnivorous taxa, such
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
4
as water boatmen (Corixidae) and water beetles (e.g. Dytiscidae) may also thrive in the
absence of acid sensitive higher predators such as salmonids. Specialised grazers
feeding on attached algae (e.g. several mayflies) are often absent from more acidic
sites. Some species previously assumed to be detritivorous may fill the niche of grazers
in acidified systems and, therefore, may be considered more “generalist” in these
circumstances (Ledger et al., 2005). While the relative balance of the assemblage may
shift to one dominated by “collector-shredders” and predators with progressively more
acid water, biological monitoring suggests that chemical recovery and consequent
expansion of the aquatic food chain may encourage the recolonisation of other
predators, such as dragonflies and caddisflies (Woodward and Hildrew, 2001; Monteith
et al., 2005) and this could lead to an overall increase in predatory species in less acid
environments.
e) The net effect of these relationships is a strong negative relationship between acidity
and diversity parameters (see, for example, Petchey et al., 2004). There is evidence to
suggest a logarithmic relationship between species richness (i.e. total number of species
observed) and mean annual pH (up to a pH of circa 7.0), and this relationship does not
appear to differ significantly between lake and stream systems (Woodward, pers.
comm).
3. PHYSICO-CHEMICAL INDICATORS OF THE PRESSURE OF ACIDIFICATION
3.1 Distinguishing between the effects of Acidity and Acidification
If the relationship between the aquatic fauna and acidity is understood adequately it should be
possible to predict the acidity of a site given the biological data. This forms the basis of the
biological acidity-indicator systems discussed in further detail below. However, the remit of this
project is to develop a tool with which to classify the biological impact of acidification, i.e. the
extent to which the aquatic macroinvertebrate fauna has been affected by anthropogenically-
driven change in acidity.
It is important, therefore to distinguish between the role of acidity in determining differences in
biotic composition between sites, and the effects of the process of acidification on aquatic
organisms and communities. pH per se is not an indicator of acidification; the pH of acid-
sensitive lakes and streams in non-acidified regions (i.e. regions with low sulphur and nitrogen
deposition) varies substantially depending on the amount of geological buffering and the
contribution of organic acids from catchment soils.
In order for pH to be of value as a physico-chemical indicator under WFD60 it would first be
necessary to estimate pH reference conditions for the sites in the study sites. This is feasible
where palaeoecological information is available, since robust diatom-pH inference models are
available with quantifiable and low levels of predictive error. However, there is currently an
insufficient number of acid-sensitive lakes for which palaeoecological and macroinvertebrate
data are available. At the same time, the ability of process-based physico-chemical models to
predict pH is questionable as most do not account for possible links between changing acid
deposition and organic acid solubility (see for example Battarbee et al., 2005). It is clear,
therefore, that alternative physico-chemical metrics of acidification pressure are required for this
project.
3.2 Acid Neutralising Capacity (ANC)
ANC is the preferred response variable in process-based acidification models concerned with
Critical Loads. Representing the balance between base cations and strong acid anions, it is
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
5
relatively easy to model (in comparison with pH) and has been shown to be a robust predictor of
damage to salmonid populations (Lien et al., 1996). Relationships have also been shown for
macroinvertebrates (Raddum and Fjellheim, 1984; Raddum and Skjelkvåle, 1995), and diatoms
(e.g. Juggins et al., 1995). A negative ANC (i.e. a surfeit of acid anions over base cations)
implies elevated concentrations of acid cations, i.e. hydrogen and aluminium ions, and hence
acidic water. For water with a positive ANC, the surfeit of base cations may be accounted for by
organic anions (i.e. DOC), bicarbonate and (at higher values) carbonate. In the field, ANC may
fall during periods of high precipitation or snow melt as a result of dilution of high ANC
groundwater or the delivery of mineral acids stored in the snowpack. ANC is also depleted
during episodes of seasalt deposition, when marine cations temporarily displace hydrogen ions
from soil exchange sites, particularly during winter storms.
ANC is normally determined as follows:
i.e. Ion balance ANC = [Ca2+] + [Mg2+] + [Na+] + [K+]) - ([SO4
2-] + [NO3-] + [Cl-]
with all parameters expressed as equivalent concentrations.
However, Evans et al. (2001) suggested that this calculation was sensitive to the compound
errors associated with the measurement of the 7 constituent ions, and that the resulting “noise”
might hamper the detection of trends in ANC time series. An alternative expression of the same
balance (Alkalinity-based, or Cantrell, ANC) is derived from Gran, or dual endpoint alkalinity,
and organic acid (DOC) concentration. Here, an assumption is made regarding a standard
charge per milligrame of DOC. In the UK the most commonly applied standard is 4.5 µeq l-1 g-1
C for samples with a pH less than 5.5 or otherwise, 5.0 µeq l-1 g-1 C. Recent AWMN data
assessments include a further modification of the alkalinity-based expression to account for the
influence of Allab. An assumption is made that all labile inorganic aluminium is trivalent and the
conversion is 3 µeq l-1 µg-1 Allab.
Theoretically these approaches should yield similar results although this is not always the case.
Estimation of ANC from alkalinity, DOC and Allab is less sensitive to compound errors but carries
uncertainty with regard to the amount of charge (protonation) attributed to DOC which is known
to be pH dependent, and to a lesser extent Allab. Uncertainties associated with the ion balance
method for estimating ANC should become less important where annual means are determined
from a number of samples, as the effect of random errors should be dampened. Discrepancies
between methods are currently under investigation by the AWMN.
Unlike pH, that has been proposed elsewhere as a possible physico-chemical standard for
acidification pressure, ANC does convey information on the likelihood that a water body has
been damaged by acidification. In upland acid-sensitive systems Cl- is normally considered to
be derived from charge-neutral sea salt while the other two strong acid anions SO42- and NO3-
largely reflect inputs from acid deposition. ANC thus reflects, predominantly, the capacity of
base cation leaching to withstand the input of these latter two anions.
Natural waters almost invariably exhibit a surplus of base cations over strong acid anions, and
waters with negative ANC are, therefore, very likely to have acidified. Low pH waters are
common in unacidified regions where the influence of organic acids derived from organic soils is
strong. However, ANC is normally positive in these systems, regardless of acid-sensitivity.
Importantly for this project, Dangles et al. (2004) found diverse and functional macroinvertebrate
communities in naturally acid (i.e. high DOC) waters but not in acidified waters of similar pH.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
6
3.2.1 Estimating change in ANC using the Steady State Water Chemistry model
The steady state water chemistry (SSWC) model underpins UK critical load assessments and
deposition scenario modelling. The Henriksen SSWC model formulation assumes that the
greater the current base cation concentration:
i) the greater the likely source of bicarbonate weathering and ANC generation; and therefore,
ii) the lower the likelihood of acidification having taken place for a given deposition load.
SSWC uses measured SO42- concentration as an index of deposition, on the assumption that
this anion is mobile within catchments. The ratio of SO42- to base cation concentration is used
to determine what proportion of contemporary base cation concentration is due to acid
deposition, and this is then used implicitly to back-calculate baseline ANC for an assumed pre-
industrial sulphate concentration.
The key equation is:
[BC]0* = [BC]t* - F([NO3-] + [SO42-]t* - [SO42-]0*)
[BC]0* pre-industrial non-marine base cation concentration
[BC]t* current sum of non-marine base cations
F correction factor: the "F-factor"
[SO42-]t* current non-marine sulphate
[SO42-]0* pre-acidification non-marine sulphate: "sulphate zero"
In the SSWC model the components of this equation are calculated as follows:
a) Current total non-marine base cations;
[BC]t* = [Ca2+]* + [Mg2+]* + [Na+]* + [K+]*
b) The F-factor;
F = sine{ 90[BC]t* / S }
F is defined as the "change in base cation concentration per unit change in excess acid anions".
It is a function of current base cation concentration and may vary for a given lake over a period
of time. It is assumed that a high measured concentration of base cations in a lake indicates
high "weatherability" of soils within the catchment and a large pool of exchangeable base
cations in catchment soils.
S is an empirically derived value of [BC]t* for which F=1; i.e. when current total non-marine base
cations equal S, all acid deposition results in base cation leaching. Previous studies have found
that S may vary from 200-400 µeql-1 from site to site, but that generally speaking, at S=400 µeq
l-1 the pH of the water will be in the range 6.5-7.0 and the leaching of base cations causes no
change in pH. For critical loads work in the UK a value of S=400 µeql-1 is therefore used.
As a sine function the value of F can range from 0-1; in practice F ranges from near zero in
dilute lakes to F=1 for lakes with high levels of base cations. Because F is a sine function there
has to be a cutoff value of base cations above which F is taken to be 1; otherwise its value
would start to decrease again at base cation levels greater than the constant S. In the SSWC
model, F is therefore set to unity for any lake with [BC]t* > 400 µeql-1.
c) Sulphate zero;
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
7
The "background" concentration of SO42- is taken as 15 µeq l-1, the minimum figure observed in
a study of near pristine Norwegian lakes (Braake, 1989), plus an extra component which is
related to the "weatherability" of catchment soils and therefore proportional to [BC]t*. This figure
has been derived empirically in other critical loads studies:
[SO42-]0* = 15 + 0.16[BC]t*
When [SO42-]t* > 500 µeql-1, SO42- is removed from the SSWC model calculation to avoid
deriving improbably low critical load values for insensitive lowland sites. This means that base
cation zero is taken to be the current total non-marine base cation value, i.e. [BC]0* = [BC]t*.
This cutoff is intended to exclude catchments where SO42- concentrations are too high to have
been caused by atmospheric deposition alone.
It is assumed that pre-industrial nitrate concentration [NO3-] in acid sensitive lakes was
negligible, so that pre-industrial, baseline ANC can be calculated as:
ANC0 = [BC]0* - [SO42-]0*
The change in ANC according to the assumptions of the SSWC model can be calculated as the
difference between current measured ANC and baseline ANC (ANC0).
3.2.2 ANC as a direct indicator of acidification pressure
We applied the SSWC model to estimate ANC0 for 830 UK lakes and streams with pH < 7.0
from the DEFRA Freshwater Umbrella database held at UCL. This demonstrated the
relationship between contemporary chemistry and pre-industrial ANC, as determined by SSWC
with its various assumptions, on a wide spatial basis.
Contemporary pH showed a relatively weak relationship with ANC0. Unsurprisingly however,
given the model assumptions, the relationship between contemporary ANC and ANC0 is much
stronger (Figure 3.1). At the high ANC end of the plot the data largely fit the red 1:1 line, or
show various degrees of positive deviation but no strong tendency for departure between
current and pre-acidification ANC. With declining ANC the tendency for deviation from linearity
increases. This plot demonstrates that, according to the SSWC model:
a) pre-industrial ANC would rarely have been negative, even for sites which are strongly
negative ANC today;
b) there is little indication that sites with high ANC today (i.e. >100 µeq l-1) would have had
higher ANC in the past, i.e. these sites are unlikely to have acidified;
c) as ANC falls below 100 µeq l-1, there is an increasing likelihood that a site would have had
higher ANC in the past, i.e. that it will have acidified;
d) the likelihood of a site having experienced a large decline in ANC – e.g. 50 ueq/l – increases
as contemporary ANC declines towards zero and beyond.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
8
Figure 3.1 Relationship between contemporary ANC and SSWC inferred pre-industrial ANC based
on 830 water samples from UK acid sensitive waters. Red line = 1:1.
-200
0
200
400
-200 -100 0 100 200 300 400
measured ANC (µeq/l)
pre-industrial ANC (µeq/l)
Furthermore, the model implies that the amount of acidification predicted by the SSWC model
for a given contemporary ANC is dependent on the contemporary base cation (e.g. Ca2+)
concentration. Figure 3.2 illustrates that, according to the SSWC model:
a) the extent to which ANC is predicted to have declined for a given contemporary ANC is
positively related to contemporary Ca2+ concentration;
b) even at an ANC of 80-100 µeq l-1, sites with a relatively high current Ca2+ i.e. (80-100 µeq l-1)
may have lost ANC - although this mostly equates to a less than 10% a reduction and is unlikely
to be of great physico chemical or biological significance;
b) sites with a contemporary ANC as low as 10 µeq l-1 may not have acidified providing that the
current Ca2+ concentration is very low (i.e. below 20 µeq l-1);
c) despite large variation in the acidification threshold between Ca2+ classes, all types of site
with a current ANC of 0 µeq l-1 or less are modelled to have undergone a substantial reduction
in ANC;
d) the discrepancy between Ca2+ classes in the amount of ANC change for a given current ANC
is greatest at around 0 µeq l-1 and the discrepancy declines to negligible levels as current ANC
approaches -100 µeq l-1.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
9
Figure 3.2 The amount of ANC reduction as a result of anthropogenic sulphur deposition
(inferred by the SSWC model) related to current (measured) ANC. Data grouped into 5 calcium
concentration classes (Ca2+ units µeq/l).
-100
-80
-60
-40
-20
0
20
40
-100 -50 0 50 100 150 200
measured ANC (µeq/l)
SSWC inferred absolute reduction in ANC (µeq/l)
Ca <20
Ca 20-40
Ca 40-60
Ca 60-80
Ca 80-100
Thus, according to SSWC assumptions, the extent to which a site has acidified can be
approximated from contemporary ANC in the context of the base cation concentration.
Alternatively, the amount of deviation could be described in terms of acid anion concentration.
Hence, sites with very low but positive ANC (i.e. approaching zero) and a low sum of
concentrations of sulphate and nitrate may be unacidified, while sites with the same ANC but
larger concentrations of sulphate and nitrate are more likely to have been impacted.
3.3 Labile inorganic aluminium
Labile inorganic aluminium is mobilised as soils acidify and is known to be highly toxic to many
types of aquatic fauna. It has been shown to be the primary cause of salmonid death in
Scandinavia (Rosseland et al., 1990) and is thought to exert a strong control on
macroinvertebrate species composition. Rosseland et al. (1990) proposed that Allab became
toxic to fish at concentrations between 25 and 75 µg l-1. Recently, Allab was found to be the most
important direct chemical predictor of change in AWMN macroinvertebrate communities
(Monteith et al., 2005).
According to ECRC-ENSIS data holdings Allab concentrations rarely rise above 10 µg l-1 in UK
waters which have not been acidified by acid deposition (rare exceptions include sites in
catchments with unusual, sulphur-rich mineralogy, or exposed mine workings). In general terms,
therefore, Allab concentrations of over 10 µeq l-1 in upland systems will normally be indicative of
acidification and larger concentrations imply a greater likelihood of biological impacts.
Unfortunately, Allab is not routinely measured by the UK environment agencies and, beyond
AWMN/ECRC-ENSIS data holdings (and additional data generated and held by the FRS
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
10
Laboratory, Pitlochry), we are not aware of other datasets (combined water chemistry and
biology) that could be used for Allab calibration purposes in this project.
Broad classes of Allab concentration may be predicted on the basis of ANC. Figure 3.3 draws on
all water chemistry samples collated for all AWMN sites. This shows that samples with an ANC
above 40 µeq l-1 often have undetectably low Allab concentrations, and rarely exceed 10 µg l-1.
An ANC of circa 40 µeq l-1 might therefore be considered indicative of high status with regard to
the physico-chemical normative definitions for the pressure of acidification through the
mobilisation of aluminium. At the other extreme, water samples with an ANC below 0 µeq l-1
almost invariably have concentrations of Allab above the lower biological limit of 25 µg l-1
(Rosseland et al., 1990), and most samples with an ANC below -20 µeq l-1 have Allab
concentrations above the higher limit of 75 µg l-1. On physico-chemical evidence and published
biological information, therefore, water with an ANC of less than 0 µeq l-1 might be considered to
be in a condition ranging from “poor” to “bad”.
Figure 3.3 Relationship between labile inorganic aluminium concentration and ion balance ANC
for all water samples in the UK Acid Waters Monitoring Network database (comprising 24 acid
sensitive lakes and streams). Red line = 10 µg l-1 (theoretical maximum for unacidified UK waters);
blue line and black line represent lower and upper limits for toxicological effects on fish
(Rosseland et al., 1990).
0
20
40
60
80
100
120
140
160
180
200
-200 -100 0 100 200 300 400
ion balance ANC (µeq l
-1
)
inorganic aluminium (µg l
-1
)
Recently, Lawrence et al. (2007) investigated the relationship between inorganic aluminium
mobilisation and ion balance ANC with respect to DOC concentration for streams in the
Adirondack Mountains, USA. They demonstrated that the point at which Allab concentrations
become significant in an acidifying system equates to a “base cation surplus” (BCS) of 0 µeq l-1,
where BCS represents ion balance ANC minus the charge provided by DOC, using a single
charge estimate of circa 6 µeq mg-1 C. In effect this implies that for waters with negligible DOC
concentration (e.g. less than 1 mg l-1) Allab will only become mobilised, and therefore of
biological importance as ANC falls below 0 µeq l-1. In waters with higher DOC, mobilisation will
occur at higher ANC values. This observation should not be confused with the general
observation that DOC offers some protection to aquatic biota at low pH. However, it again
illustrates that ANC in isolation may not be sufficient to ascertain damage. For AWMN data a
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
11
similar relationship can be demonstrated between Allab and ANC and Ca2+ concentration. In
samples with Ca2+ concentrations below 20 µeq l-1, Allab mobilisation is only apparent in waters
at an ANC of around 0 µeq l-1 or below; whereas for the Ca2+ class of 60-80 µeq l-1, mobilisation
may occur at an ANC as high as circa 40 µeq l-1.
3.4 Palaeoecological support for the use of ANC as an indicator of acidification
Some of the first work to reveal the geographical extent and timing of freshwater acidification in
Scotland was based on the palaeoecological analysis of trends in diatom species composition in
sediment cores (e.g. Flower and Battarbee, 1983). Freshwater diatoms are excellent indicators
of lake acidity as most species occur only over very narrow pH ranges. The silicious outer-
casing of diatoms preserves in lake sediments so it is possible to use changes in the relative
abundance of species down a sediment core to infer how the pH of a lake has changed through
time. Techniques for calibrating the diatom change to a pH change have developed from simple
classification approaches, which grouped species into pH classes, to the more statistically
rigorous weighted averaging or maximum likelihood methods (Birks et al., 1990). A crude but
cost-efficient approach to estimate how much a lake has acidified is simply to determine the
difference between the diatom inferred pH of the surface (i.e. modern) sample and a sample
taken from circa 20 cm or more core depth, normally assumed (on the basis of sedimentation
rates) to represent conditions prior to the onset of anthropogenic acidification in the mid 19th
century.
Data provided by the DEFRA Freshwater Umbrella Programme demonstrate a striking
relationship between diatom inferred pH change and current ANC (Figure 3.4). Given the model
error of circa 0.3 pH units there is no obvious indication of acidification for sites with a current
ANC of more than approximately 40 µeq l-1. The relatively moderate inferred change in pH for
some sites with negative ANC may reflect the increased role of aluminium as a buffer in these
very acidic waters. These results are therefore fully consistent with physico-chemical
observations discussed earlier with respect to the value of ANC as an indicator of acidification
pressure.
Figure 3.4 Change in pH (as inferred from the difference in diatom inferred pH between samples
from the top and bottom of sediment cores) in the context of contemporary ANC. Blue dotted lines
indicate the Root Mean Square Error of Prediction for the weighted averaging based diatom pH
transfer function.
-0.6
-0.3
0.0
0.3
0.6
0.9
1.2
1.5
1.8
-50 0 50 100 150 200 250 300
Acid Neutralising Capacity ( µeq l-1)
diatom inferred pH change
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
12
3.5 Summary of physico-chemical indicators of acidification
In summary, both physico-chemical and palaeoecological models suggest that the relationship
between ANC and Ca2+ may be used both to infer the likelihood and extent of lake acidification,
and to predict the biologically crucial effect of elevated Allab concentration.
4. MACROINVERTEBRATES AS INDICATORS OF ACIDIFICATION
A principal challenge of this report is to relate WFD biological normative definitions for
macroinvertebrates (Table 1.1) to physico-chemical indicators of acidification. A number of
schemes and multivariate statistical techniques are currently available with which water acidity
can be predicted on the basis of the macroinvertebrate community structure. These are
summarised as follows.
4.1 Raddum Indices
The Raddum 1 Index (Raddum et al. 1988) is based on a simple classification system, whereby
a range of taxa are ascribed to one of four classes (see Table 4.1) depending on their pH
tolerance.
Table 4.1 Original classification framework defined by Raddum (1988)
Category species tolerating pH score
A >5.5 1.00
B >5.0 0.50
C >4.7 0.25
D <4.7 0.00
The community is then classified according to the highest scoring taxon present. Hence, the
presence of one or more individual Baetis sp. (Score = 1) results in a classification of 1.0, i.e.,
the highest possible rating for the sample. If no species representing the first three classes is
present the sample is rated as zero. The approach was developed by Fjellheim and Raddum
(1990) to include more taxa. The Raddum I Index is routinely deployed in monitoring
assessments such as the UNECE International Cooperative Programme on the Assessment of
Acidification of River and Lakes Programme (e.g. Raddum, 1999). Application of the Index to
the UK would require a revision of the taxon lists to allow incorporation of taxa not found in
Scandinavia. While this system is logically based on tolerance limits and is simple to apply, it is
likely to lack sensitivity, particularly in high DOC systems where reference condition
communities might only be expected to classify between 0.25 - 0.50.
The Raddum II Index is normally only applied to river samples. This Index is based on the
relationship between two groups of macroinvertebrates which show markedly different acidity
distributions, i.e. the ratio of the total number of individuals of the mayfly genus Baetis sp. (a
species of flowing waters) and the total number of individuals of acid tolerant stonefly species. A
major limitation of both these schemes is that they are pH based and, for reasons outlined in
Section 3.1, are not appropriate for the prediction of acidification pressure in landscapes where
water chemistry is influenced strongly by organic acids
4.2 The Henriksson and Medin Index
This multi-metric approach is based on a large Swedish macroinvertebrate dataset from humic
influenced rivers (Henriksson and Medin, 1986). The Henriksson and Medin Index integrates
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
13
presence/absence of indicator taxa and ratios of sensitive to non-sensitive taxa and an estimate
of species richness. The score for a sample is derived by calculating the sum of the score of 5
components representing: 1) the highest 0-3 Index score of a range of mayflies, stoneflies and
caddisflies; 2) the presence/absence of amphipods (score 0 or 3); 3) the presence of sensitive
groups (Hirudinea, Elmididae, Gastropoda and Bivalvia) (score 0-4); 4) the ratio of numbers of
individuals of the Genus Baetis sp. to stoneflies (score 0-2); and, 5) the total number of species
in a comprehensive list of 517 aquatic macroinvertebrate species (score 0-2). The criteria above
are highly compatible with the major requirements of WFD Normative definitions, i.e. with regard
to the need to use information on taxonomic composition and abundance (1 & 2), ratios of
disturbance sensitive to insensitive taxa (4), presence/absence of major taxonomic groups (1,3
& 5) and estimates of levels of diversity (5). It has been proposed that an Index of 6.0
represents a threshold below which the probability of effects of acidification are “likely”.
The Henriksson-Medin Index has been demonstrated by Johnson et al. (2004) to correlate with
the pH and ANC of 48 lakes in a mixed forest ecoregion of Sweden. They observed that the
Index showed a “funnel-shaped” relationship with these acidity variables and proposed that the
reduction in variance about the regression lines with increasing acidity was indicative of
increased acid stress. Classification of class boundaries was based on the extent of this
variance and resulted in five classes for pH (<5.0, 5.0-5.6, 5.6-6.2, 6.2-6.8 and >6.8)
representing extremely acid, very acid, acid, weakly acid and neutral-alkaline lakes respectively.
The “acid” class was defined by the first obvious increase in residual variance and its upper
boundary was selected to intersect with an Index of 6.0.
Four classes were defined for the relationship with ANC. All sites with an ANC < 20 µeq l-1
showed little residual variance, had Index scores below 3.0 and were deemed to be in the most
acid class. This, they argued, was consistent with the findings of Lien et al. (1996), which
suggested that an ANC 20 µeq l-1 represents a significant tolerance level for fish. Once again an
Index of 6.0 was used to define the lower boundary for the non-acid class, corresponding
approximately to an ANC of 150 µeq l-1.
Johnson et al. (2004) found that, in contrast with pH and ANC, relationships between the Index
and Allab concentrations were non-linear. However there were clear patterns in the data. Almost
all sites where Allab concentration was below 20 µg l-1 had relatively high Henriksson - Medin
scores (i.e. >5.0). This was taken as the threshold below which aluminium effects would be low
and is consistent with observations of Rosseland et al. (1990) that concentrations of less than
25 µg l-1 have negligible effects on aquatic biota. All sites but one, which lay above the upper
threshold of 75 µg l-1 (Rosseland et al., 1990), had scores of around 5.0 or less. These findings
are particularly interesting, given the comments in Section 3.3, and suggest that this Index may
have value as an indicator not only of acidity but also acidification status. On the basis of the
observations for aluminium concentration, however, a Henriksson - Medin score of 5.0 might be
a more appropriate threshold for good status than the more conservative value of 6.0.
4.3 AWIC – Acid Water Indicator Community
The AWIC, or Acid Water Indicator Community, classifications were developed by staff at the
Centre for Ecology and Hydrology, Dorset, primarily to assist the UK Environment Agency in
their assessment of the extent of ecological damage caused by the acidification of running
waters. Two classifications, for family level and species level data, were based on an extensive
biological and chemical database (487 samples, 410 sites) drawn from several regions of
England and Wales (Davy-Bowker et al., 2003; Davy-Bowker et al., 2005).
Both classifications are based on partial Canonical Correspondence Analysis (pCCA) in which
biological data are constrained by mean pH (based on a minimum of 5 samples taken over
three years), with significant physical factors such as altitude and slope included as covariables.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
14
The first axis scores for each taxonomic group are then allocated to one of six “bins” depending
on their relative position on this axis. The sample score is determined according to an average
score per taxon method (ASPT), termed AWIC (fam)–ASPT or AWIC (sp)–ASPT, for the family
and species classifications respectively. Initial testing of the AWIC (fam)–ASPT approach
(Davy-Bowker et al., 2003; Ormerod et al., 2006), using a “partially independent” dataset
demonstrated that this Index is strongly correlated with pH. However, the relationship is heavily
influenced by sites with a mean pH greater than 7.0 (which form the vast majority of the
dataset). For sites with pH <7.0 the relationship shows considerable scatter. While the
approach, at least at the species level, has potential for inferring the mean pH of a system, it
does not allow for the differentiation of acidified and naturally acidic waters which is central to
the WFD60 project.
4.4 Weighted Averaging based approaches
Being based on CCA, the methods applied in the AWIC classification are related to Weighted
Averaging (WA) regression, a commonly used environmental diagnostic procedure which has
been found to perform particularly well in diatom-pH calibration exercises (used to infer pH from
fossil remains in sediment cores). In essence, weighted averaging is used to determine the
optimal value of an environmental variable for individual species, and then the abundance
weighted average of the optima of all species present in a sample is used to infer that
environmental variable for a given site. The predictive error of the method may be assessed
using “bootstrapping” procedures, in which individual samples are taken from the dataset to test
the predictive power of models based on the remaining data.
Weighted averaging was used to investigate the relationship between macroinvertebrate
assemblages and stream minimum pH by Hämäläinen and Huttenen (1990). They compared it
with a “Tolerance limit” approach (TL), the tolerance limit of individual species being defined by
the lowest pH of water in which each occur. Hämäläinen and Huttenen (1996) found that WA
performed better than TL, resulting in lower root mean squared error of prediction. Larsen et al.,
(1996) also demonstrated the power of the WA approach for Norwegian streams, and showed
that macroinvertebrates are as good predictors of pH (Root Mean Standard Error of Prediction
for WA = 0.309 pH units) as more conventionally used diatoms. In their assessment of species
distributions across the pH gradient they found a variety of patterns (i.e. unimodal, sigmoidal or
indicative of either high or low pH) suggesting that a combination of Gaussian regression and
direct gradient analysis might be necessary to provide a complete overview of indicator taxa.
4.5 Diversity based indices
It is widely recognised that environmental stress influences species diversity, a term often taken
to encapsulate information both on the total number of species (species richness) and their
relative distribution (or evenness). Species richness per se has been shown to increase across
a gradient of pH in acid sensitive systems for a range of trophic levels (Petchey et al., 2004).
However, estimation of the true number of species in a population of interest is subject to
problems of rarefaction, i.e. the number of species in a sample will be dependent on the size of
the sample taken. Possibly one of the most robust and widely used diversity indices is
Shannon’s Index, defined as the sum of the product of the proportional abundance of each
species and its natural logarithm, converted to a positive number by a prefixed negative sign.
This can be translated into more meaningful values by determining its exponent (Hill’s N1; Hill,
1973). Hill’s N1 represents the “effective number” of abundant species in a sample and is thus
more readily interpretable than the original index.
An alternative and commonly used diversity index, which does not include information on
species proportions, is Margalef’s Index, defined as the number of species divided by the
natural logarithm of the total number of individuals. This makes the assumption that species
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
15
proportions will become less even as the total abundance of individuals increases. Resh and
Jackson (1993) found that this was the only community-based index which showed a significant
response in macroinvertebrates to an acid impact.
4.6 The need for a new macroinvertebrate-based acidification tool under WFD60
While most classification schemes reviewed above illustrate the potential for macroinvertebrate
communities to predict the acidity of a freshwater system, few meet the fundamental WFD
classification requirement to infer the pressure of acidification. The Henriksson Medin Index is
the only system which shows real potential in this respect, since relationships with ANC and
Allab have been demonstrated explicitly. However, the information necessary for us to assess
underlying model assumptions and the relative importance of each component are not
available, and we are unclear about the necessity of the current complexity of the model. Ideally
the WFD60 tool should be based on as simple a model as possible, to minimise potential
complications resulting from variation in sampling effort and taxonomic skill. Furthermore the
Henriksson Medin Index is specifically calibrated for the Swedish boreal eco-region. There is a
clear need, therefore, for the development of a new UK-based classification procedure under
WFD60, and this requires a new physico-chemical/biological database to cover acid-sensitive
UK lakes.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
16
5. PRELIMINARY DATA ASSESSMENT
A central part of the WFD60 project was the collation of a database to contain the biological and
physico-chemical information necessary to build and test the tool. At the outset it was
envisaged that a substantial database, comprising several hundred macroinvertebrate and
water samples, would result.
5.1 Data sources
Five main sources of data have been used within the WFD60 project
1) A lake macroinvertebrate – water chemistry dataset, derived from data from projects funded
by DEFRA, including the AWMN, and currently held in databases at ECRC-ENSIS;
2) Macroinvertebrate data generated in variations to the WFD60 contract and based on samples
collected by SEPA between 2003-2006
3) Water chemistry data for lochs in Galloway provided by the FRS Freshwater Laboratory,
Pitlochry.
4) Water chemistry data from the SEPA lake water chemistry database to accompany biological
data collated under point 2.
5) Macroinvertebrate and water chemistry data for lakes from the Environment Agency of
England and Wales.
Only macroinvertebrate and water chemistry data that could be applied to specific components
of the calibration exercise were included in the final WFD60 database. A substantial amount of
SEPA macroinvertebrate data could not be used in this project since the bulk of this was
collected in Autumn, whereas the dependency of this project on establishing links between
water chemistry and ECRC-ENSIS data holdings has meant that the focus must be on Spring
samples.
5.2 Quality and screening of the macroinvertebrate – water chemistry dataset
Protocols for macroinvertebrate sample collection and the water chemistry sampling and
analysis differ inevitably between the projects and programmes outlined in Section 5.1. We
have settled for “lowest common denominator” criteria as follows:
• Macroinvertebrate data must be derived from kick samples from lake stony littoral
habitats using a standard long-handled pond net. Each kick sample must be of at least
of one minute duration. Ideally more than one sample, from separate locations, should
be taken to represent a lake on any sampling visit;
• Macroinvertebrate data must represent a full count, or at least an estimate, of all
individual animals in the sample identified to Mixed Taxon Level;
• Owing to the relative paucity of macroinvertebrate data collected in Autumn that could be
related to the full required suite of water chemistry data, WFD60 focuses on Spring
sampled macroinvertebrates only. The Spring season was deemed to extend into June
for sites in northern Scotland. Samples were thus to be collected between February and
the first 10 days of June.
• Owing to considerable problems with the quality of water chemistry data provided by the
environment agencies, the range of essential chemical determinands (in the final
classification exercise) was restricted to the following:
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
17
pH
dissolved organic carbon (DOC)
calcium (Ca2+)
conductivity
ANC (determined either by ion balance or the Cantrell method depending on the
availability of constituent data)
(Cantrell ANC (µeq l-1) = Gran Alkalinity (µeq l-1) + f*DOC (mg l-1) (where f = 4.5 for
samples with pH < 5.5, and 5.0 when pH > 5.5)
• Sufficient water samples were required to allow the estimate of annual mean chemistry
over any one year period, within one year (prior to or after) the collection of the
macroinvertebrate sample. A series of water samples taken prior to the collection of the
macroinvertebrate sample was preferred. Again, due to the paucity of data of acceptable
quality, as few as three samples were accepted for the estimation of an annual mean.
Where five or more samples were available these had to be distributed approximately
evenly within the course of one year.
Despite relatively modest compliance requirements our final dataset, comprising Spring
sampled macroinvertebrate data and matching mean annual water chemistry, consisted of only
107 sites. Due to concerns that this rather small number of samples might restrict model
development we compiled a second dataset based on Spring sampled water chemistry only. In
this second dataset we included sites represented by one Spring water chemistry sample only
although if more data were available within this season then a mean value was determined.
This resulted in a small increase in the number of sites to 120. However, preliminary data
analysis suggested relatively poor relationships between the macroinvertebrate assemblages
and water chemistry, possibly due to problems presented by short-term variability in water
chemistry. Consequently we were unable to develop this further.
5.3 The WFD60 database
The data used in this project are stored in a Microsoft Access relational database housed at
ECRC-ENSIS. Due to several concerns with water chemistry data quality from different sources,
the database is built around the available macroinvertebrate samples. Chemistry data (for the
determinands listed in Section 5.2) are only included in the chemistry data tables provided there
are sufficient measurements to meet the annual mean estimate requirements for specific
macroinvertebrate samples in the database. All macroinvertebrate data generated through
WFD60 contract variations are included whether or not there is sufficient supporting water
chemistry. Macroinvertebrate samples which do not have sufficient supporting chemistry are
used at the end of the project to test the WFD60 tool (with respect to geographic distribution of
lake classes). The database also includes tables providing information on macroinvertebrate
and water chemistry samples (e.g. provenance, sample date, etc.), a table detailing geographic
information on sites, a “species dictionary” which relates species names to macroinvertebrate
Furse codes, and a series of Access queries enabling the determination of annual average
water chemistry, selection of appropriate macroinvertebrate sample data and the generation of
biological summary statistics.
5.4 The interim dataset
At the onset of the WFD60 project we explored the relationship between macroinvertebrate
community structure for a wider range of physico-chemical variables than were available for
later stages of the project. These data were drawn from ECRC-ENSIS data holdings and
comprised 38 sites (described from now as the Interim dataset). The macroinvertebrate samples
for these sites were all taken during Spring. While these data are of high quality, the relatively
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
18
low number of samples restricts the potential power of the resulting analyses. Details of the
samples included in the Interim dataset are provided in Appendix 1.
5.4.1 Indirect ordination
All multivariate analyses were conducted using the vegan ecological statistics package
(Oksanen et al., 2007) in the R statistical software package (R Core Development Team, 2007).
Macroinvertebrate data, in the form of raw counts, were first ordinated by detrended
correspondence analysis (DCA). The gradient length was approximately 3.0 indicating that
unimodal rather than linear techniques were most appropriate for subsequent data analysis.
Correspondence Analysis (CA) revealed three major outliers, Burnmoor Tarn, Llyn Llagi and
Llyn Cwellyn, resulting from the occurrence of small numbers of individuals of a limited number
of taxa which were found at no other sites. Since these had a disproportionate influence on the
ordination the sites were removed from this analysis. A CA ordination plot of site scores for the
modified dataset is presented in Figure 5.1. This shows a satisfactory distribution of sites across
the first two CA Axes. High elevation sites, such as Lochnagar (NAG), Scoat Tarn (SCOATT)
and Llyn Glas (GLAS) cluster in the upper left of the plot but otherwise there is no broader
indication of an influence of altitude on the ordination of sites on these axes.
Figure 5.1 Correspondence Analysis (CA) of macroinvertebrate assemblages for 35 acid-sensitive
UK lakes
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
19
5.4.2 Direct ordination with chemical variables
The macroinvertebrate data for the 35 remaining lakes were then subjected to Canonical
Correspondence Analysis (CCA) with the chemical parameters listed in Table 5.1 available as
explanatory variables. CCA derives a set of ordination axis scores for species and samples. For
the first axis, species scores and sample scores are chosen to maximise the correlation
between them. Scores on subsequent axes are also maximally correlated, but uncorrelated with
species and sample scores of the previously derived axis. In the following analyses all chemical
data were standardised.
First, CCA was performed for each chemical variable individually, to determine the maximum
amount of variance each could explain, regardless of potential covariant effects.
Table 5.1. Variance of the 35 lake macroinvertebrate dataset explained by chemical variables
applied individually in Canonical Correspondence Analysis (CCA). P-value determined by Monte-
Carlo permutation test.
Variable % total variance
explained
p-value (1000 permutations)
H+ (hplus) 7.91 0.017
Alkalinity (alk) 5.73 0.019
Conductivity (cond) 3.66 0.284
calcium (Ca) 6.66 0.001
magnesium (Mg) 4.08 0.210
potassium (K) 3.80 0.235
nitrate (NO3) 6.83 <0.001
sulphate (SO4) 3.29 0.337
labile inorganic aluminium (labileAl) 9.01 0.006
dissolved organic carbon (DOC) 8.48 <0.001
ion-balance ANC (ionANC) 7.82 <0.001
Hydrogen ion, alkalinity, calcium, nitrate, Allab, DOC and ion-balance ANC all showed significant
(p<0.05) relationships with the species data. DOC, ion-balance ANC and nitrate concentration
were most highly significant, while Allab, followed by DOC, explained the largest amount of the
variance. Ion-balance ANC and hydrogen ion concentration explained very similar amounts of
variance (approximately 8%).
CCA was then repeated to include all the above as explanatory variables. In this analysis the
variables explained 47 % of the total variance in the species data. 22.4 % of the total variance
was accounted for in the first two axes of the ordination (see Table 5.2).
Table 5.2 Summary statistics for CCA of macroinvertebrate assemblages for 35 acid-sensitive UK
lakes. Mean squared contingency coefficient = 3.528.
CCA Axes 1 2 3 4
Eigen values 0.4543
0.3372
0.2139
0.1809
Cumulative variance of
species data 0.1288
0.2244
0.2850
0.3363
Biplot scores (correlations between ordination axes and environmental variables) are provided
in Table 5.3. The first axis was related predominantly to DOC, while there was also some
interaction with hydrogen ion and Allab. The second axis appeared to represent the main acidity
axis with a particularly strong correlation with ion-balance ANC in addition to hydrogen ion and
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
20
Allab. Nitrate was also strongly correlated with this axis, suggesting that this anion, rather than
sulphate, may exert the more important anthropogenic acidity effect on this small dataset.
Table 5.3 Biplot scores for chemical variables and ordination axes for the CCA of
macroinvertebrate assemblages for 35 acid-sensitive UK lakes
CCA1
CCA2
CCA3
CCA4
H+ 0.43
-0.68
-0.05
-0.37
alkalinity -0.09
0.67
0.17
0.25
conductivity 0.19
0.31
0.05
-0.19
Calcium 0.27
0.71
-0.05
0.01
magnesium 0.03
0.46
0.07
-0.22
potassium -0.03
0.45
0.14
-0.19
Nitrate 0.02
-0.66
-0.46
0.36
sulphate 0.19
-0.03
-0.33
-0.13
labile inorganic aluminium 0.48
-0.76
0.00
-0.18
dissolved organic carbon 0.70
0.38
-0.02
-0.23
ion balance ANC -0.05
0.86
0.13
-0.19
Figure 5.2 CCA Ordination plot for Axes 1 and 2 of the macroinvertebrate – water chemistry
dataset for 35 acid-sensitive UK lakes
−2 −1 0 1 2
−2 −1 0 1
CCA1
CCA2
ACH
ARR
BLU
CADH
CHAM
CHN
CLAI
CLYD
CORN
CREI
DCAL
DOI
DUBH
EDNO
FEOI
FHIO
GLAS
GLOY
HIR
IRD
LACH
LGR
LNEI
LOCH
LOD
LOSG
ME04B
MYN
NAG
NEUN
RLGH
SCOATT
TINK
VNG9402
hplus
alk
cond
Ca
Mg
K
NO3
SO4
labileAl
DOC
ionANC
−1 0 1
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
21
Species scores, which describe the extent of the correlation of individual species with the
ordination axes, are given in Table 5.4. Several species showed relatively extreme scores on
both axes 1 and 2, possibly representing influences of organic and mineral acidity respectively.
The stoneflies Leuctra inermis and Nemoura sp., (known for their acid tolerance) showed strong
positive scores on Axis 1 (associated with high DOC) and strong negative scores on axis 2
(associated with low ANC). The same relationships were observed for the caddisflies Oecetis
ochracea and Anabolia nervosa and the beetle Oulimnius tuberculatus. The corixids Sigara
dorsalis and Callicorixa wollastoni, on the other hand, only showed strong (negative) scores
with Axis 2 indicating these species are most abundant in relatively clear acidified systems. The
abundance of these predators may be indicative of the absence of fish (perhaps lost through
acidification) in these lakes
In contrast to the observations above, few species showed both strong negative Axis 1 scores
and strong positive Axis 2 scores. Species with associations with low DOC sites included
several which are commonly associated with “non-acidified” systems, including the mayflies,
Ecdyonurus sp. and Siphlonuridae, and the limpet Ancyclus sp.. Species associated with high
ANC lakes included the acid sensitive stonefly Isoperla grammatica and the mayfly family
Baetidae.
This analysis on a small but high quality interim dataset therefore provided evidence for:
i) strong acidity controls on the macroinvertebrate dataset;
ii) differentiation between species indicative of acidified rather than naturally acid lakes;
iii) Allab as a potentially important indicator of species composition;
iv) ANC to be at least as powerful a predictor of macroinvertebrate assemblage
structure as hydrogen ion concentration (or pH).
Table 5.4 Macroinvertebrate species scores for the CCA of macroinvertebrate assemblages for 35
acid-sensitive UK lakes. Data are sorted by the score on the second CCA axis, deemed to be the
dominant acidity axis and particularly strongly correlated with ANC.
Taxon
Axis 1 Axis 2 Axis 3 Axis 4
Oecetis ochracea 1.77 -2.30 -0.24 -0.95
Oulimnius tuberculatus 1.70 -2.14 -0.17 -0.87
Agrypnia obsoleta 1.40 -1.93 -0.09 -0.64
Callicorixa wollastoni 0.53 -1.92 0.41 -0.16
Anabolia nervosa 1.53 -1.78 -0.18 -0.61
Deronectes griseostriatus 0.80 -1.77 0.40 -0.42
Leuctra inermis 1.62 -1.76 -0.21 -0.54
Agabus chalconotus -0.03 -1.71 1.21 0.25
Sigara dorsalis -0.03 -1.71 1.21 0.25
Nemoura sp. 1.31 -1.60 -0.12 -0.35
Agabus bipustulatus -0.68 -1.59 0.46 0.55
Capnia -1.57 -1.08 -2.48 1.15
Oreodytes davisii -1.57 -1.08 -2.48 1.15
Baetis sp. -0.32 -0.98 2.92 2.03
Centroptilum pennulatum -0.32 -0.98 2.92 2.03
Rhithrogena semicolorata -0.32 -0.98 2.92 2.03
Hydroporus palustris -0.67 -0.87 0.27 0.26
Diura bicaudata -1.46 -0.80 -1.24 0.84
Halesus -0.24 -0.76 0.19 0.32
Leuctra hippopus -0.58 -0.74 -0.05 0.55
Glaenocorisa propinqua 0.67 -0.66 0.17 0.73
Agrypnia varia 0.67 -0.62 0.11 -0.05
Annelida (Oligochaeta) 1.04 -0.58 0.61 0.09
Chaetopteryx villosa -0.58 -0.57 0.94 -0.38
Cordulegaster boltonii -0.86 -0.57 0.26 -1.17
Chaoboridae -0.34 -0.57 0.54 -0.03
Polycentropus sp. 0.12 -0.55 0.94 0.05
Limnephilus sp. 0.90 -0.54 0.28 0.54
Oreodytes sanmarkii -1.56 -0.54 1.21 -0.63
Chloroperla -1.14 -0.48 0.17 0.73
Hesperocorixa castanea -0.20 -0.47 -0.39 1.18
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
22
Ancylus sp. -1.54 -0.46 -0.43 0.64
Asellus aquaticus -1.54 -0.46 -0.43 0.64
Ecdyonurus sp. -1.54 -0.46 -0.43 0.64
Potamophylax latipennis -1.54 -0.46 -0.43 0.64
Plectrocnemia sp. 0.25 -0.46 0.22 0.65
Empididae -0.88 -0.43 1.19 0.29
Chironomidae -0.68 -0.36 -0.07 0.03
Tipulidae -0.53 -0.36 0.19 0.45
Cyrnus -0.24 -0.29 0.78 0.13
Siphlonuridae -1.41 -0.29 1.64 0.17
Halesus radiatus -1.17 -0.28 0.67 -0.42
Cordulia 1.35 -0.22 0.32 1.45
Erpobdella octoculata 1.35 -0.22 0.32 1.45
Mesophylax impunctatus 1.35 -0.22 0.32 1.45
Aeshna -0.43 -0.18 0.17 -0.20
Nemurella picteti -0.65 -0.17 -0.41 0.21
Agabus arcticus -0.60 -0.15 0.21 -0.21
Hygrotus novemlineatis 0.17 -0.14 -0.43 1.29
Arctocorisa germari 0.96 -0.12 -0.30 1.02
Polycentropodidae -1.06 -0.11 -0.13 0.24
Amphinemura sulcicollis -0.15 -0.11 1.52 -0.13
Limnephilidae -0.69 -0.11 -0.05 0.06
Simuliidae -0.31 -0.11 -0.73 0.11
Holocentropus sp. 1.28 -0.10 0.19 1.46
Plectrocnemia conspersa -0.94 -0.04 -0.05 -0.05
Haliplidae -1.01 -0.04 0.72 0.32
Tubificidae -0.87 -0.03 1.62 0.00
Centroptilum luteolum -0.37 -0.01 1.38 0.62
Enchytraeidae -0.98 -0.01 1.44 0.27
Deronectes depressus 0.84 0.00 -0.03 1.15
Dicranota sp. -1.02 0.02 0.18 -0.17
Capnia bifrons -0.68 0.02 0.60 0.53
Sigara scotti 1.21 0.03 0.05 1.47
Brachycentrus subnubilus 1.11 0.03 0.89 0.49
Lymnaea peregra -0.29 0.04 0.63 0.08
Pisidium sp. 0.20 0.05 0.41 -0.03
Chloroperla torrentium -0.76 0.05 0.01 0.17
Psychomyiidae -0.95 0.05 -0.61 -0.19
Siphlonurus lacustris -0.20 0.07 1.28 0.40
Collembola -0.86 0.08 -0.45 -1.41
Cordulegasteridae -0.86 0.08 -0.45 -1.41
Triturus sp. -0.86 0.08 -0.45 -1.41
Ceratopogonidae -0.07 0.09 0.69 0.28
Nemoura cambrica -0.37 0.13 0.32 -0.60
Deronectes -0.98 0.13 0.55 -0.71
Caenis moesta -0.04 0.14 -0.47 1.18
Leuctra nigra 0.53 0.16 0.03 0.96
Plecoptera -0.05 0.19 0.17 0.22
Agrypnia picta -0.56 0.26 -0.20 -1.02
Pyrrhosoma nymphula 1.06 0.28 -0.23 1.49
Hydracarina -0.65 0.28 0.36 -0.25
Anisoptera sp. -0.21 0.29 0.27 -0.54
Diptera -0.46 0.30 0.33 -0.35
Cymatatia bonsdorffi 1.04 0.32 -0.26 1.50
Elmis aenea -0.53 0.33 1.88 0.03
Erythromma -0.53 0.33 1.88 0.03
Rhabdiopteryx acuminata -0.53 0.33 1.88 0.03
Siphlonurus armatus -0.53 0.33 1.88 0.03
Trichoptera -0.23 0.33 0.22 0.09
Glossiphoniidae -0.11 0.34 0.43 0.31
Siphlonurus sp. -0.52 0.34 1.83 0.01
Mystacides sp. 0.85 0.34 0.84 0.25
Dytiscidae -0.30 0.35 0.53 -0.31
Agrypnia sp. -0.23 0.37 0.13 -0.60
Leptophlebiidae 0.52 0.37 -0.16 0.22
Polycentropus flavomaculatus -0.47 0.38 -0.08 -0.23
Limnephilus lanatus -0.53 0.41 1.53 -0.18
Lepidostoma hirtum -0.48 0.42 1.54 -0.32
Coleoptera -0.32 0.43 0.42 -0.60
Naididae -0.10 0.45 0.21 -0.19
Oulimnius sp. -0.16 0.46 0.64 -0.33
Athripsodes sp. 0.69 0.49 -0.37 0.97
Leptophlebia marginata -0.42 0.49 -0.34 -1.22
Zygoptera -0.37 0.50 -0.48 -1.43
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
23
Sialis lutaria 0.43 0.50 -0.21 0.19
Heptageniidae -0.14 0.51 1.11 -0.11
Lumbriculidae -0.04 0.53 0.46 -0.20
Callicorixa praeusta 0.92 0.53 -0.50 1.52
Deronectes assimilis 0.92 0.53 -0.50 1.52
Gyrinus aeratus 0.92 0.53 -0.50 1.52
Hesperocorixa sahlbergi 0.92 0.53 -0.50 1.52
Ischnura elegans 0.92 0.53 -0.50 1.52
Sigara distincta 0.92 0.53 -0.50 1.52
Sigara nigrolineata 0.92 0.53 -0.50 1.52
Sigara nimitata 0.92 0.53 -0.50 1.52
Enallagma cyathigerum 0.13 0.54 -0.18 -0.40
Coenagrionidae -0.41 0.57 1.18 -0.10
Cyrnus insolutus -0.36 0.57 -0.80 -1.02
Caenis horaria -0.09 0.57 1.07 -0.37
Gyrinus caspius 1.14 0.58 1.28 0.31
Sialis fuliginosa 1.14 0.58 1.28 0.31
Stylodrilus heringianus -0.19 0.60 0.94 -0.42
Sericostoma personatum 0.32 0.61 0.57 0.06
Limnius volckmari -0.31 0.62 0.35 -0.23
Nemouridae 0.13 0.63 0.56 -1.08
Cyrnus flavidus -0.13 0.63 -0.55 -1.31
Erpobdellidae -0.49 0.64 -0.27 -1.11
Leptophlebia vespertina 0.02 0.65 -0.23 -0.91
Plectrocnemia geniculata 0.28 0.66 -0.43 -0.21
Cyrnus trimaculatus 0.18 0.66 0.41 -0.39
Mystacides azurea 0.26 0.67 0.90 -0.01
Tinodes waeneri -0.41 0.68 -0.21 0.14
Pericoma -0.10 0.70 1.88 -0.19
Nematoda 0.19 0.72 0.00 -0.57
Helobdella Stagnalis 0.31 0.72 0.20 -0.02
Stylaria lacustris 0.12 0.73 -0.09 -0.86
Corixidae -0.27 0.78 0.07 -0.68
Mystacides longicornis 0.35 0.82 0.26 -0.78
Polycentropus kingi 0.35 0.82 0.26 -0.78
Sympetrum 0.35 0.82 0.26 -0.78
Leuctra sp. 0.64 0.86 0.61 0.20
Baetis rhodani 0.64 0.88 -0.43 -0.38
Ernodes articularis 0.64 0.88 -0.43 -0.38
Gammarus pulex 0.64 0.88 -0.43 -0.38
Tabanoidea 0.64 0.88 -0.43 -0.38
Ameletus inopinatus -0.37 0.91 0.10 0.40
Rantus exsoletus 0.92 0.92 -0.18 0.15
Gyrinidae 0.28 0.92 0.09 -1.17
Elminthidae 0.43 1.01 -1.66 -0.83
Dreissenidae -0.16 1.04 -0.21 -0.35
Leuctra fusca -0.16 1.04 -0.21 -0.35
Sympetrum nigrescens -0.16 1.04 -0.21 -0.35
Tabanidae -0.16 1.04 -0.21 -0.35
Ephemeroptera 0.09 1.05 -0.17 0.35
Lumbriculus variegatus 0.92 1.11 -0.02 -0.54
Rantus bistriatus 0.92 1.11 -0.02 -0.54
Haliplus obliquus 0.29 1.19 -0.52 0.01
Oxyethira sp. 0.11 1.23 -0.34 0.15
Micropterna sp. 0.59 1.24 -0.22 -0.22
Baetidae -0.11 1.30 -0.38 0.37
Isoperla grammatica -0.06 1.49 -0.60 0.41
5.4.3 Species distributions and labile inorganic aluminium
The relationship between species distributions and Allab is of particular interest in WFD60 since
Allab is known to be highly toxic to many aquatic animals, while its presence in toxic
concentrations is highly indicative of acidification (see Section 3.3). Unfortunately Allab is not
routinely measured by the UK environment agencies and we are therefore only able to examine
relationships for this restricted dataset, enhanced by the inclusion of data for eight extra sites
provided by the Fisheries Research Services Freshwater Laboratory, Pitlochry. Figure 5.3
provides some indication of a relationship between Allab concentration and species richness (as
determined by the total number of species identified to species level). Of the nine sites with an
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
24
annual average concentration of more than 25 µg l-1 none contain more than 10 identifiable
species. Conversely, the seven sites with 14 or more species all have Allab concentrations of
less than 20 µg l-1. Clearly, however, low species richness is also common in a range of lakes
with low Allab concentrations.
Figure 5.3 The relationship between labile inorganic aluminium concentration and the number of
species identified to species level. Lines represent a fitted GAM model.
0 50 100 150
5 10 15 20
Labile Al (µg L−1)
No. of Taxa
Relationships between individual taxa and Allab are provided in Figure 5.4, with respect to
presence absence distributions and probability of occurrence (as modelled by the GAM curves).
Several taxa provide an indication of acute Allab sensitivity. Unfortunately the size of the dataset
is too restrictive to draw firm conclusions, but overall these plots are consistent with the
hypothesis that Allab may exert a strong influence over species distributions in acidic lakes.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
25
Figure 5.4 Presence/absence of the more common taxa in the interim dataset in relation to labile
inorganic aluminium concentration. Curves illustrate GAM functions for presence/absence (i.e.
probability of occurrence); dotted lines indicate 95% confidence intervals. Species names
provided in Appendix 2
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0 1.2
CHIRODAE
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8
DYTISCID
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0 1.2
HALE.SUS
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8
HYDRACAR
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8
LEPT.VES
Labile Al
0 50 100 150
0 1 2 3 4
LEPTOPAE
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0
LIMNEPAE
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0
LUMCLDAE
Labile Al
0 50 100 150
0.0 0.1 0.2 0.3 0.4 0.5
MYST.AZU
Labile Al
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
26
Figure 5.4 continued
0 50 100 150
0.0 0.2 0.4 0.6 0.8
TIPULDAE
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0 1.2
MYST.DES
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0 1.2
NEMO.URA
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8
PISI.IUM
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6
TINO.WAE
Labile Al
0 50 100 150
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
CORIXIDA
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0
DIPTERA
Labile Al
0 50 100 150
0 1 2 3
EMPIDIDA
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0
TRICLADI
Labile Al
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
27
Figure 5.4 continued
0 50 100 150
0.0 0.1 0.2 0.3 0.4 0.5
CERATOPO
Labile Al
0 50 100 150
0.0 0.1 0.2 0.3 0.4 0.5 0.6
LEPT.MAR
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0
SIPH.LAC
Labile Al
0 50 100 150
0.0 0.2 0.4 0.6 0.8 1.0
SIAL.LUT
Labile Al
0 50 100 150
0.0 0.1 0.2 0.3 0.4 0.5
ENAL.CYA
Labile Al
0 50 100 150
0.0 0.1 0.2 0.3 0.4 0.5
NEMO.CAM
Labile Al
0 50 100 150
0.0 0.1 0.2 0.3 0.4 0.5 0.6
TUBIFDAE
Labile Al
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
28
5.5 Exploratory Analysis of the full WFD60 dataset
5.5.1 Site ordination
Figure 5.5 is a non-metric multidimensional scaling (NMDS) ordination plot of the 105 sites in
the WFD60 training set based on the species chord distances (i.e. dissimilarities). Overlain on
these plots are fitted surfaces representing contours for gradients of ANC, Ca2+, DOC and pH,
based on an additive model which relates NMDS axes scores to these variables. These plots
illustrate a generally satisfactory distribution of sites across the key acidity gradients and
obvious relationships between species dissimilarity and acidity. However, the plots also
emphasise the particularly poor coverage of more acidic sites.
Figure 5.5 Non-metric multidimensional scaling (NMDS) ordination plots of the 105 sites in the
WFD60 database, based on macroinvertebrate species chord distances. Hence sites with the most
similar species assemblages lie closest together. Contours represent gradients of ANC, Ca2+, DOC
and pH. Sites names are represented by WBID codes.
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
ANC
NMDS1
NMDS2
31104
29153
21790
21723
46279
22125
99999
38409
18209
34400
35578
27927
14293
35561
46232
14202
38390
27872
27808
27900
24745
20712 17379
20633
29000
11424
5350
29215
22395
24892
27912
38394
8266
29021
36405
12606
29062
22223
34987
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
Calcium
NMDS1
NMDS2
31104
29153
21790
21723
46279
22125
99999
38409
18209
34400
35578
27927
14293
35561
46232
14202
38390
27872
27808
27900
24745
20712 17379
20633
29000
11424
5350
29215
22395
24892
27912
38394
8266
29021
36405
12606
29062
22223
34987
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
DOC
NMDS1
NMDS2
31104
29153
21790
21723
46279
22125
99999
38409
18209
34400
35578
27927
14293
35561
46232
14202
38390
27872
27808
27900
24745
20712 17379
20633
29000
11424
5350
29215
22395
24892
27912
38394
8266
29021
36405
12606
29062
22223
34987
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
pH
NMDS1
NMDS2
31104
29153
21790
21723
46279
22125
99999
38409
18209
34400
35578
27927
14293
35561
46232
14202
38390
27872
27808
27900
24745
20712 17379
20633
29000
11424
5350
29215
22395
24892
27912
38394
8266
29021
36405
12606
29062
22223
34987
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
29
While sites with a mean pH below 6 are common, most sites in the dataset have an ANC above
60 µeq l-1. Figure 5.6 is based on the same NMDS scaling but has macroinvertebrate weighted
average site scores superimposed. The plots show that the most acidic sites are dominated by
taxa such as the Hemipteran Arctocorisa germari, the Dytiscid beetle Agabus arcticus and the
Corixid Glaenocorisa propinqua, while those at the other extreme are represented by the
Dytiscid beetle Oreodytes septentrionalis, the caddis species Agraylea multipunctata and
Tinodes waeneri. Unfortunately most taxa names are hidden from view in this plot due to size
limitations.
5.5.2 Macroinvertebrate distributions on ANC and Ca2+ gradients
The presence/absence of taxa which occur in 10 or more sites (full dataset), across gradients of
ANC, and Ca2+ (the two key variables for the WFD60 physico-chemical classification) are
presented in Figures 5.7 and 5.8 respectively. The two plots show many similarities, which is
not surprising given the strong covariance between ANC and Ca2+ in the full dataset. As in
Figure 7.4 a number of taxa show very clear distributions. The majority of species show a rise in
probability of occurrence with increasing ANC, but some, such as Amphinemura sulicollis
(AMPH.SUL), Oulimnius sp. (OULI.IUS) and Hydracarina sp. ((HYDRACAR) appear to group
toward the middle of the gradient and are found neither in very acid or high ANC waters. The
stonefly Nemoura cambrica (NEMO.CAM) is one of very few species to be confined to the most
acidic sites only.
Unsurprisingly, given the tendency for an increase in the probability of occurrence of individual
taxa, species richness, as determined by the number of species identified to species level, also
shows a very marked relationship with ANC and Ca2+ (Figures 5.9 – 5.10). Beneath an ANC of
40 µeq l-1 most sites contain no more than ten defined species, whereas above 60 µeq l-1 the
vast majority of sites contain more than ten. Most sites with an ANC < 10 µeq l-1 contain six
defined species or fewer. Whereas the relationship with ANC appears stepped, the relationship
with Ca2+ is more linear, if more scattered, on most sections of the gradient.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
30
Figure 5.6 Non-metric multidimensional scaling (NMDS) ordination plots of the 105 sites in
the WFD60 database, based on macroinvertebrate species chord distances. Species positions
represent a weighted average of the scores of the sites in which they occur.
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
ANC
NMDS1
NMDS2
ARCT.GER
OREO.DAV
AGAB.BIP
HYDROPHL
LEST.SPO
CORD.LIA
AGAB.ARC
PLEC.MIA CALL.WOL
LIMN.ORA PARA.CNC
PHOX.PHO
CALL.PRA
GLAE.PRO
ILYB.FUL
OREO.SEP
BRAC.SUB
STIC.MUL
PEDICIID
TRIT.RUS
DUGE.SIA
CYRN.INS
AGRA.MUL
GYRI.AER
ISCH.URA
CYRN.FLA
AMEL.INO
SIGA.SCO
TIPU.LA
DICR.OTA
GYRI.NUS
DIPTERA
SIGA.DIS
PISI.IUM
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
Calcium
NMDS1
NMDS2
ARCT.GER
OREO.DAV
AGAB.BIP
HYDROPHL
LEST.SPO
CORD.LIA
AGAB.ARC
PLEC.MIA CALL.WOL
LIMN.ORA PARA.CNC
PHOX.PHO
CALL.PRA
GLAE.PRO
ILYB.FUL
OREO.SEP
BRAC.SUB
STIC.MUL
PEDICIID
TRIT.RUS
DUGE.SIA
CYRN.INS
AGRA.MUL
GYRI.AER
ISCH.URA
CYRN.FLA
AMEL.INO
SIGA.SCO
TIPU.LA
DICR.OTA
GYRI.NUS
DIPTERA
SIGA.DIS
PISI.IUM
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
DOC
NMDS1
NMDS2
ARCT.GER
OREO.DAV
AGAB.BIP
HYDROPHL
LEST.SPO
CORD.LIA
AGAB.ARC
PLEC.MIA CALL.WOL
LIMN.ORA PARA.CNC
PHOX.PHO
CALL.PRA
GLAE.PRO
ILYB.FUL
OREO.SEP
BRAC.SUB
STIC.MUL
PEDICIID
TRIT.RUS
DUGE.SIA
CYRN.INS
AGRA.MUL
GYRI.AER
ISCH.URA
CYRN.FLA
AMEL.INO
SIGA.SCO
TIPU.LA
DICR.OTA
GYRI.NUS
DIPTERA
SIGA.DIS
PISI.IUM
−0.4 −0.2 0.0 0.2 0.4 0.6
−0.4 −0.2 0.0 0.2 0.4
pH
NMDS1
NMDS2
ARCT.GER
OREO.DAV
AGAB.BIP
HYDROPHL
LEST.SPO
CORD.LIA
AGAB.ARC
PLEC.MIA CALL.WOL
LIMN.ORA PARA.CNC
PHOX.PHO
CALL.PRA
GLAE.PRO
ILYB.FUL
OREO.SEP
BRAC.SUB
STIC.MUL
PEDICIID
TRIT.RUS
DUGE.SIA
CYRN.INS
AGRA.MUL
GYRI.AER
ISCH.URA
CYRN.FLA
AMEL.INO
SIGA.SCO
TIPU.LA
DICR.OTA
GYRI.NUS
DIPTERA
SIGA.DIS
PISI.IUM
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
31
Figure 5.7 The presence/absence of taxa which occur in 10 or more sites (full dataset), across the
ANC gradient (in µeq l-1). Black line represents a GAM function (Poisson error distribution and
logit function).
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
CHIRODAE
ANC
−50 0 50 100 150 200 250
0.00 0.10 0.20 0.30
DICR.OTA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
DYTISCID
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
HALE.SUS
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
HYDRACAR
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
LEPT.VES
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6
LEPTOPAE
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
LIMNEPAE
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
LUMCLDAE
ANC
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
32
Figure 5.7 continued
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
MYST.AZU
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
NEMATODA
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
OLIGOCHA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
OULI.IUS
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
PLEC.CON
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
POLY.FLA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
POLYCENT
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
RADI.PER
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
SERI.PER
ANC
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
33
Figure 5.7 continued
−50 0 50 100 150 200 250
0.00 0.10 0.20 0.30
AMEL.INO
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
CAEN.LUC
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
ELEC.LAT
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
HYDR.ILA
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
LIMN.VOL
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
OULI.TUB
ANC
−50 0 50 100 150 200 250
0.00 0.05 0.10 0.15 0.20 0.25
OXYE.IRA
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
SIPH.TOR
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
TIPU.LA
ANC
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
34
Figure 5.7 continued
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
TIPULDAE
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
CENT.LUT
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
LIMN.LUS
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
MYST.DES
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6
NEMO.URA
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
PISI.IUM
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
SIGA.SCO
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
TINO.WAE
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
ANAB.NER
ANC
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
35
Figure 5.7 continued
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
BAETIDAE
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
CORIXIDA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
DIPTERA
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
EMPIDIDA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
HELO.STA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3
ISOP.GRA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
LEUC.TRA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
TRICLADI
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
CERATOPO
ANC
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
36
Figure 5.7 continued
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
LEPT.MAR
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
LEPT.BIA
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
LIMN.LUN
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
LIMN.VIT
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
NEMO.CIN
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
SIPH.LAC
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
CYRN.FLA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
SIAL.LUT
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6
ATHR.CIN
ANC
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
37
Figure 5.7 continued
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
ATHR.DES
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
CAEN.HOR
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
CHAE.VIL
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6
CYRN.TRI
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
DIUR.BIC
ANC
−50 0 50 100 150 200 250
0.0 0.2 0.4 0.6
LEPI.HIR
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
POLY.PUS
ANC
−50 0 50 100 150 200 250
0.00 0.10 0.20 0.30
ENAL.CYA
ANC
−50 0 50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
NEMO.CAM
ANC
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
38
Figure 5.8 The presence/absence of taxa, which occur in 10 or more sites (full dataset), across
the calcium gradient (in µeq l-1). Black line represents a GAM function (Poisson error distribution
and logit function).
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
CHIRODAE
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
DICR.OTA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
DYTISCID
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
HALE.SUS
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
HYDRACAR
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
LEPT.VES
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6
LEPTOPAE
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
LIMNEPAE
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
LUMCLDAE
Ca
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
39
Figure 5.8 continued
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
MYST.AZU
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
NEMATODA
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
OLIGOCHA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
OULI.IUS
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
PLEC.CON
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
POLY.FLA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
POLYCENT
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
RADI.PER
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
SERI.PER
Ca
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
40
Figure 5.8 continued
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
AMEL.INO
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0 1.2
CAEN.LUC
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
ELEC.LAT
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
HYDR.ILA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
LIMN.VOL
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
OULI.TUB
Ca
50 100 150 200 250
0.00 0.05 0.10 0.15 0.20 0.25 0.30
OXYE.IRA
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
SIPH.TOR
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
TIPU.LA
Ca
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
41
Figure 5.8 continued
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
TIPULDAE
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
CENT.LUT
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3
LIMN.LUS
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
MYST.DES
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
NEMO.URA
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
PISI.IUM
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
SIGA.SCO
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0 1.2
TINO.WAE
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
ANAB.NER
Ca
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
42
Figure 5.8 continued
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
BAETIDAE
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
CORIXIDA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
DIPTERA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
EMPIDIDA
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
HELO.STA
Ca
50 100 150 200 250
0.00 0.10 0.20 0.30
ISOP.GRA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
LEUC.TRA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
TRICLADI
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
CERATOPO
Ca
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
43
Figure 5.8 continued
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
LEPT.MAR
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
LEPT.BIA
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
LIMN.LUN
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
LIMN.VIT
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
NEMO.CIN
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
SIPH.LAC
Ca
50 100 150 200 250
0.00 0.10 0.20 0.30
CYRN.FLA
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
SIAL.LUT
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8 1.0
ATHR.CIN
Ca
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
44
Figure 5.8 continued
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
ATHR.DES
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
CAEN.HOR
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
CHAE.VIL
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
CYRN.TRI
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5 0.6
DIUR.BIC
Ca
50 100 150 200 250
0.0 0.2 0.4 0.6 0.8
LEPI.HIR
Ca
50 100 150 200 250
0.00 0.10 0.20 0.30
POLY.PUS
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4 0.5
ENAL.CYA
Ca
50 100 150 200 250
0.0 0.1 0.2 0.3 0.4
NEMO.CAM
Ca
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
45
Figure 5.9 The number of taxa identified to species level related to mean ANC for the 105
sites in the WFD60 training set. Lines represent a GAM function (Poisson error distribution and
log-link function.
−50 0 50 100 150 200 250
5 10 15 20 25 30 35
ANC (µeq L−1)
No. of Taxa
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
46
Figure 5.10 The number of taxa identified to species level related to mean Ca2+ for the 105 sites
in the WFD60 training set. Lines represent a GAM function (Poisson error distribution and log-link
function.
50 100 150 200 250
5 10 15 20 25 30 35
Calcium (µeq L−1)
No. of Taxa
5.5.3 Summary of exploratory analysis
This analysis shows that macroinvertebrate structure can be tightly linked to ANC. The fact that
the majority of taxa show an increasing probability of occurrence with rising ANC results in a
very strong gradient in species richness. However this gradient shows a clear stepped
distribution and is steepest between 0 – 60 µeq l-1. The overall pattern in macroinvertebrate
species diversity indicates that “High” biological status (according to WFD normative definitions)
might be the norm for sites with an ANC greater than 60 µeq l-1, whereas the abrupt reduction in
diversity as ANC approaches zero suggests that the fauna of sites with an ANC of less than
zero will be in a condition that could be described as between “Poor” and “Bad”.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
47
6 WFD TOOL DESIGN PRINCIPLES
6.1 Other WFD Schemes under development
Several WFD classification tools are currently being developed in the UK under the supervision
of UKTAG. The design methodology for most of these is similar and involves four main steps.
1) Reference sites (i.e. a sub-set of those assumed to be of High status) are identified
for each typology, using “expert judgment”.
2) Biological assemblages in a “training set”, including those for reference sites, are
related to a physico-chemical pressure gradient, such as phosphorus concentration,
using multivariate ordination methods such as canonical correspondence analysis
(CCA).
3) Sample scores derived from the ordination procedure are divided by sample scores
for reference lakes within the same typology to provide an ecological quality ratio or
EQR for each lake.
4) EQRs are then related to biological normative definitions, such as the relationship
between stress tolerant and intolerant species, and this is used to divide up the
gradient into the five WFD classes introduced in Section 1.
Class membership is then subjected to uncertainty analysis to ascertain the likelihood that a
biological sample will be allocated the appropriate damage class, given the known susceptibility
of the sample data to variability in sampling effort, time, space, etc..
Thus EQR derivation follows WFD guidance outlined in Annex 5 in a very literary manner, i.e.:
“..the results …..shall be expressed as ecological quality ratios for the purposes of classification
of ecological status. These ratios shall represent the relationship between the values of the
biological parameters observed for a given body of surface water and the values for these
parameters in the reference conditions applicable to that body.”
However, there is no explicit requirement in the Directive that an EQR must be calculated
mathematically by dividing sample scores in the way outlined above. We argue that this clause
may alternatively be interpreted as a qualitative requirement that the EQR must be based on
comparisons of biological condition of a site with what is considered to be reference condition.
If certain biological characteristics of high status can be considered universal, then any
deviation from these characteristics may be used to derive an EQR score.
We argue that the commonly adopted procedure is prone to a priori uncertainties which are not
subjected to rigorous analysis. First, no two lake ecosystems are identical now or in the past,
however WFD “reference lakes” tend to be few. Uncertainties arise immediately, therefore, with
regard to representativity of these lakes and hence the relative level of damage for lakes within
the same typology for which the same reference condition is used. While a continuous EQR
score is generated there are no grounds to believe a lake which has been allocated an EQR of,
for example, 0.5 is less damaged than one with an EQR of 0.4, even within the same typology.
The apparent continuity of the score, therefore, has limited ecological merit with respect to
damage assessment.
Second, the procedure by which the EQR gradient is divided into damage classes is often
based on subjective criteria such as the cross-over point between two biological indicator
classes (themselves defined by relating to “pressure gradients” used in previous steps), rather
than a mechanistic understanding of how the pressure is likely to influence the biological
community.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
48
Finally, this cross-over point, whilst being the optimal discriminator between the two groups of
biological indicators, may not be an optimal decision threshold for discriminating between site
damage classes. Again, the use of the cross-over point to fix the Good-Moderate boundary is
not required by Annex V of the Directive and we argue that the widespread use of this criterion
for setting this important boundary results from an overly prescriptive adoption of statements in
the WFD. The cross-over point is used solely to describe what the Commission meant by
“moderate” change over reference, not that this will necessarily be adopted as part of member
states classification schemes. Overall, therefore, we feel this process is self-referential and of
restricted ecological validity.
6.2 Classification under WFD60
We have proposed an alternative approach to lake classification to that discussed above. This
is based on the following observations outlined in previous sections of this report:
a) In contrast to other pressures of concern to the WFD, lakes of “reference condition” may
be particularly difficult to identify within the same biogeographic region; the most
appropriate “reference conditions” for acidified lakes in north Wales, the Pennines and
the English Lake District, may only be found in unacidified parts of the far north-west of
Scotland, but these may not be sufficiently analogous to lakes much further south for
climatic and geological reasons;
b) unlike other pressures of concern to the WFD the pressure of acidification can be
predicted from current physico-chemistry;
c) high status, according to physico-chemical normative definitions, can be identified with
some confidence on the basis of ANC;
d) a physico-chemical “good-moderate” boundary may be defined according to our
understanding of the importance of Ca2+ in determining the likely ANC threshold for
biological damage through acidification (based on physico-chemical and
palaeoecological models); this threshold can be considered to be the point at which
biological communities begin to differ “moderately” from reference and where major
taxonomic groups are first likely to disappear, according to WFD normative definitions;
e) “poor” to “bad” status may also be defined on the basis of biological toxicity thresholds
for aluminium, and relationships between Allab and ANC our data demonstrate that many
species and taxonomic groups typical of acid sensitive lakes are excluded from lakes
with an ANC below zero and the dominant reason for this is likely to be due to the
coincidence of this threshold with substantially elevated levels of Allab in addition to low
pH.
We are confident, therefore, that we can classify lakes in our training set using physico-
chemistry in a manner that accords with biological normative definitions and evidence of the
degree of departure of biological communities from reference state. In this report we go on to
show how such classes can be predicted by the macroinvertebrate community, and how this
can then be used as the basis for the WFD60 tool.
6.3 WFD60 proposed classification approach
Rather than generate EQRs for individual sites using assumptions on appropriate reference
sites, and then attempt to divide the EQR gradient according to normative definitions into the
five WFD classes, we set out to test whether it was possible to use our understanding of the
relationship between physico-chemical indicators of damage and aquatic biology to derive a
priori EQR compatible classes for sites, and then use a classification approach to predict
membership of a class based on the macroinvertebrate characteristics of each site.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
49
For example, according to our preliminary analysis of the data and information available in the
scientific literature, it is reasonable to assume that the macroinvertebrate community of sites
with an annual mean ANC of >60 µeq l-1 are unlikely to differ significantly from reference
condition (with respect to acidification). Providing there are no other major environmental
constraints: taxonomic composition should correspond totally or nearly totally to unacidified
conditions; there should be no sign of alteration in the ratio of sensitive to insensitive taxa; and,
there should be no sign of any reduction in diversity from that found in acidified sites. The EQR
of such a site should approach a value of 1, and could, for the sake of convenience, be
allocated a score of 0.9.
Conversely, the macroinvertebrate community of any site with an ANC of < -50 µeq l-1, is highly
likely to exhibit very low pH and highly toxic concentrations of Allab. Such a site is highly likely to
support a very limited number of highly acid-tolerant taxa only and will deviate profoundly from
reference condition with respect to taxonomic composition, abundance, ratios of sensitive to
insensitive taxa and diversity. Its EQR must therefore approach zero and could again, for
convenience, be allocated a score of 0.1. If we can derive classification rules which can
determine the likelihood of a biological assemblage falling into these classes then we have the
basis for a robust classification system which is compatible with WFD requirements.
6.3.1 Generation of a “damage matrix”
We used our understanding of the relationships between current ANC and calcium
concentrations, and Allab concentrations, predictions of the extent of acidification, and our
understanding of critical biological thresholds, to derive a damage matrix with which any site
could be classified. This is presented in Table 6.1. Each of the 107 sites in the WFD60
database was therefore assigned a class according to these categories.
Table 6.1. Damage matrix, based on understanding of relationships between ANC and calcium
concentrations and evidence from palaeoecological and hydrochemical models of acidification,
and contemporary relationships with Allab and macroinvertebrate assemblage characteristics.
Letters represent expert judgement on likely ecological status with respect to damage from
acidification. H = High, G = Good, M = Moderate, P = Poor and B = Bad.
ANC group (µeq/l)
Ca2+
(µeq/l)
MGHH
80-100
MHHH
60-80
BBBPMGHHH
40-60
BBBPMGHHH
20-40
BBPMGHHHH
0-20
-20-(-40)-10-(-20)0-(-10)10-020-1040-2060-4080-60100-80
ANC group (µeq/l)
Ca2+
(µeq/l)
MGHH
80-100
MHHH
60-80
BBBPMGHHH
40-60
BBBPMGHHH
20-40
BBPMGHHHH
0-20
-20-(-40)-10-(-20)0-(-10)10-020-1040-2060-4080-60100-80
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
50
Following preliminary data analysis, the paucity of samples in classes “Poor” or “Bad” resulted
in modelling problems. The training data set was heavily biased to classes “High” and “Good”
(50% and 30% of samples respectively). Moderate sites comprised just over 11% of the
samples with the remaining c. 9% of samples in Poor or Bad status. This is illustrated in Figure
6.1, where there a very few samples in the areas coloured red (Bad), orange (Poor) or yellow
(Moderate).
To work around this problem, classes “Poor” and “Bad” were merged into one “Poor-Bad” class.
Also, as a result of our relative lack of confidence we had in defining the “High-Good” boundary
it was also decided to merge these two classes. Consequently our classification would be based
on fitting to only three status classes.
Figure 6.1 Distribution of sites according to ANC and Ca2+ concentration. Diagram shaded
according to the damage matrix (Table 6.1).
−20 0 20 40 60 80 100 150 200 250
20 40 60 80 100 150 200 250
ANC (µeq L−1)
Calcium (µeq L−1)
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
51
6.3.2 Macroinvertebrate input data
Macroinvertebrate data were processed for subsequent analysis in 6.3.3 as follows:
Total abundances were first calculated for each Mixed Taxon Level taxa for each sample date
for each site. The bulked data thus represented differing sample sizes depending on the origin
of the samples. In order to account for this, species data were converted to proportions of the
full species count.
In addition, the following summary data were collated for each bulked sample:
1) The minimum possible number of species present (described as minimum species richness
or MSR; i.e. all species plus genus and family level groups where there are no higher
order members present) within the following groups);
The whole assemblage
Trichoptera
Ephemeroptera
Ephemeroptera not of the family Leptophlebiidae (known to be an acid tolerant family)
Plecoptera
Odonata
Hemiptera
Coleoptera
3) Total number of individuals within the following groups (identified to any level) expressed as a
proportion of the total number of individuals in the sample;
Trichoptera
Ephemeroptera
Plecoptera
Trichoptera + Ephemeroptera + Plecoptera
Odonata
Hemiptera
Coleoptera
Tricladida
Chironomidae
6.3.3 Decision trees
Decision trees are popular in many fields as a way of encapsulating and structuring the
knowledge of experts for use by the less experienced. They are commonly used in botany and
medical decision making for example. Automatic tree construction was first developed in the
social sciences, but the work of Breiman and colleagues in the late 1970's and early 1980's
(encapsulated in their monograph; Breiman et al., 1984) placed tree-based models firmly within
a statistical framework. Since then, the properties of tree-based classifiers have been well
studied and are widely regarded as being a powerful tool for supervised classification purposes.
Recent advances such as bagging (Breiman 1996), boosting (Freund and Schapire, 1997) and
random forests (Breiman, 2001) have been developed that extend the tree concept to so-called
ensemble methods to improve predictions from trees, but do so at the expense of simplicity and,
to some extent, interpretability.
Tree-based methods partition the “feature space” into a set of regions and then fit a simple
model, such as a constant one, in each one. Tree-based models are computationally intensive
methods that are ideally suited to situations where there are many explanatory variables to
choose from and it is not known a priori which of them are most important. The main virtues of
tree-based models are that they are that they are excellent for initial data inspection, they give a
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
52
clear picture of the structure of the data and they provide a highly intuitive insight into the kinds
of interactions between variables.
Models are fitted using binary recursive partitioning, where the data are successively split along
features of the environmental data so that at any node the split which maximally distinguishes
the response variable in the left and right branches is selected. Splitting continues until the
nodes are “pure”, i.e. comprising one class only, or the data are too sparse.
Where the response variable is a factor (i.e. grouped in classes), the tree is known as a
classification tree. Where the response is continuous, the tree is known as a regression tree.
Because the recursive partitioning continues until pure nodes are reached or the samples in
each node are too sparse, there is a danger of over fitting the response. Tree-based models are
generally pruned back to a minimal, adequate model. This is done via a cross-validation
procedure to obtain “honest” estimates of the true prediction error. Plotting this prediction error
against tree-size allows the selection of the tree with the minimum error. An alternative is to
select, as the best tree, the smallest tree whose estimated error rate is within one standard
deviation of the minimum error rate. A simple introduction to the use of classification and
regression trees in ecological data analysis is given by De'ath and Fabricius (2000).
In this report classification trees were used to predict a class status for High-Good, Moderate
and Poor/Bad from the macroinvertebrate training set data:
Trees were fitted in the R computer software (Version 2.5.1; R Core Development Team 2007)
using the rpart package by Therneau and Atkinson (2007). Splits were determined via
minimising the Gini index measure of node impurity:
D
i
=1−
∑
k
p
ik
2
where pik is the observed proportion of class k within node i, and Di is the node impurity for node
i. The total measure of node impurity is then:
∑
Di
Trees were fitted to their maximal extent and then a 10-fold cross-validation (CV) procedure
was used to identify the smallest size of tree within 1 standard error of the tree with the lowest
cross-validated misclassification error. In 10-fold CV, the training data are randomly assigned to
one of ten groups. In turn, each group is excluded from the CV training set whilst the remaining
nine groups are used to grow an “unpruned” tree. This tree is then used to predict the class
membership of the samples in the group left out. This is repeated for each group in turn. At
each of the 10 stages, the misclassification error is computed for each tree of size n, n = 1, ...,
m, m = number of samples. The average error across the 10 CV stages is used as a measure of
tree performance, with an associated standard error.
Predictions from the tree are governed by the terminal nodes or leaves of the tree. Predictions
are based on a “majority rules” concept, whereby the predicted class for a target sample is
determined by the most abundant class of the node the target sample ends up in. The
proportion of samples classified into a particular node can also be used as an indicator of class
probability; a target sample ending up in a node containing samples mainly of class k will have a
high probability of belonging to class k. Conversely, a target sample ending up in a node with a
mixture of classes will have a lower probability of belonging to the majority class.
SNIFFER WFD60: Macroinvertebrate Classification Diagnostic Tool August 2007
53
Due to the uneven sampling of sites within each class and a desire to minimise “High” or “Good”
status sites being classified as less than “Moderate” or worse (and vice versa), two additional
parameters were used to grow the tree.
The prior probabilities of class membership are one of the important parameters controlling tree
growth. If these are unspecified, the prior probabilities are taken as the number of samples in
each class expressed as a proportion of the total number of samples in the training set. Given
the unbalanced structure of our dataset, simply by assuming that any site was in High-Good
class one would be correct over 50% of the time – a considerably higher probability than
randomly guessing site class. In data sets that are heavily biased to one or a few classes, the
priors will be biased towards well-represented classes and as such misclassification rates can
be minimised by concentrating on correctly assigning the well-represented classes at the
expense of the poorly-represented ones. This will often result in low misclassification rates for
the well represented classes but high misclassification rates for the poorly represented classes.
In effect, the classification tree will be a poor predictor of class for the poorly-represented
classes.
To balance weights of each class in the fitting algorithm, we defined the prior probability of
belonging to a given class to be 0.33 for all three classes. This means that, a priori, any sample
has an equal chance of belonging to any of the three classes.
Variable misclassification costs can be used by supplying the fitting routine with a loss matrix.
This loss matrix describes the relative “costs” of classifying a site of a known class to a different
class. By specifying the relative misclassification costs in a loss matrix, we can control the fitting
algorithm by penalising certain types of misclassification more strongly than others. For
example, in the case of WFD classification, classifying a “Poor-Bad” site as “Moderate” may be
considered less costly than classifying it as “High-Good”.
The loss matrix used was:
site class prediction
High-Good Moderate Poor-Bad
High-Good 0 3 3
Moderate 3 0 3
Poor-Bad 3 1 0
The diagonal elements of the loss matrix are equal to 0, indicating that there is no
misclassification cost associated with correctly classifying a sample. The off-diagonal elements
of the loss matrix describe the relative misclassification costs. Here we have said that
classifying a “High-Good” site as “Moderate” (or vice versa) incurs a relative misclassification
cost of 3 – this is the default misclassification cost in the rpart fitting algorithm. The value 3 was
chosen simply to add additional penalty to this ty