Reliable Biometrical Analysis

in Biodiversity Information Systems

Martin Drauschke

, Volker Steinhage

, Artur Pogoda de la Vega

1??

Stephan M

uller

, Tiago Maur

ıcio Francoy

2? ? ?

and Dieter Wittmann

Institute of Computer Science, University of Bonn

omerstr. 164, 53117 Bonn, Germany

Institute of Crop Science and Resource Conservation, University of Bonn

Melbweg 42, 53127 Bonn, Germany

Abstract. The conservation and sustainable utilization of global biodiversity ne-

cessitates the mapping and assessment of the current status and the risk of loss of

biodiversity as well as the continual monitoring of biodiversity. These demands in

turn require the reliable identiﬁcation and comparison of animal and plant species

or even subspecies. We have developed the Automated Bee Identiﬁcation System,

which has not only been successfully deployed in several countries, but also sup-

ports taxonomical research as part of the Entomological Data Information Sys-

tem.

Within this framework our paper presents two contributions. Firstly, we explain

how we employ a model-driven extraction of polymorphic features to derive a

rich and reliable set of complementary morphological features. Thereby, we em-

phasize new improvements of the reliable extraction of region-induced and point-

induced features. Secondly, we present how this approach is employed to derive

new important results in biodiversity research.

1 The Role of Identiﬁcation in Biodiversity Research

Genera and species with ecological or economical use are a very relevant topic in bio-

diversity research. The bees, Apoidea, are of special interest, because about 70% of the

world’s food crops are pollinated by them. The annual value of this pollination service

has been reported to be about 65 billion US-$, cf. [6]. Since there is an alarming ex-

tinction of crop pollinating bees due to diseases, parasites, and pests, as well as due to

human interventions like clearing, pesticides, etc., bees are specially focused by biodi-

versity research.

Martin Drauschke is now with the Department of Photogrammetry, Institute of Geodesy and

Geoinformation, University of Bonn, Nussallee 15, 53115 Bonn, Germany

Artur Pogoda de la Vega is now with CA Computer Associates GmbH, Emanuel-Leutze-Str.

4, 40547 D

usseldorf, Germany

? ? ?

Tiago Maur

ıcio Francoy is now with de Gen

etica, Faculdade de Medicina de Ribeir

ao Preto,

USP, Av. Bandeirantes, 3900, 14049-900 Ribeir

ao Preto, SP, Brazil, tfrancoy@rge.fmrp.usp.br

Drauschke M., Steinhage V., Pogoda de la Vega A., Stephan Müller S., Maurício Francoy T. and Wittmann D. (2007).

Reliable Biometrical Analysis in Biodiversity Information Systems.

In Proceedings of the 7th International Workshop on Pattern Recognition in Information Systems, pages 25-36

DOI: 10.5220/0002430300250036

 SciTePress

Due to the worldwide lack of bee exerts (less than 50 experts, so-called bee taxonomists,

worldwide) on the one hand and the urgency of monitoring and protecting endangered

pollinating bee species, the Automated Bee Identiﬁcation System ABIS has been devel-

oped. After training, ABIS performs a fully automated analysis of images taken from

the forewings of bees. This fully automated analysis is the outanding quality of ABIS

compared with all other approaches, which demand for intensive interactive processing

of each specimen, e.g. [2] or [7]. Compared with these interactive approaches ABIS

has been reported to be at least 50 times more effective since there is no need to adjust

the wings into an exact position of the camera ﬁeld. The fully automated approach pro-

vides together with the ABIS database an information system for the surveillance and

management of biometrical data of one of the most important pollinating group world-

wide. It has been successfully deployed for monitoring purposes in Germany, Brazil and

the United States, cf. [12]. ABIS is also part of Entomological Data Information Sys-

tem (EDIS), a co-operative research project jointly undertaken by a number of leading

German natural history museums and university institutions in the area of biodiversity

informatics.

Fig. 1. Left: classiﬁcation on demand: Bee experts use ABIS during their excursions. Middle:

input image of an Africanized bee. Due to the image aquisition, the scale of all wings is in a

small range. The wing’s position and orientation does not impact the feature extraction. Right:

detected cells and vein net of wing image.

The success of ABIS on identifying bee species leads now to a new challenge: the iden-

tiﬁcation within the next taxonmical levels, i.e. subspecies and races, which are far more

similar and therefore more difﬁcult to separate. To face this difﬁcult task of subspecies

identiﬁcation, we improve the extraction algorithms of morphological features, espe-

cially of the region-induced and point-induced features, i.e. cells and vein junctions.

We demonstrate our improved approach within a fascinating taxonomical (oder biodi-

versity) research project on the so-called killer bees. In 1957, swarms of the African

honey bee Apis mellifera scutellata accidentally escaped near Rio Claro, Brazil, and

initiated a series of crosses and backcrosses with the pre-introduced European honey

bee races, cf. [5].

This hybridization process resulted in the Africanized honey bee, which is better adapted

to the tropics and broadly more productive than the European bees. Nevertheless, many

beekeepers abandoned their activities in Brazil at the beginning of the Africanization

process, because these bees presented undesirable traits like a high swarming rate and a

strongly defensive behaviour (which led to their popular, but actually wrong name killer

bees). In the meanwhile, the beekeepers have learnt how to work with these aggressive

bees. Thus, the Africanized honey bee has spread throughout the (sub)-tropic Americas.

However, the evolution of the genetic and morphometric proﬁles of the bees from the

beginning of the Africanization process in Brazil till now is still unknown. The results

of comparing these proﬁles are not only important in taxonomic terms of biodiversity

research, but may also show important practical implications for pollination manage-

ment and beekeeping.

2 Robust Extraction of Reliable Features

2.1 Basic Steps

The traditional and manual classiﬁcation approach of bees by taxonomists uses mor-

phological characteristics of their whole body. If a subset of these features enables

an automated, efﬁcient recognition process, we call it a ﬁngerprint. We derive such

a ﬁngerprint from a highly resolved digital image of a bee’s wing as shown in Fig. 1.

The images of the wings typically show under backlight illumination light regions, the

transparent cells, which are surrounded by a dark net: the opaque veins and their junc-

tions. We call the polymorphical features, i. e. the region-induced cells, the line-induced

veins, and the point-induced junctions, primary features. From these primary features,

we compute a set of derived features as listed in Table 1. The set of derived features

forms one part of the ﬁngerprint. On basis of the extracted features, we normalize the

position and orientation of the wing and determine selected grayvalues for the second

part of the ﬁngerprint. The combination of both feature types, the derived features and

the grayvalues, returns the best classiﬁcation results, cf. [9].

Table 1. List of derived features of the ﬁn-

gerprint.

derived from feature

cells area, circumference,

distance of centers

(absolute and relative values),

eigenvalues and eccentricity

veins length of curvature

junctions distances and angles (along

adjacent cells)

Fig. 2. Demonstration of clusters formed by

ﬁngerprints. Therefore, each feature vector

is projected into a 2D-plane. The samples

of the same subspecies of honey bees are

colored identically.

The expert knowledge is inserted into ABIS within a training step, where genus speciﬁc

wing models have to be learnt under supervision of a taxonomist. The wing models con-

tain the expected cell contours including their variance as well as the expected positions

of the vein junctions. Then, ABIS performs a fully automated feature extraction which

consists of the following steps:

1. edge detection using gradient information and

2. cell extraction based on grouping of cell border segments until an initial substruc-

ture of cells could be veriﬁed by the system, cf. section 2.2,

3. adjusting the trained wing model to the present wing by an afﬁne transformation T ,

4. model-driven cell extraction for difﬁcult cells and

5. precise location of the vein junctions and vein detection, cf. section 2.3,

6. derivation of rotational stable and translation invariant features from extracted cells

and vein net, cf. Tab. 1,

7. statistical analysis of feature vector for bee species classiﬁcation, cf. [9].

Before starting the classiﬁcation, we perform a PCA to reduce the dimension of our

feature vector and eliminate redundant information, i.e. if some features are highly cor-

related. Then, the classiﬁcation can be operated by various discrimination algorithms.

We gained the best results by using the kernel discriminant analyis as proposed by [8].

The integration of the expert knowledge will be done in an ofﬂine training step, which

is necessary only once for each classiﬁcation project. The following tasks must be done

manually within the training step:

1. naming of the extracted cells,

2. identiﬁcation and naming of the junction positions,

3. ﬁne tuning of classiﬁcation parameters.

Then, the identiﬁcation of bee species and subspecies is able to run in a fully automated

way.

In this section, we want to point out the improvements of our feature extraction. Firstly,

star-shaped polygons form an appropriate model of all possible cell shapes. This model

ensures a reliable cell detection, which is based on the grouping of extracted edges.

Secondly, we model the localization of the junctions more precisely. The junctions are

of special importance, because identiﬁcation by bee experts and by ABIS relays heavily

on the junction positions, i. e. many ﬁngerprint features are derived from them.

2.2 The Use of Star-Shaped Regions for Cell Extraction

Since the border of the cells appear much darker than their inner points, the cell de-

tection starts with an edge detection algorithm, which is based on an approach to line

detection proposed by [11]. This approaches estimates the ﬁrst and second derivatives

of the image intensity by convolving the image with the derivatives of the Gaussian

smoothing kernel, since this smoothing kernel makes the inherently ill-posed problem

of estimating derivatives of noisy images well posed. In contrast to the line detection

task tackled by Stegers approach, the line-shaped veins of the wing images show lines

with highly variable width - even degenerating into region-shaped areas near the vein

junctions. Therefore, we modify the line detection approach into an edge detection

approach by searching for zero-crossings of the second derivative instead of the ﬁrst

derivative.

Generally, the detected edge pixels, so-called edgels, do not yield in continuous region

borders. Pseudo veins and disturbances due to hairs and polls on the wings lead to

locally positive false hypotheses of cell borders. Furthermore, mutations and low con-

trasts lead to locally negative false hypotheses of cell borders. Therefore, we developed

Fig. 3. Left: all detected edge pixels. Middle: all detected straight edges of a minimum length.

Right: all detected cell hypotheses.

a domain speciﬁc grouping strategy. Firstly, we determine all straight edgels of a mini-

mum length, secondly, we classify all other edgels into left and right curved edgels, cf.

Fig. 3.

If some straight edgels are directly connected to right curved edgels at their ends, we

join them to construct cell hypotheses for our cell detection. Our iterative grouping

strategy focuses on the gaps between the current hypotheses on the one hand and the

remaining edgels on the other hand, cf. Fig. 3.

Bee taxonomists can tell us, that the cells in the bee wings are convex or star-shaped.

Hence, we derive the following strategy. Thus, we call a set H of concatenated border

edgels a cell hypothesis, if these concatenated edgels are clockwise ordered around

their center, the mean of all edgels, and if they do not contain a cycle. H denotes the

closed cell hypothesis, where both endings of the cell hypothesis have been connected

by a straight line. The grouping algorithm enlarges each hypothesis H until the closed

border of a region is constructed by the edgels. Two constructive elements of the cell

hypotheses are of interest for the cell detection: the center c(H) and the kernel ker(H)

with

ker(H) =



p ∈ H | ∀x ∈ H : xp ∈ H



, (1)

where xp encodes all points on the straight line between x and p.

The grouping of the border edgels is comparable to a best ﬁrst search: the algorithm

is looking for the best aggregation with another cell hypothesis for closing the border

of the original cell hypothesis. If the best ﬁtting cell hypothesis does not close the

hypothesis, it must reduce the gap between its endings. The gap size is measured by

the angle α, which is deﬁned by the end point, the center and the start point of the cell

hypothesis.

More precisely, let H

be the current cell hypothesis, which has a gap size of α

. Then,

in the next iteration step, we determine the new cell hypothesis from H

by concatena-

tion with an other appropriate cell hypothesis H

, i.e. which has a nearly smooth transi-

tion to H

. If there is no such cell hypothesis that closes H

to a completely closed cell

border, we append the candidates to resulting hypotheses H

s+1

at either the beginning

or the ending of the start hypothesis, respectively:

s+1

= H

◦ B ◦ H

or H

s+1

= H

◦ B ◦ H

, (2)

where B encodes the bridging pixels, which form a smooth transition between H

and

. The gap sizes α

s+1

of the resulting hypotheses H

s+1

are determined with the old

center c(H

) of the start hypothesis, because we want to ensure a decreasing gap size.

We do not recompute the center c

s+1

until we have decided for the best ﬁtting cell

hypothesis for aggregation.

When the centers of the cell hypotheses lie in the kernel of the reconstructed cell, this

local inspection works very well. Then, the gap size α

decreases monotonously, since

all border points of the region are visible from the center, cf. the corollas on visibility

polygons discussed by [3]. For convex cells, the center of each hypothesis lies in the

kernel of the cell. Since several cells of a bee’s wing are only star-shaped, we need to

shift the center of a cell hypothesis iteratively with each hypothesis update towards the

also iteratively changing kernel of the cell. This task can only be done in a heuristic

way, combining the following two steps:

1. We compute the grayvalue gradient ∇g

c(H)

from the neighborhood of center c(H)

and shift c(H) in the direction of the maximum gradient, until a minimal distance

to the cell border is reached. Especially, this is necessary to move the centers away

from nearly straight hypotheses.

2. The shifted center (from step 1) gets replaced again, now by the mean of the center

itself and both hypothesis’s endings. This step has almost no impact on smaller,

straight hypotheses, but a big one on hypotheses which already form a big part of

a cell. Then, the center of such a hypothesis is pulled towards the opening of the

hypothesis and is now much closer to the center of the extracted cell.

Fig. 4. Five cell hypotheses are shown as black

lines, their centers are visualized by black dots.

Finally, these cell hypotheses should be ar-

ranged together to build the cell. To show the

importance of shifting the centers, we depict

the kernel of the ﬁnally reconstructed cell by

the gray ﬁlled polygon. The left image shows

the original centers, which are all outside the

kernel - the cell extraction fails in such a case.

The right image shows the centers of the same

cell hypotheses after their transformation - the

cell extraction performs well.

Fig. 5. Wing Model as created in the training

step. The average cell contours are shown by

the darker lines which are surrounded by the

minimum and maximum bounds of the cell

contours. The junctions are set manually into

the wing model.

We show an example in Figure 4, where we need to construct the cell shape from ﬁve

cell hypotheses. None of the centers of those hypotheses lies in the kernel of the cell

(left). Our closing procedure originally failed in these cases. Then we extended our

approach and improved the positions of the centers of the cell hypotheses (right). If

the centers were shifted towards the kernel, we were successful with reconstructing the

cell. Nevertheless, in almost all cases, there is a cell hypothesis, whose center lies in the

cell’s kernel after the shifts have been done.

We tested our algorithms on about 25,000 cells. Due to the improvements on closing in

the case of non-convex, but star-shaped cells, we are now able to detect all cells of a

bee’s wing in over 99.9% of all tested samples. Fig. 1 shows all detected cells of a wing

image.

The extracted cell borders of all specimens of the training set form the basis for com-

puting the wing model a genus or a group of genera. The wing model describes the

expected cell borders as well as their geometric variances as learnt from the traing set,

cf. Fig. 8. After training the wing model will be employed as a ”road map” for the ex-

traction of cell borders to identify the species or subspecies of speciemens of the trained

genus or group of genera.

2.3 Positioning of the Vein Junctions Using the Wing Template

Classical biometrical approaches show that the appropriate junction positioning is cru-

cial for the successful identiﬁcation, because a large number of ﬁngerprint features (i.e.

distances, lengths, angles) will be derived from the junction positions, cf. Table 1. Thus,

we have to place the junctions after the complete cell detection.

Therefore, during the training of a wing model, the taxonomist deﬁnes exactly the posi-

tions of the junctions into the model, cf. Fig. 5. These positionings will be aligned to the

wing under investigation by the afﬁne transformation T from step 3 of the feature ex-

traction process in section 2. If p is the position of the j-th junction in the wing model,

then p

= T p is the estimated position of the junction in the wing.

We estimate the parameters of the afﬁne transformation T from the centers of the cells

in the wing model and the centers of initially extracted cells. These initially extracted

cells are quite robustly to detect without the help of a wing model. They form the basis

for the model-based extraction of the remaining cells together with the generic wing

model. Intutionally, one might say that the inner cells of the wing image are the initially

extracted cells, while the outer cells are the cells, which are more difﬁcult to extract and

therefor we need the wing model. Although, this estimation proved to be locally precise

enough for our purpose, there is a need for readjusting the junction’s coordinates, as

seen in Figure 6. The ﬁnal position is shown in the right part of Figure 6.

Fig. 6. Left: inprecise localization of junctions by afﬁne transform. Right: ﬁnal positions of the

junctions after local rearrangements.

The localization of the junctions is improved with the use of explicit, spherical junction

templates, which the bee expert deﬁnes also in the wing model. These templates do not

only contain the predicted position of a junction, but also its local neighbourhood

(p) := { x | d(x, p) = ||x − p||

≤ r

} , (3)

and a analogously deﬁned uncertainty area of a junction U

(p). A junction’s environ-

ment shows the information about the adjacent cells and veins, see Figure 5, Fig. 6, and

Figure 7. Its uncertainty area encodes all valid positions for the junction.

Fig. 7. Visualization of junction templates: E

contains the adjacent cells in black. The infor-

mation of the veins is derived from the the extruded cells (dilation with a circle, marked in light

gray). Pixels outside of these extruded cells are unknown and marked in dark gray. The examples

show (f.l.t.r.) one, two or three adjacent cells. Thereby, the values of the radii (f.l.t.r. 17, 19 and

23 [pxl]) are chosen by the taxonomic expert and encode the expert knowledge about the relevant

local feature neighbourhood for placing the junctions.

With adjusting of the cells of the wing model to the investigated wing, we also transform

the junction templates. So, we work with afﬁne stable templates for deriving the optimal

positions of the junctions. The two radii r

and r

, which are chosen by the bee expert,

do not change much by this transformation, so we approximate r

by r

and r

by r

respectively.

Considering each pixel in the uncertainty area of a junction z ∈ U

(p) as a candi-

date for the junction’s position, we perform a template matching by determining the

weighted correlation between the environment E

(z) and the transformed junction

template. Thus, we shift the junction template such that its center maps to z. Since E

and the template have the same size, we are able to compare them pixel-wise. We deﬁne

the following three sets:

– C contains all pixels of E(z), which have been classiﬁed as points of cells in the

wing,

– R contains all pixels of E(z), which are marked as points of cells in the shifted

junction template,

– V contains all pixels of E(z), which are marked as points of veins in the shifted

junction template.

Then, we evaluate two functions ρ and ω which are deﬁned on E(z) by

ρ(x) :=







1 , if x ∈ R and x ∈ C

0 , if x /∈ R and x /∈ V

−1 , if x ∈ V and x ∈ C,

(4)

ω(x) :=



1 , if x ∈ V and x /∈ C

0 , else.

(5)

Fig. 8. Selectable derived features: The bee expert has inserted the three junctions a, b and c into

the genus model, e. g. the model for the honey bees. The positions of the three junctions in the

current wing have been placed automatically during the feature extraction. Various ﬁngerprint

features could be derived from these junctions. In a training step for the classiﬁcation, the bee

expert may select different sets of the derived features. Here, one may choose between two sets

of distances to perpendicular point x for subspecies identiﬁcation of Apis bees, ax and bx, or ab

and cx. The feature selection may vary between distinct bee genera.

A correspondence between the prediction of the junction template and the classiﬁcation

of the pixels as cells yields a positive response of ρ(x). Since we do not have any

information on the vein course, we can only consider wrongly classiﬁed pixels, where

the template predicts a vein instead of a part of the cell. Then, ρ(x) returns a negative

value for these pixels. The response of ρ(x) is zero, if no information of the pixel is

given by the template.

We obtain the similarity between E and the junction template by

s(z) =

x∈E(z)

I(x)

ρ(x) + I(x)

ω(x), (6)

where I(x) is the intensity value and I(x) is its inverted value. Thus, we combine the

intensity information of the image with ρ- and ω-responses of the template matching.

2.4 Vein Detection

When the detection of the cells succeeded and the placing of the junctions is well done,

the extraction of line-induced veins is relatively simple. As the detected cells and their

shapes have impact on the positional precision of the junctions, they together also in-

ﬂuence the peculiarity of the veins. So, the veins form the complementary space of the

other two feature types. Each vein starts from a junction, runs along the cell borders,

and it ends at a junction. Thus, the thinning approach, which was already proposed in

[9], combined with the idea to connect adjacent junctions around a cell, leads to a highly

reliable detection of the veins.

2.5 Selection of Reliable Fingerprint Features

The classiﬁcation of bees is based on the ﬁngerprint features, which we derive from the

primary features, i.e. the extracted cells, veins and junctions. Our approach to feature

selection is ﬂexible to cope on different opinions and approaches on relevant feature sets

for morphological identiﬁcation within the community of taxonomists. Therefore, we

are generally able to derive different ﬁngerprints from the same set of primary morpho-

logical features. Fig. 8 illustrates such a case where we allow the bee expert to derive

different feature sets concerning the honey bees Apis.

Fig. 9. Data quality inspection: result of the cell extraction of an image with a big pollen inter-

secting a vein. With help of the visualisation, errors like this can be found very easily.

Within the last years, ABIS proved to work successfully using the proposed ﬁngerprint

features as shown in Table 1. Furthermore, the ﬁngerprints are conﬁrmed by our recent

results in subspecies identiﬁcation.

3 Data Management and Visualization

All the wing images, the primary and derived features, the downsampled images, the

models, identiﬁcation results and georeferences are maintained with the help of the

ABIS database. As an user-friendly front-end of ABIS and the ABIS-Database, the-

AbisCommander, was developed for the efﬁcient and multimedia-based support of project-

related data management and quality inspection. With the help of AbisCommander, data

can organized easily by drag’n’drop-operations in different projects, thereby providing

the multiple usage of the same data by reference in different projects and roles (i.e.

training, identiﬁcation), cf. [10].

Visual inspection of input data as well as of all derived data (i.e. features and models)

is important, particularly during the training of wing models. Figure 9 depicts an ex-

tracted deformed cell region caused by a big pollen intersecting one of the veins. Using

this image for training of a wing model, will obviously corrupt the wing model and

therefore decrease classiﬁcation rates. Such errors can be found easily with help of the

visualization. Generally, AbisCommander can visualize the derived results of each step

in training and identiﬁcation. Generally, the detection of errors will improve the quality

and reliability of feature extraction and classiﬁcation.

For the implementation of ABIS and AbisCommander, we paid attention to use open

source software (Java, MySQL, Google Earth) only. This is important, to make ABIS

achievable without great ﬁnancial investments, which i.e. permits the use in locations

of the third world, where the expert knowledge stored in ABIS is needed mostly.

4 Results, Conclusion, and Outlook

We will now come back to our taxonomical prominent identiﬁcation project of sub-

species, i.e. the taxonomical research project on the so-called killer bees introduced

in section 1. We use the enhanced ABIS to compare the morphological proﬁle of the

Africanized honey bee population from 1965 with the corresponding proﬁles of the

most important ancestors populations and the actual one. We work with six groups:

Apis mellifera mellifera, Apis mellifera ligustica, Apis mellifera carnica, Apis mellifera

scutellata, Africanized Bees Ribeir

ao Preto (1965), Africanized Bees Ribeir

ao Preto

(2002).

The ﬁrst three groups represent the most important European ancestor subspecies, while

the fourth group represents the most important African ancestor subspecies. The ﬁfth

and sixth group represent Africanized honey bee populations from 1965 and 2002, re-

spectively. While ABIS was employed successfully for species identiﬁcation with about

5000 specimens, in this investigation in subspecies level about 100 individuals of each

group were available. After the training ABIS reveals a recognition rate of about 94%

(leave one out test), which is on the one hand “only” 94% due to the very close rela-

tionship of all six groups, but on the other hand sufﬁcient and clearly better than results

of other morphometric approaches.

Table 2. Left: confusion matrix of classiﬁcation with ABIS. Right: Mahalanobis square distances

between the cluster centroids in feature space. We considered the following subspecies of Apis

mellifera: carnica (A.c.), ligustica (A.l.), mellifera (A.m.), scutellata (A.s.), and the Africanized

honey bees from 1965 (A 65) and 2002 (A 02), resp.

A.c. A.l. A.m. A.s. A 65 A 02

A.c. 14.8 0.5 0.0 0.4 0.0 0.0

A.l. 2.5 3.3 0.0 0.0 0.0 0.0

A.m. 0.4 0.0 15.7 0.2 0.0 0.0

A.s. 0.0 0.0 0.2 21.4 0.5 0.0

A 65 0.0 0.0 0.0 0.0 15.8 0.5

A 02 0.0 0.0 0.0 0.2 0.5 23.1

A.c. A.l. A.m. A.s. A 65

A.l. 9.32

A.m. 34.54 29.68

A.s. 27.08 23.65 22.47

A 65 32.98 29.83 21.60 12.40

A 02 37.68 34.04 24.18 15.14 12.43

Fig. 10. Dendrogram showing the Mahalanobis square distances between the two Africanized

honey bee populations of 1965 and 2002 and the ancestor populations.

The tables in Tab. 2 show a more detailed analysis in form of a confusion matrix and

the Mahalanobis square distances between the cluster centroids in feature space, resp.

The very similar and very hardly separable subspecies of Apis m. carnica and Apis m.

ligustica yield the least distance in feature space and the most misclassiﬁcations. On the

one hand this missclassiﬁcation of about 6% will motivate future work on ABIS. On the

other hand the correct results on the other 94% of specimens allow a new taxonomical

interpretation on the basis of the Mahalanobis distances as follows, cf. Tab. 2, and Fig.

10 and [1]:

– The African honey bee A.m. scutellata contributed most to the Africanized popula-

tions.

– A.m. mellifera is the European subspecies that contributed the most to both African-

ized populations. This conﬁrms results from allozymes,

– The “mellifera” genes were probably added to the Africanized proﬁle only at the

beginning and spread together with the population.

– The 2002 Africanized population at Ribeir

ao Preto acquired a unique morphometric

proﬁle.

– Since 1965 its distance to the A.m. scutellata and the European ancestors has in-

creased.

The changes in wing venation indicate that the African honey bees continued to dif-

ferentiate from 1965 to 2002 with the effect that the actual populations of the African-

ized bees evolved a unique proﬁle which will keep on changing. The ﬁngerprint of the

Africanized honey bee, which we derived by morphometric analysis of the bee’s wings

using ABIS, shows a signiﬁcant difference to all the other introduced Apis subspecies

in Brazil. Our results are partially in agreement with results obtained on genetic mark-

ers, e. g. [4] and are initiating further research using genetic markers to conﬁrm our

morphometric results.

References

1. Francoy, T.M., Goncalves, L.S., Wittmann, D.: Changes in the Patterns of Wing Venation

of Africanized Honey Bees over Time. In: VII ENCONTRO SOBRE ABELHAS, 2006,

Ribeir

ao Preto. Anais do VII Encontro sobre Abelhas (2006) 173–177

2. Gauld, I.D., O’Neill M.A., Gaston, K.J.: Driving Miss Daisy: the performance of an auto-

mated insect identiﬁcation system. In: Austin, A.D., Dowton M. (eds.): Hymenoptera: Evo-

lution, Biodiversity and Biological Control. CSIRO, Canberra (2000) 303–312

3. Goodman, J., O’Rourke, J.: The Handbook of Discrete and Computational Geometry. CRC

Press LLC (1997) 467–479

4. Lobo, J.A., Krieger, H.: Maximum-Likelihood Estimates of Gene Frequences and Racial

Admixture in Apis melifera L. (Africanized Honeybees). Heredity, Vol. 68 (1992) 441–448

5. Perez-Castro, E.E., May-Itz

a, W.J., Quezada-Eu

an, J.J.G.: Thirty Years after: a Survey on

the Distribution and expansion of Africanized honey bees (Apis mellifera in Peru. J. Apic.

Res., Vol. 41 (2002) 69–73

6. Pimentel, D., Wilson, C., McCullum, C., Huang, R., Dwen, P. , Flack, J., Tran, Q., Salt-

man, T., Cliff, B.: Economics and environmental beneﬁts of biodiversity. Biological Sci-

ences, Vol. 47 11 (1997) 747–757

7. Rohlf, F.J.: tpsDig, digitize landmarks and outlines, version 2.0. Department of Ecology and

Evolution, State University of New York at Stony Brook (2004)

8. Roth, V., Steinhage, V.: Nonlinear Discriminant Analysis Using Kernel Functions. In: Ad-

vances on Neural Information Processing Systems, NIPS ’99, Denver (1999) 568–574

9. Roth, V., Pogoda, A., Steinhage, V., Schr

oder, S.: Integrating Feature-Based and Pixel-Based

Classiﬁcation for the Automated Identiﬁcation of Solitary Bees. In: Musterkennung, DAGM

’99, Bonn (1999) 120–129

10. Seim, D., M

uller, S., Drauschke, M., Steinhage, V., and S. Schr

oder: Management and Visu-

alization of Biodiversity Knowledge. In: Proc. 20th International Conference on Informatics

for Environmental Protection, EnviroInfo 2006, Graz, Shaker (2006) 429–432

11. Steger, C.: An Unbiased Detector of Curvilinear Structures. PAMI, Vol. 20 2 (1998) 113–125

12. Steinhage, V., Schr

oder, S., Roth, V., Cremers, A.B., Drescher, W., Wittmann, D.: The

Science of “Fingerprinting” Bees. In: German Research - Magazine of the Deutsche

Forschungsgesellschaft 1 (2006) 19–21