Content uploaded by Erik Sulanke
Author content
All content in this area was uploaded by Erik Sulanke on Jul 25, 2022
Content may be subject to copyright.
Edited by Erik Sulanke and Dr. Jörg Berkenhagen
Workshop organized and hosted by:
Thünen Institute of Sea Fisheries, Bremerhaven, Germany, part of
Johann Heinrich von Thünen Institute
Federal Research Institute for
Rural Areas, Forestry and Fisheries
28 – 30 March 2022
Online Workshop
REPORT
OF THE SECOND WORKSHOP ON AN
ALTERNATIVE APPROACH TO THE
SEGMENTATION OF FISHING FLEETS
Introduction
Our second workshop on a newly developed, alternative approach to the segmentation of shing
eets was held from 28th to 30th of March 2022. Due to travel restrictions caused by the global
Covid-19 pandemic, the workshop took place online. Cisco Webex provided the online platform.
The workshop was hosted by the Thünen Institute for Sea, Fisheries, part of the Johann Heinrich
von Thünen Institute, Germany's Federal Research Insitute for rural areas, forestry, and sheries.
Thirty-seven experts representing 16 nations, the ICES, and DG MARE, participated in the
workshop. Fifteen national sheries data sets were analyzed. The list of participants can be found
in Annex 1. Erik Sulanke and Jörg Berkenhagen chaired the workshop. The agenda can be found
in Annex 2.
Executive summary
A new approach to the segmentation of shing eets was developed in a DCF pilot project at the
Thünen Institute of Sea Fisheries and transferred to an R package, referred to as "FS-package" in
the following. In March 2021, the rst workshop on the novel approach and the FS-package was
held with 34 experts representing 15 nations, and major progress in improving the package was
made. After implementing the suggestions made by the attendants of the rst workshop and
identifying the next urgent steps for further developing the novel approach, the creators of the
approach organized a second workshop and formulated the following ToRs:
1. Harmonize the data preparation and develop a standardized protocol.
2. Clarify the application of the eet segmentation approach and evaluate newly developed
tools.
3. Establish regionally consistent eet segments over multiple member states operating in the
same shing regions.
Besides the mentioned ToRs, various technical issues needed to be addressed before the workshop,
the most signicant being the publication of the FS-package in the publically available repository
GitHub. Regarding ToR 1, newly established, standardized thresholds in a key function assigning
stocks to catch data as well as a novel function for automatic data preparation were well received
by the participants. However, an appropriate way to pre-separate eet data to improve the
clustering result was not agreed upon. The remaining issues include the grouping of demersal
seiners and demersal trawlers in the DCF, no clear agreement on appropriate length pre-
segmentation, and the possibility of a regional pre-segmentation, e.g., ICES and non-ICES area
fleets.
The workshop participants gave overall positive feedback on the additional new functions of the
FS-package. Especially the new functions making use of target assemblages were highlighted, as
they have the potential to foster comparability between economic and metiér-based biological eet
classication methods. The regional group work, which was specied in ToR3, comprised the
- 1 -
major work conducted during the workshop. A full list of the analyzed regions can be found in
Annex 4. Due to the composition of workshop participants, some regions were more well-
represented than others, i.e., a comprehensive analysis could be conducted for, e.g., the North Sea
and the Northeast-Atlantic, but not, e.g., for the Black Sea. Despite this imbalance in the analysis,
the overall feedback of the regional group analyses was positive, and regionally consistent eet
segments could be identified in all adequately represented regions. Nevertheless, more detailed,
extensive, and inclusive work in separate regional sessions is necessary for an adequate regional
analysis of shing eets.
In conclusion, we highlight the necessity of a standardized workow to classify and name eet
segments. In addition, an articial intelligence analysis approach developed in cooperation with the
technical university of Kaiserlautern signicantly improves the applicability of the FS-package. A
corresponding publication is in preparation, and the AI approach needs to be included in the terms
of reference of a future eet segmentation workshop. All considerations and discussions of
reforming the DCF eet segmentation procedure need to take into account the criteria of the
suitability of a eet segmentation procedure. We identied a close connection of segments to
sheries, a homogenous cost structure of the resulting eet segments, and the feasibility of the
approach as the most important.
In the following, we present the Terms of Reference.
Terms of Reference
The overarching aim of our workshop was to test the amendments made to the newly developed
approach to the segmentation of shing eets. This approach was created in a DFC pilot project
of the Thünen Institute between September 2019 and December 2020 and transferred to an R
package. After testing the approach in cooperation with various national partners, a workshop on
the novel approach was held from March 29th to 31st, 2021. The results of this rst workshop led
to the following ToRs:
1. Harmonize the data preparation and develop a standardized protocol.
2. Clarify the application of the eet segmentation approach and evaluate newly developed
tools.
3. Establish regionally consistent eet segments over multiple member states operating in the
same shing regions.
Background
Under the Data Collection Framework (DCF), economic data of European fishing eets have to
be provided by eet segment. The current eet segmentation scheme is based on the vessel length
class and the main shing gear, which are both technical parameters of the vessels. This
segmentation method is well dened and easily applicable, but it does not adequately represent
- 2 -
target sheries. In addition, the length classes used for the classication often fail to group vessels
of matching operational scales correctly. Vessels with similar technical parameters are often active
in varying sheries that differ in terms of catch composition, shing activity, and cost structure.
To improve reporting with respect to individual target sheries, a transferable, systematic approach
based on multivariate statistics methods was developed in the pilot project 'Fleet Segmentation' at
the Thünen Institute for Sea Fisheries in Bremerhaven, Germany, from September 2019 to
December 2020. The statistical framework was transferred to a user-friendly function package for
the R statisticssoftware, which was tested by multiple partners in the developmental process. From
the 29th to the 31st of March 2021, the rst workshop on the novel approach to eet segmentation
was held. The workshop was well attended by national sheries experts, and based on the
discussions, feedback, and suggestions gathered, the ToRs for the second eet segmentation
workshop were formulated.
Statistical Framework and Technical Amendments
A detailed description of the statistical background of the alternative eet segmentation approach
can be found in the report of the rst eet segmentation workshop. For an overview, please see
gure 1. This section will focus on the amendments made to the approach prior to the workshop.
First and foremost, the R package containing the novel approach was made available in a public
programming repository, GitHub. This is in accordance with one major conclusion of the rst
workshop. Publishing the package makes it available to all potential users, simplies the
implementation of updates, and creates a citable source for reports and publications. On the
technical side, rstly, the standard distance measure used in the clustering algorithm was changed
from the Euclidean distance to a metric conversion of the Bray-Curtis dissimilarity. This distance
measure is less sensitive to pairs of zeros in the data. It, therefore, creates more balanced clusters
and helps reduce the abundance of single-ship clusters, which tend to be a problem, especially in
large, heterogenic data sets.
As suggested in the rst workshop, a function allowing standardized data preparation was created.
The user no longer needs to create multiple specific data sets, only one comprehensive data set is
needed from which all necessary subsets of data are created. This data preparation function and
possibilities for improving it further will be discussed in the next section regarding ToR 1;
additional newly developed or updated functions aiding the applicability of the package will be
discussed in the section of ToR 2.
- 3 -
ToR 1. Harmonize the data preparation and develop a standardized protocol
The hurdles in installing the FleetSegmentation R package from a .zip le that were observed in
the rst eet segmentation workshop illustrated the necessity of having the package publicly
available. Hence, the FleetSegmentation R package was publically released on the GitHub
repository in August 2021. As a consequence, updating the package has become considerably easier.
However, installing a package from a non-quality checked repository like GitHub can be an
obstacle for inexperienced users. Therefore, publishing the FleetSegmentation package in the
official, quality-checked R repositoryCRAN (The Comprehensive R Archive Network) is the next
step in improving availability and ensuring a high code quality that meets scientic standards.
During the rst eet segmentation workshop, the ICES stock database used to create the distance
matrices for the clustering algorithm was extended by stock lists of the ICES and the GFCM. This
and the commonly agreed thresholds for automatic creation of species-area combinations
considerably improved the clustering result. Nevertheless, especially the stock list of the
Mediterranean Sea appears to be still incomplete and requires validation from experts of
Mediterranean Fisheries Management.
The function to automatically assign stocks to the national data sets was nested in a new function
that was created to harmonize the data preparation procedure. Thisnew function requires only one
comprehensive data frame, including all necessary eet variables (vessel ID, main gear, species,
area, catch weight), and computes all data frames for subsequent analysis per gear class. While this
step reduces the preparatory work and the potential for errors, the pre-segmentation by main gear
class requires further elaboration. In the second eet segmentation workshop, most participants
used the DCF main gear, as an alternative gear classication was not yet agreed upon. This DCF
main gear is assigned based on gear use time, and any gear being used in more than 50% of the
shing time is considered the main gear. This not only conceals patterns of polyvalent gear use but
also classies some very heterogenic shing gears, specically demersal trawlers and demersal
seiners (DTS). Based on our observations of shing patterns and cost structure, we highly
- 4 -
recommend separating demersal trawlers and demersal seiners in the DCF just like pelagic trawlers
(TM) and pelagic purse seiners (PS). Regarding polyvalency, we emphasize the necessity to consider
polyvalent gear classes like polyvalent active gears (MGP), a mix of active and passive gears (PMP).
Thresholds for assigning polyvalent gear use based on time of use need to be agreed upon by
the experts of the STECF.
Especially the large, heterogenic eets of major shing nations analyzed in the workshop elucidated
the necessity of additional pre-segmentation measures. In these cases, analyzing vessel groups
based only on the gear classes might lead to the grouping of vessels with very different technical
parameters and cost structures, especially when these vessels are targeting widely distributed sh
stocks. Participants tested various strategies of pre-separation during the workshop, e.g., size-based
pre-separation of passive geares over and under 12m vessel length or area-based pre-separation
between vessels that operate in non-ICES areas and those that do not. The different measures of
pre-separation are currently collected from the participants, and surveys on which of those
measures appear to be most promising will be held. It might be advisable that member states with
large, diverse eets introduce additional characteristics to further segment the eet without
initiating condentiality issues.
Conversely, it was suggested to group different active gears or passive gears where necessary in a
way that there is evidence for a comparable cost structure within that group. One idea was to test
grouping all passive gears, all active gears, and all dredgers. However, this has to be further
analyzed. Moreover, it was suggested to apply a pre-segmentation based on the technical equipment
of the vessels, e.g., net drums, beams, net hauler, line hauler, etc. It is intended to include several
pre-separation protocols in an updated version of the FleetSegmentation package.
ToR 2. Clarify the application of the eet segmentation approach and evaluate newly
developed tools
Both the newly developed and the updated functions were generally reviewed positively by the
participants, especially by those who were familiar with the FleetSegmentation package from the
rst workshop. As suggested in the rst workshop, the bar plot displaying the stock-based catch
composition of the clusters (clustering_stockshares_plot) was added with the option to display the
number of vessels in the corresponding clusters. This feature helps identify major shing strategies
and supports the user in decision-making on which clusters should be joined. The generic plots
displaying cluster catch composition and technical variables (clustering_stockshares_plot and
clustering_plotgrid, as well as all sub-plot functions included in the latter) were updated with an option
to subset the plot to a selection of clusters to give additional advice for cluster joining and segment
identication. This helps the user not only in getting an overview of the cluster features but also in
saving results in a more structured and clear way. In addition to those updates, the target
assemblage was introduced as a new feature for analysis. It is now possible to create a bar plot
(clustering_assemblageshares_plot) and a multi-dimensional scaling (clustering_assemblage_MDS) of the
target assemblages linked to the species caught by vessels. The assemblages help identify vessels
with similar shing strategies (e.g., vessels targeting small pelagic sh), especially if the user is not
familiar with all species codes and stock units. Since ICES alone considers 269 stock units and
shing eets generally target several hundred species, this is quite common. Including the
assemblages in the segmentation procedure has the additional advantage that resulting eet
- 5 -
segments are more comparable to the métier-based biological analysis, where the target assemblage
is included as a feature of the six-level métier denition. A complete list of the target assemblages
can be found in Annex 3.
ToR 3. Establish regionally consistent eet segments over multiple member states
operating in the same fishing regions
The primary focus of the second eet segmentation workshop was the regional group work to
identify regionally consistent shing strategies and, ultimately, also eet segments between multiple
shing nations. Major shing regions considered relevant for the participant nations were the Baltic
Sea, the Black Sea, the Eastern Arctic, the Mediterranean, the Northeast Atlantic, the North Sea,
and Other shing regions (e.g., NAFO areas or Southern Atlantic). Due to the fact that not all
European shing nations were represented in the workshop, some regions were more diversely
analyzed than others. All available regional plot sheets and participant case studies can be found in
Annexes 5.1-5.4. The regional group work in the Mediterranean was mainly characterized by the
comparison of the Maltese and Greek shing eets. For the Baltic Sea analysis, eets from
Germany, Finland, Sweden, Denmark, Lithuania, and Estonia were compared. In the North Sea
region, analysis was carried out for shing eets from Germany, Denmark, the Netherlands,
Sweden, and France. The Northeast Atlantic shing eets hailed from Portugal, Spain, France,
Ireland, and Germany. In all shing regions, consistent eet segments could be, to some extent,
identied among shing nations. Of course, identifying targeted and distinct fishing patterns like,
e.g., the Brown Shrimp shery in the North Sea is less challenging than comparing diverse,
heterogenic segments like mixed demersal sheries. Yet, the regional group work highlighted that
different shing nations are in many cases engaged in comparable, regionally consistent shing
activities and that those activities can be identied using the novel eet segmentation approach.
We, therefore, emphasize the necessity for additional, detailed analysis of regional shing patterns,
ideally in regional subgroups. This analysis will help not only to identify major European shing
eets but also to nalize the application protocol of our alternative approach to the segmentation
of fishing eets.
Overall workow perspective
The alternative segmentation package is, in general, a necessary step but not entirely sufcient for
a regular segmentation routine. The main purpose of the segmentation package is to provide
support in dening and describing alternative, sheries-related segments. This step is a manual one
that requires expert knowledge. Expert knowledge is needed for pre-segmenting the eet and for
checks if the clusters generated are appropriate or can/need to be combined or split.
Thus far, participants were, in many cases, able to derive alternative eet segments from the output
of the segmentation package more or less intuitively. However, general principles for these steps
still need to be elaborated. The characteristics of alternative segments should be clearly
documented, e.g., in a table. The description should include
- 6 -
Segment name
Number of vessels
Main gears
Main stocks or stock assemblages targeted
Vessel size range
Engine power range
DCF segments included
Effort and/or landings measures were regarded as informative
…
Once the eet has been segmented for one year, this procedure should be repeated for 2-4 more
years. The results should be checked for consistency. Where necessary, segment denitions and
descriptions can be adjusted. For transparency reasons, it is advisable to cross-check the result of
this step within regional working groups.
As soon as appropriate segments are characterized, the raw data will be used in an Articial
Intelligence (AI) package, which can identify patterns in vessel, logbook, and landings data and
then assign vessels to alternative segments automatically that were dened in the aforementioned
way. The AI package was developed by Verena Dully and Prof. Thorsten Stoeck of the Technical
University of Kaiserslautern. The results of this tool are highly convincing, and a corresponding
publication is in preparation. The AI tool will be part of a future workshop.
Further considerations – criteria for evaluating the suitability of segmentation principles
The existing segmentation approach, which is based on the dominant gear class and vessel length,
has been in place for several years, with an introduction of two additional length class thresholds
(10m and 18m) and a split of one gear class (PTS to TM and PS) in 2008 and a shift in length
threshold from 10m to 8m in the Baltic in 2022. Length classes are easy to derive from the eet
register, and the main gear is usually determined using logbook information, where available. This
approach has been used for many years and has proven to be feasible for all eets. Still, the only
application of eet economics data with the same resolution as provided in the DCF has been the
Annual Economic Report. Hence, the same resolution is used in the "Balance Report".
However, sheries management often requires eet economic data at a different resolution. One
way to address this mismatch is to re-allocate economic data proportional to, e.g., effort or landings.
In order to overcome this auxiliary step, the existing segmentation principle should be reviewed
with a focus on linking segments more closely to stocks or target assemblages.
It appears desirable to aim at keeping the total number of segments similar to the DCF
segmentation. If this is the case, the approach with a closer link to stock is to be preferred. If
deemed necessary, the number of segments can be increased, though. Even in that case, the
sampling effort might not need to be increased as alternative segments can be less heterogeneous
- 7 -
in cost structure and thus require lower sampling rates per segment. In order to compare an
alternative segmentation approach with the existing (DCF) one, several criteria should be taken
into consideration. The following criteria should be taken into account but do not claim to be
complete:
Any alternative approach should aim for a closer link of segments to stocks or groups of stocks.
This aim can compete with the requirements of data condentiality and potentially with cost-
effectiveness. The closer the link to specic sheries, the higher the number of segments and the
smaller the number of vessels per segment. If segments get too small, data might become
condential.
The need for an amendment of the current DCF eet segmentation is based on the observation
that DCF segments, in some cases, combine vessels with different cost structures. Hence, any
alternative approach should result in segments with a more homogeneous cost structure. Like the
DCF segmentation, any segmentation will combine vessels with a broad range of landings and
effort. These characteristics are therefore not in the focus of an analysis of homogeneity. Likewise,
cost items should not be analyzed based on absolute values but rather on indicators or proxies.
E.g., if a group of shrimp beam trawlers has a broad range of fuel cost per year, this will be mainly
due to a broad range of shing days and probably engine power. Therefore, the indicator fuel
cost/kwday might be appropriate to describe homogeneity.
The segmentation procedure has to be clear, doable without excessive extra burden, and repeatable.
It is desirable if the segmentation is compatible with any existing time series. However, a time series
is only as good as the information that comes with it. The link between the current DCF
segmentation and particular sheries can be quite loose. In this case, the time series is not very
informative in terms of fisheries management. If the link between the DCF segment and certain
sheries is close, then the link to the alternative segmentation is close as well, and the time series
is more or less stable. It has to be borne in mind that the recent EUMAP legislation has introduced
a length class threshold at 8m instead of 10m for the Baltic Sea, thus introducing a break in time
series anyway.
- 8 -
Annex
1. List of participants
2. Workshop agenda
3. List of target assemblages
4. List of analyzed regions
5. Case studies
5.1. Finland
5.2. Greece
5.3. Romania
5.4. France
5.5. Germany
Name Institu�on Nation
- 9 -
5.6. Portugal
- 10 -
Second workshop on a novel
approach to the segmentation of fishing fleets
Agenda*
* all times are in CEST and might be changed in the progression of the workshop
Monday
10:00 Welcome and housekeeping
10:30 Brief introduction to the current fleet
segmentation system and its issues (Jörg
Berkenhagen)
11:00 Summary of the last fleet segmentation
workshop and ammendments made to the
fleet segmentation approach (Erik Sulanke)
11:30 Data preparation seminar and getting started
13:00 Lunch break
14:00 Regional subgroups –forming and running
national analyses, if missing
17:00 End of the first day
Tuesday
09:30 Regional subgroup work
13:00 Lunch break
13:00 Regional subgroup work cont.
16:30 Impressions from data work
17:00 End of the second day
Wednesday
10:00 Summary and accomplishments of data work
day (Erik Sulanke)
10:30 Regional subgroups present their results
13:00 Lunch break
14:00 Experience presentation, pt. II
15:00 Key results and implications for the future
16:00 End of the workshop
- 11 -
Target assemblage - code Target assemblage Species (example.)
Region Extent / Remarks
- 12 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 1/13
Alternative Approach to Fleet Segmentation
Workshop report - Finland
Antti Sykkö, Natural Resources Institute Finland (antti.sykko@luke. (mailto:antti.sykko@luke.)), Joonas Valve,
Natural Resources Institute Finland (joonas.valve@luke. (mailto:joonas.valve@luke.))
2022-04-20
Experiments with The FIN 2020 catch data
The following clustering analysis was carried out via eetSegmentation() -package and with the 2020 catch data by Finnish trawlers ( )
and vessels which used passive gears (n). The average vessel lengths in catch data were 25 and 6 meters for trawlers and vessels
using passive gears, respectively (Figure 1).
The pelagic trawlers gear class TM consist of trawlers, which mostly target Baltic herring (Clupea harengus membras) in the Northern Baltic
Sea, with additional target species being sprat (Sprattus sprattus), smelt (Osmerus eperlanus) and vendace (Coregonus albula). Other species
are caught in minor amounts and, such as Herring, are mostly sold as sh fodder. All the vessels sh in the Baltic Sea, (27.d), with majority of
the shing occurring in sub-areas 27.d.32,27.d.31, 27.d.29 and 27.d.30.
The passive gear class PG involves small coastal vessels targeting their shing to, e.g., Baltic herring (Clupea harengus), atlantic salmon
(Salmo salar), pikeperch (Stizostedion lucioperca) and European perch (Perca uviatilis), which are economically the most important target
species for the coastal sheries. The vessels using passive gears sh in the Northern Baltic Sea, mostly in Gulf of Bothnia and Gulf of Finland at
the coast of Finland.
Figure 1: The length of ships in Finnish eet in 2020 in two gear classes - Passive Gears (PG) and Pelagic Trawlers (TM).
Pelagic trawlers - TM
We utilized the segmentation_datapreparation() -function for trawler data preprocessing. The amount of catch in kilograms was chosen as the
catch measurement. We exploited numberclust_table() and numberclust_plot() to receive statistics related to the optimal k, and visualized
the suggested clusters via numberclust_dendrogram() and numberclust_clustree() .
n = 4 3
=1245
- 13 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 2/13
Figure 2: Visual diagnostics for optimal number of clusters in TM catch data.
The initial analysis were performed in order to nd the optimal number of clusters k. Based on statistics (Avg. Silhouette score for
, Mantel score for ), the optimal choice would have been either or . The visual diagnostics (Figure 2) show
that elbow suggests to place ksomewhere between four and eight, silhouette proposes two or three while Mantel encourages to try between
four and seven. However, since we only had 43 vessels and we knew beforehand that the vast majority of herring would be shed, we wanted to
examine the clustering with a sufciently large k. Thus, we chose for further analysis.
Figure 3: Clustering dendogram with range-method for pelagic trawlers
=0 .699
k = 2 = 0 . 8 4 k = 6 k = 2 k = 6
k = 8
- 14 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 3/13
Figure 4: Clustering tree for TM trawler data
TM with
The clustering analysis was performed with number of clusters . We deployed segmentation_clustering() with thee catch data of
trawlers and examined the visualizations provided by clustering_stockshares_plot() ,clustering_assemblageshares_plot() ,
cluster_assemblages_MDS() and clustering_plotgrid() . The results were in line with our initial thoughts as of the vessels
were placed into one, herring-weighted, cluster.
Figure 5: Stockshare plot of TM catch data (n=43)
k = 8
k = 8
31/43≈70%
- 15 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 4/13
Figure 6: Assemblageshares plot of TM catch data (n=43)
Figure 7: Assemblages MD5 plot of TM catch data (n=43)
We looked the properties of the vessels in each cluster (gure 8) before nailing down the nal clustering. While the clusters 1,2,3,6,7 and 8 are
not similar in terms of catch composition (gure 7), they are very similar in terms of vessel diagnostics. We can see that with respect to vessel
length, annual catch per ship and annual catch per cluster, our choice to combine the previously mentioned group of clusters is well justied.
- 16 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 5/13
Figure 8: Diagnostics and statistics of TM catch data vessels (n=43)
TM Final clustering
Final clustering for pelagic trawlers was formed by combining the initial clusters 1,2,3,6 and 7 together and leaving clusters 4 and 5 as they were
suggested by the algorithm. We identied three different clusters.
Herring cluster having 31 trawlers of length between 20-40 meters and targeting their shing heavily on herring (initial cluster 4, gure 9).
Herring & Sprat cluster, which contains 5 relatively large vessels targeting on herring and sprat (initial cluster 5, gure 10).
Mixed cluster holding 6 vessels with length less than 20 meter. These vessels sh a quite of a number of different species, none of which
seems to surpass the others (initial clusters 1,2,3,6,7 , gure 11).
It is interesting to notice that these three nal clusters are in practice identical when comparing the nal clusters dened in these analysis to the
nal clustering determined in previous workshop (2021) for .
k = 1 0
- 17 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 6/13
Figure 9: TM Cluster - Herring
Figure 10: TM Cluster - Herring + sprat
- 18 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 7/13
Figure 11: TM Cluster - Mixed
Passive Gears - PG
The passive gear class analysis was implemented by rst running the segmentation_datapreparation() -function with PG data ( )
followed by the actual clustering analysis. We used the amount of catch in kilogram as a cath measurement. The diagnostics and statistics were
then examined in order to choose kfor further analysis. Based on elbow and silhouette means (score for and ), we can
see that, overall, the optimal number of clusters seems to be either or (gure 1). However, Mantel test makes an exception
suggesting to try clustering with clusters (highest score for ). Moreover, we also deployed numberclust_clustree() -
function to create a clustering tree to provide additional visualizations.
Figure 12: Visualization - Estimated number of clusters for passive gear (PG) class.
n = 1 2 4 5
≈ 0 . 3 8 8 k = 6 k = 7
k = 6 k = 7
k = 1 5 0 . 7 4 9 k = 1 5
- 19 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 8/13
Figure 13: Clustering three with PG data
PG with
We chose clusters for further analysis. We utilized passive gear catch data and performed clustering with segmentation_clustering() -
function. We started our experiments by looking the stock shares and assemblage shares over the seven clusters. The initial results were found
quite promising, as we can see (gure 14) that there is relatively clear distinctive specie representation among the clusters.
Figure 14: Stockshares of Passive Gear data with n=1245 and k=7
k = 7
k = 7
- 20 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 9/13
Figure 15: Assemblage shares plot of PG data
We utilized cluster_assemblages_MDS() and clustering_MDS() functions to gain more information of the cluster candidates. From the left
hand side (of gure 16) can be seen that cluster 2 stands out from the crowd, but clusters 5,6,7 and 1,3,4 appear to form relatively similar
groups. The same phenomena can be seen from the right hand side (gure 16)). That is, it looks like there is a major overlap among vessels.
However, if we look at the three dimensional plot (gure 17) and rotate it appropriately, we can notice that the ostensible overlap is just a result
of visualizing the ships in two dimensional space.
Figure 16: Cluster assemblages and clustering MD5 of PG gear class - Multidimensional scaling to two dimensions.
cluster 7
cluster 6
cluster 5
cluster 4
cluster 3
cluster 2
cluster 1
- 21 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 10/13
Figure 17: Clustering MD5 with PG catch data - Multidimensional scaling to three dimensional space
Then we looked at the properties of the clusters in terms of the vessel length, cluster size and annual catches (gure 18). From the annual catch
point of view, it is interesting to see that the biggest share of catch is landed by cluster 2 (gure 18) and the vessels in this cluster are the
longest. This notable amount of catch is natural, as the vessels in cluster 2 are targeting herring, and herring forms undoubtedly the largest
share of annual landings among Finnish sheries.
The vessels in clusters 1,3,4 and 5 have approximately equal lengths, but the annual catch in cluster 1 is considerably larger. However, this is
clearly explained by the number of vessels in cluster 1 versus other clusters. Furthermore, it is interesting to see that there are some outliers in
clusters 1 and 3 in terms of ship length. Thus, we can notice that, from the similarity of shing activities point of view, the vessel length is not a
one classication feature above the others yet it is an important factor in many extent.
Figure 18: PG Vessel properties over different clusters
PG Final clustering
We introduce the nal PG clustering in this section. Based on the previous diagnostics and analysis, the seven clusters provided by
eetSegmentation were categorized as follows:
Herring cluster holding in total of 84 medium size coastal vessels (initial cluster 2, gure 19), which use passive gears and target their
shing heavily on herring.
Salmon cluster having 128 shing small coast ships (initial cluster 3, gure 20) targeting mostly salmon but other species, e.g, european
perch, beam and roach, to some extent as well.
Northern pike cluster consists of 28 small coastal vessels (initial cluster 4, gure 21) targeting their shing primarily to northern pike and
secondary to burbot.
Smelt & Mixed cluster is by far the largest cluster with 986 vessels (initial cluster 1, gure 22) of which most are small coast ships. These
vessels catch mostly smelt, but also sh other species (bream, perch, roach) on side.
Mixed cluster involves 19 vessels (initial clusters 5,6,7, gure 23) of which 15 of them are targeting mostly unknown species and smelt.
Three vessels are catching cod heavily. Furthermore, this cluster has one ship shing solely ounder.
- 22 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 11/13
Figure 19: PG Cluster - Herring
Figure 20: PG Cluster - Salmon
- 23 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 12/13
Figure 21: PG Cluster - Northern pike
Figure 22: PG Cluster - Smelt & Mixed
- 24 -
29.4.2022 12.03 Alternative Approach to Fleet Segmentation
le:///G:/Luke2/Stat_kala_yleinen/PG_ECON_Fleet_segmentation_workshop_03_2022/eet_wsreport_FIN.html 13/13
Figure 23: PG Cluster - Mixed
Discussion
We found the eetSegmentation -package and the results provided by the package very interesting. The problems related to package’s wider
usage in ocial reporting and applications are most likely to be related to small cluster sizes (privacy and applicability of the analysis). In spite of
this, we encourage the further development and usage of the package. We think that utilizing this tool should not be restricted only to ofcial
reporting. Instead, this approach could have a potential analysis usage at those stages where, e.g., number of vessels is invisible to end users.
One example could be to examine the performance in imputation with, e.g., economical data - the traditional way to form stratum versus
clustering approach.
Some keynotes that were not considered in these analysis due to time limitations but good to take account further.
Value of landings Would we see more variety if we include the value of landings into the analysis. This would take into account whether
herring is used to feed or food production, which can lead to drastic differences in price and value. We are aware that value-option is
already available in the package, but reliable results would require some data processing in terms of value calculations. Unfortunately we
had no time to experiment this during the workshop.
Twin-trawlers It would be interesting to test if twin-trawlers could be identied as an independent cluster.
Merging PG data We know there are few small trawlers that report their catch with coastal landings report and are omitted from the data
used here. They should be included.
PG’s targeting Cod It would have been nice to separate cod sheries into their own cluster, but unfortunately the amount of vessels
targeting cod is relatively small.
PG shing on ice We have an additional PG data containing the shing without a ship ( ). This is mostly shing on ice, but these
were not analyzed in the workshop because of time limitations.
Finally, some suggestions for the future development.
Clustering features Some alternative clustering variables would be nice to include in the future. Our understanding is that currently the
package supports a xed list of clustering variables. Would be interesting to see what type of results we would get by using, e.g., a
combination of horse power and catch measurement.
Combining the clusters The strategy in combining clusters likely involves using the vessel properties at least to some extent. An
additional function, which would leverage the cluster combining automatically w.r.t. vessel properties would be an interesting extension to
the package.
n = 3 7 3
- 25 -
- 26 -
- 27 -
- 28 -
- 29 -
- 30 -
- 31 -
- 32 -
- 33 -
- 34 -
- 35 -
- 36 -
•
•
•
- 37 -
• •
- 38 -
• •
- 39 -
• •
- 40 -
• •
- 41 -
• •
- 42 -
• •
- 43 -
• •
- 44 -
•
- 45 -
• •
- 46 -
• •
- 47 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 42 sur 88
5 Applications of the alternative segmentation tool on the French
national fleets operating in the supra-region Atlantic
5.1 The French fleet operating in the supra-region Atlantic
As mentioned above, the application exercise of the alternative segmentation tool is applicable
only to the French fleet operating in the supra-region Atlantic (FAO area 27) based on the vessel
(individual) landings data available in the SACROIS data-set (see before for details). The application
exercise is neither feasible for the Mediterranean fleet operating in FAO 37 nor for the Outermost
fleet operating in Indian Ocean (area 51) and Western Atlantic (FAO area 31 & 41).
In 2020, French fishing fleet operating in the supra-region Atlantic consisted of 2 900 vessels; 187
being inactive. The 2 713 active vessels presented a high variability in term of vessel length from
less than 4 meters to more than 90 meters vessels (see next figure). The majority (52%) were less
than 10 meters vessels (1495 vessels) when the more than 24 meters vessels represented less
than 5% of the total fleet (125 vessels).
Figure 13: Number of vessels per vessel length ranges (Atlantic area 27) in 2020
In 2020, The Atlantic fleet was distributed as follows according to the EU-MAP and Ifremer-FIS
segmentation.
Table 20: Number of vessels per EU-MAP fleet segment (Atlantic area 27) in 2020
*Please note that this case study is an extract of the original Ifremer document:
Demaneche Sebastien, Guyader Olivier, La Grand Christelle, Merzereaud Mathieu, Vigneau Joel,
Quentin Laurent (2022). Alternative approaches to the segmentation of EU shing eets. Workshop
II - 28-30th March 2022. Previous experiences, tests for application in the French context and
recommendations. PDG-RBE-HISSEO, PDGRBE-EM, PDG-RBE-HMMN-LRHPB.
https://doi.org.10.13155/89336
- 48 -
*
- 49 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 43 sur 88
Table 21: Number of vessels per Ifremer-FIS fleet segment (Atlantic area 27) in 2020
A high diversity of the fishing gears in used by these vessels was observed which lead to a high
distributed fleet by fleet segment. The Ifremer segmentation allows to assess the exclusive or non-
exclusive nature of fishing strategies of the vessels highlighting that the combination of two (or
more) fishing gears during the year is very common. This should be considered to carry out a fleet
segmentation of interest.
5.2 Methodology to tests the alternative segmentation tool
Different tests of the alternative segmentation tool (clustering approach) were applied on the
French fleets operating in the supra-region Atlantic based on the “Fleet Segmentation manual”35.
In a last step, results of the clustering approach obtained by EU-MAP fleet segment were also
compared with other alternative pre-existing fleet segmentations using intra vs inter variances
indicators and stability indicators over the years.
Two different methods were tested to pre-segment the full dataset by:
1) EU-MAP fleet segment,
2) Ifremer fleet segment
Then the clustering approach (R-package provided in the context of the workshop) was tested
based on the following metrics:
Catch composition profiles in weight (Ldgs/species*sect) and value (val/species*sect)
Total landings by “fishing gear” (métier DCF level4) in weight (Ldgs/metDCF4) and value
(val/metDCF4)
35 Fleet Segmentation - Package Manual. Erik Sulanke Thuenen-Institute for Sea Fisheries, Bremerhaven,
Germany.
https://rdrr.io/github/ESulanke/FleetSegmentation/f/vignettes/FleetSegmentation_vignette.Rmd
- 50 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 44 sur 88
Total landings by “métier” (métier DCF level5) in weight (Ldgs/metDCF5) and value
(val/metDCF5)
Total landings by “métier” * “ICES division” in weight (Ldgs/metDCF5*sect) and value
(val/metDCF5*sect)
Figure 34: Methodology used to compare different segmentation and different metrics
Finally, the clustering approaches results obtained by EU-MAP fleet segment were compared with
the following alternative pre-existing fleet segmentations:
Ifremer-FIS fleet segmentation (FLEET_IFR),
Capacity regions of the vessels (REG_CAP),
Vessel length ranges (VSL_LGTH),
Ifremer-FIS fleet segmentation * vessel length ranges (VSL_LGTH/FLEET_IFR),
Vessel length ranges * capacity regions (VSL_LGTH/REG_CAP),
Ifremer-FIS fleet segmentation * capacity regions * vessel length ranges
(VSL_LGTH/REG_CAP/FLT_IFR)
At the end, 14 different fleet segmentations were compared.
- 51 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 45 sur 88
5.2.1 “Demersal trawlers and/or demersal seiners (DTS)” EU-MAP fleet segment, metric
= catch composition profiles in weight
First application exercise of the alternative fleet segmentation tool was applied by EU-MAP fleet
segment on catch composition profiles in weight (default approach proposed by the tool). The
result obtained for “Demersal trawlers and seiners (DTS)” EU-MAP fleet segment is briefly
presented hereafter. The results of the other EU-MAP fleet segments are available in Annex I.
Figure 22: Results for the Demersal trawlers and/or demersal seiners (DTS) EU-MAP segment, metric = catch
composition profiles in weight.
A first conclusion of this exercise is that the alternative fleet segmentation tool seems not well
adapted to the specificity and the diversity of the French fishing fleets. One of the major issues
from this application is that the tool tends to highlight some very specific/specialized vessels
designing fishing segments with less than 5 to 10 vessels and keeping the majority of the other
fishing vessels in 2 to 3 large diverse groups where the principal stocks landed are grouped. The
segmentation carried out with the proposed clustering R-package failed to achieve the objective.
This seems to be linked to the high diversity observed in the different French EU-MAP fleet
segments. Pre-segmentation of the data before applying the approach is therefore a key issue to
- 52 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 46 sur 88
consider. To try to avoid the issue linked with the polyvalence of the vessels belonging to the same
EU-MAP fleet segment, same application exercise was carried out on the pre-segmented data set
by Ifremer fleet segment (segmentation based on gear or combination of gears practiced).
5.2.2 “Demersal trawlers exclusive” Ifremer-FIS fleet segment. metric = catch
composition profiles in weight
The results achieved for the application of the alternative fleet segmentation tool by Ifremer-FIS
fleet segment for “Demersal trawlers exclusive” Ifremer-FIS fleet segment is briefly presented
hereafter, other segments could be found in Annex II.
Figure 23: Results for Demersal trawlers exclusive Ifremer-FIS segment, metric = catch composition profiles in weight.
- 53 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 47 sur 88
The application exercise gives better results on Ifremer-FIS Fleet segmentations (especially for
specialized vessels e.g. “Demersal trawlers exclusive” or “Netters exclusive”) but the tool seems to
have still difficulties to identify a segmentation adapted to the polyvalent/diversified vessels
considered and tends to continue to group specific/specialized vessels into small groups and to
keep other fishing vessels in 2 to 3 large diverse groups.
This issue should possibly be linked to the technical statistical parameters considered in the tool
as the “distance” or the “segmentation method” (e.g. hierarchical agglomerative cluster analysis
(HAC)). Maybe should be valuable to propose in the tool an alternative choice 1) for the distance
as a “denormalized distance”, 2) for the segmentation method as a “k-means clustering method”
or 3) to parameter a “minimum cluster size control”. Furthermore, it is not obvious if the
classification tool considers the “absolute value” or recalculate the data in “percentage”. The two
different possibilities should be possibly allowed and tested.
5.2.3 “Demersal trawlers exclusive” (Ifremer-FIS-segment)- catch composition profiles
in value
The metric to perform the segmentation could be questioned especially considering the landings
weight vs the landings value. To classify the vessels into fleet segment, like to define the metier
and for the same reasons, it seems that value landed should be a better metric to consider.
Actually, same considerations apply that the ones approved in the DCF WK Métier Workshop36:
“However, it is the recommendation of this group that if target assemblage is defined as describing
the fisher intent then value is the metric that should be used, as fisheries are conducted for
economic gain. Likewise, when species with a low weight relative the value is the real target, then
value is a better metric. Finally, the use of value as the metric for target assemblage would help to
avoid the complication created by the implementation of the landings obligation, where
potentially large weights of low economic value could affect any post classification system based
solely on weight, resulting in incorrect definition of fishing intention. Despite this, there might be
some cases where a combination of value and weight should be used. For example, purse seiners
targeting small pelagic fish can catch a school of the target species but if some other valuable
species are caught in less weight the output of the trip can be conditioned by the more valuable
species although it was not the original target. A combination of the two criteria should be used in
these cases.”
In order to test this assumption same application exercise was applied on the French fishing fleet
considering the total value landed by species/stocks rather than the weight. See hereafter, an
example of the results obtained on the Ifremer fleet segment “Demersal trawlers exclusive” using
the value to be compared with the previous plot presenting the results for the same vessels with
the metric in weight.
36 Anonymous report: DCF Métier Workshop: Sub-group of the RCGs - North Sea and Eastern Arctic and
North Atlantic. 22 - 26 January 2018. DTU Aqua, Lyngby, Denmark.
https://webgate.ec.europa.eu/regdel/web/meetings/507/documents/1697
- 54 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 48 sur 88
Figure 24: Results for Demersal trawlers exclusive Ifremer-FIS segment, metric = catch composition profiles in value.
Considering the value of species landed allow to better segment the fleet especially into two
different ‘big’ groups and presents a better GoF (0.55 vs 0.38).
However, the two principal groups defined remain relatively big (168 and 118 vessels, groups 1 &
4) and the tool still seems to focus on very specific/specialized vessels regrouping them in small
groups (groups 2, 3, 5, 6, 7, 8 & 9). For example, the group 9 concerns only one vessel with landings
declared in the 27.1.2 which is very specific when the group 1 aggregate 168 vessels with
important landings of nephrops (NEP), sole (SOL), anglerfish (MNZ), hake (HKE) and cuttlefish
(CTC); all of them being “structuring species” for the exclusive Demersal trawlers fleet operating
in the supra-region Atlantic.
- 55 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 49 sur 88
5.2.4 Catch composition in weight vs total landings in value by fishing gear (metier DCF
level4) for polyvalent fleets (example of Netters Potters/Traps)
In order to develop a proposal flowchart to be applied to segment the fishing fleets. Another test
was done on the “Netters Potters/Traps” Ifremer fleet segment (polyvalent fleet) comparing the
results obtained by the tool directly (based on catch composition) from another approach
considering the vessel’ total landings in value by fishing gear practiced during the year (in order to
better take into consideration the polyvalent nature of the vessels considered).
Figure 25: Results for Netters Potters/Traps – Metric: catch composition profile in weight.
- 56 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 50 sur 88
Figure 26: Results for Netters Potters/Traps – Metric: total landings in value by fishing gear (metier DCF level4)
Although the groups are similarly balanced in both analysis with three big diverse groups
constituted and other groups being relatively small; it seems that the “big” diverse group are more
heterogeneous in the first process (see dimension1 * dimension 2 maps). Therefore, a first step
based on the combination of gears used by the vessels seems to better structure the fleet (pre-
segment the dataset) before getting one step further regarding the group of species targeted (i.e.
the metiers practiced during the year) and finally the species/stocks composition. For example,
here it should be useful to distinguish vessel combining “nets and pots metiers” vs “nets and fike
nets (traps)”.
- 57 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 51 sur 88
5.2.5 “Demersal trawlers exclusive” Ifremer-FIS segment, metric catch composition in
weight vs total landings in value by “métier” (metier DCF level5)
In the same way and also to test the proposal flowchart presented hereunder, a test was applied
on the “Demersal trawlers exclusive” Ifremer-FIS fleet segment (exclusive fleet) comparing the
results obtained by the tool directly (based on catch composition) from another approach
considering the vessel’ total landings in value by métier DCF level5 during the year.
Figure 27: Results for Demersal trawlers exclusive Ifremer-FIS segment, metric = total landings in value by métier DCF
level5
- 58 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 52 sur 88
The tool (even if issue stated before remains) applied on the basis of total landings in value by
metier DCF level5 (i.e. by group of species targeted) allow here to better divide the fleet into
groups more balanced than the groups obtained applying the tool on catch composition profile.
For example, considering the metiers for this exclusive fleet allow to divide the “exclusive demersal
trawlers” between “exclusive demersal trawlers targeting crustaceans”, “targeting cephalopods”
or “targeting demersal fishes”. This highlight that further steps should be then required to
segment the fleets regarding the species/stocks’ catch profile composition i.e. taking into
consideration the different métiers operating by the vessels during the years (their operating
strategy).
Indeed, métiers (regrouping fishing activity based on gear type, mesh size & target species/fish
stocks) have been defined to picture the fishing strategies of the vessels and regroup fishing trips
according to similar exploitation patterns. See following conclusions stated during the DCF WK
Métier Workshop37:
“Recently, the recast of the EU-MAP Regulation reaffirms the métier as an important domain of
interest. Today fleet and métiers are commonly employed in European fisheries to form the
building blocks which describe the heterogeneity of fishing activity in both biological and economic
terms. These building blocks allow the partitioning of landings and effort into ‘sensible’ sized units
representing the fishing activities within them (ICES, 2003). The functionality of métiers is evident
in the number of groups (i.e. DCF, ICES, RCG, GFCM, RFMO, ...) who now use them for a variety of
programs, such as the pre or post stratification/aggregation of national sampling programs, bio-
economic modelling (e.g. Ulrich, Reeves, Vermard, Holmes, & Vanhee, 2011) and management
strategy evaluations (e.g. Vermard et al., 2008). Ultimately, well-defined métiers provide the
building blocks of more effective management (Davie & Lordan, 2011) and constitute a potent tool
to improve biological and bio-economic expertise, to move towards an ecosystem-based approach
and to better estimate PETS bycatch data. The use of métiers makes it possible to describe the
fishing behaviour/fishing practices of fishermen and constitute a sound basis for the typological
classifications of vessels by fleet segment, which forms the basis of economic data collection.”
Furthermore, the matrix “metier*fleet” developed since the inception of the DCF has been defined
to consider the fact that fishing activities on a yearly basis (vessel’ operating strategy) affects the
economic performance and the fishing activity at the trip level defines the exploitation pattern to
sample, the metiers making the connection between the economic and biologic parts. Finally,
métier were harmonized between countries in order that one metier can be used in a region to
describe the same types of fishing activities across nations which reinforce the importance and
the needs to consider the métiers and eventual combination of in the process of fleet
segmentation.
37 Anonymous report: DCF Métier Workshop: Sub-group of the RCGs - North Sea and Eastern Arctic and
North Atlantic. 22 - 26 January 2018. DTU Aqua, Lyngby, Denmark.
https://webgate.ec.europa.eu/regdel/web/meetings/507/documents/1697
- 59 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 53 sur 88
5.3 Comparison of results achieved through the clustering approaches based on
different metrics with the alternative pre-existing fleet segmentations
One of the main objectives of the fleet segmentation is to build homogeneous groups of vessels
stable over the years to improve the accuracy and precision of the calculated estimates especially
in terms of revenue and cost structure and other related indicators. To test and compare the
results obtained from the tool (clustering approach based on different metrics in weight and value)
against other alternative pre-existing fleet segmentation, inter-stratum variance of key indicators
were calculated and analysed. The problem of small size clusters was also considered because of
the confidentiality issues at the stage of the estimate’s restitution. Finally, the issue of the stability
of the clusters was considered in a second step.
5.3.1 Inter and intra variance for different segmentations and metrics
Following graphical outputs were edited by EU-MAP fleet segment (example for the EU-MAP “Drift
and/or fixed netters (DFN)” fleet segment is presented hereafter; other results are available in
Annex III). Each column of the graph illustrates one of the tested segmentations, and each row
refers to a particular indicator. The first row presents an indicator assessing the importance of
small-size clusters in the result obtained. It describes the number of clusters aggregating only one
vessel, aggregating 2 to 4 vessels or aggregating more than 5 vessels. The second row completes
the first one presenting the number of vessels allocated by groups with less than 5 vessels.
The other rows present the inter-stratum (green) and intra-stratum (red) variances calculated by
fleet segmentation for some key indicators-metrics: fishing days, days at sea, hours at sea, total
landed weights and total landed values38. One of the principals aims of a fleet segmentation being
to maximize the inter-stratum variance and minimize the intra-stratum variance of economic
indicators.
Figure 28: DFN EU-MAP Fleet segment – Comparison inter/intra stratum and small size clusters between results of
alternative segmentation tools and pre-existing fleet segmentations
38 At this stage, cost indicators were not included in the analysis.
- 60 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 54 sur 88
Regarding the inter/intra-stratum variance analysis, the graph highlights the importance to first
segment the dataset by vessel length ranges which concentrate a lot of the variability (strong
contribution to the inter-stratum variance) observed between vessels and present better results
than all the other segmentations tested, whatever the variable considered. At this stage, the
clustering approach (whatever the metrics considered) failed to propose a fleet segmentation
which explain more variability than first separate vessels by vessel length ranges.
Following that, pre-segment the data set by vessel length ranges could be a way to produce better
results and to improve the homogeneity of the fleets segments obtained. Same analysis was
carried out by EU-MAP fleet segment * vessels length ranges. Example for the DFN EU-MAP Fleet
segment and the VL1012 vessel length range follows:
Figure 29: DFN EU-MAP Fleet segment – VL1012 – Comparison inter/intra stratum and small size clusters between
results of alternative segmentation tools and pre-existing fleet segmentations
As expected, pre-segmenting the EU-MAP fleet segment first by vessel length ranges allow to
decrease the total variance i.e. the process benefit from this preliminary stratification mitigating
the negative effects of too much overall heterogeneity in the population considered. This should
be regarded also considering the possible threshold issue linked with the predefine vessels length
limit to define vessels length ranges and also the usefulness to aggregate vessels from different
vessel length ranges presenting similar fishing strategies.
5.3.2 Stability of the clustering approaches
Another issue of the approach is the requirement to compare fleet segments across years. The
following graph compare the results obtained for two different years (2018 vs 2019) on the same
fleet segment and highlight the instability of the results obtained year to year. The graph presents
information about the stability of the results obtained in the two years from the different tested
segmentations i.e. segmentation provided by the tool (clustering approach based on different
- 61 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 55 sur 88
metrics in weight and value) and pre-existing fleet segmentation. Each square present a tested
segmentation and compare the results obtained in the two years (2018 & 2019).
For DFN EU-MAP fleet segment (presented as an example in figure 30 hereunder), it highlights a
high instability when using the clustering approach which is less the case regarding the pre-existing
fleet segmentation. Similar results have been observed for the other EU-MAP fleet segments. It
concludes that clustering approaches are a good statistical mean to analyse/explore the fishing
fleets studied but it is crucial from their results to define stabilized set of rules which could be
applied all along the period, year after year.
Figure 30: DFN EU-MAP Fleet segment – Comparison of the results obtained in 2018 and 2019 for the different tested
fleet segmentations. Indicator of stability/instability of the results obtained.
- 62 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 56 sur 88
5.4 Conclusions and recommendations
Based on the application of the proposed clustering R-Package to the French fleet operating in the
supra region Atlantic, our first conclusion is that the use of Principal Component Analysis and
clustering approaches is not appropriate to define fleet segmentation. If the tool is very interesting
for a preliminary understanding of the fishing fleets and activities which are complex by nature,
one of the issues with PCA analyses is that it is difficult to control how the groups are formed.
Moreover, PCA analysis do not produce stable groups/segments over time and across countries
and may even change historical perspectives upon addition of new years in the dataset. PCA
results can also lead to the definition of groups that are often too large or too small, which is also
a pitfall to be avoided (small groups) for statistical and confidential reasons.
The metric proposed by the tool is the “Catch composition profile in weight” because it is
supposed “to better represent the fishing strategies of the vessels, the stocks used and how mixed
or targeted a fishery is”. Based on our analysis, the landings in value per species or stock seems to
be a better metric than weight for the majority of the fleets. Reasons for that are similar as what
it was approved in the DCF WK on metier issues. Based on our results, it seems also crucial to
better consider the polyvalent/non-exclusive nature of the fleets in terms of fishing gears and
métiers. The specific methodology we developed for the analysis of inter/intra-stratum variance
also highlights the importance to first segment the dataset by vessel length ranges which
concentrate a lot of the variability. Indeed, vessel length ranges present better results than all the
other segmentations tested, whatever the variable considered.
A flowchart proposal for the segmentation of the EU fleet39
The following flowchart tries to synthetize the step-by-step approach proposed with the objective
to define a set of agreed and objective rules for the improvement of the EU fleet segmentation.
1. Necessary criteria for fleet segmentation
First of all, it is crucial to consider that necessary criteria for fleet segmentation should be i)
stabilized and easily replicable over time and ii) harmonized and standardized between member
states and fisheries ecoregion. Any segmentation should be stable: This means that segmentation
rules cannot be changed every year or too regularly. Obviously, if the segments change regularly,
the basis for calculating indicators evolves over time and it is therefore not possible to monitor
the economic performance or other indicators related to the vessels and fleets over time and
across countries. Furthermore, even if specific fishing activities may be operated in each member
state, the same easily identified set rules must be applied in each MS or/and each fishing
ecoregion for different member states. Moreover, the segmentation should not be too fine.
Indeed, when the segmentation is too fine, vessels can migrate from one fleet/segment to another
too easily even with minor changes in their fishing and production strategy. This can result in an
instability of the groups, which is also not desirable for the analysis of series. Another
consideration is the compatibility with previous DCF time series.
39 Complementary to fleet segmentation, it is fundamental to keep in mind the “metier * fleet” matrix
which gives the possibility to connect vessels to species and stocks through metiers.
- 63 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 57 sur 88
Finally, it is imperative on the one hand to respect the rules relating to confidentiality (for example
at least 5 vessels per segment) and on the other hand to have segments of sufficiently large size
to be able to get economic samples of acceptable size (in other terms limiting also the number of
segments).
2. Design of the set of rules
There are different options as soon as step 0: Either to reconsider completely the fleet
segmentation at global EU fleet level or to develop a new sub-segmentation of the ongoing EU-
MAP fleet segments. Vessels considered could also be regionalized by fishing ecoregion (e.g. Bay
of Biscay and Iberian waters, North Sea and Eastern Channel, …). Before any further investigation,
vessel length ranges, as key parameter, should be considered and discussed for improvement.
Then, whatever the option adopted, the first step (step 1) is to segment the fleet by fishing gear
or combination of fishing gears used by the vessels, then (step 2) by metier DCF level5 (i.e.
principal group of species targeted) or combination of and then (step 3) by catch composition in
value by species/stocks. The benefit of such an approach would be to better consider the different
dependencies to species and contribution of fishing mortality to fish stocks as well as the
polyvalence of vessels. Based on this analysis, the set of rules need then to be codified to be easily
replicable each year (threshold to be developed).
All of that, lead to consider the following fleet segmentation flowchart proposal. It takes first
(step1) into consideration the fishing gears or combination of, used by the vessels (exclusive or
polyvalent vessels), separating vessels by their exclusive (e.g. exclusive trawlers, exclusive netters,
…) or non-exclusive/polyvalent nature (e.g. trawlers-dredgers, netters-potters, …). Next step
(step2) considers the metiers or combination of metiers practiced (i.e. the principal target species
or combination of targeted) for example separating exclusive trawlers vessels between exclusive
pelagic trawlers, exclusive demersal trawlers or mixed exclusive trawlers targeting demersal and
pelagic fishes. Finally, in a third round (step3) some specificities regarding the catch composition
in species/stocks of the vessels considered could be highlighted for example separating exclusive
demersal trawlers between Nephrops specialized exclusive demersal trawlers and non-specialized
exclusive demersal trawlers. This last step allows to better define/divide the groups established.
This should be associate with considering the vessel length ranges and fishing areas. In this
method, the alternative fleet segmentation tool developed will be useful as a statistical mean to
analyse the dataset and define the set of rules to be applied in application of the flowchart.
- 64 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-30th March 2022. Previous experiences, tests for
application in the French context and recommendations. Page 58 sur 88
Figure 31 Flowchart proposal to define a set of rules to segment the EU fleet
- 65 -
Workshop on alternative approaches to the segmentation of the EU fishing fleets (II) - 28-
30th March 2022. Previous experiences, tests for application in the French context and
recommendations. Page 59 sur 88
3. Methodology
To define the set of rules, preliminary analyses of fishing fleets (by length category including the
evolution of the number of vessels, fishing effort and landings by species in weight and value)
including the regulatory contexts should be developed40. PCA tools and clustering approaches are
very interesting for a comprehension of the fleets and fishing activities which are complex by
nature. Based on these different approaches, the results should be translated into
stabilized/standardized and harmonized set of decision rules shared between member state,
easily reproducible year by year in order a vessel will be allocated to one fleet segment in the
same way in each MS for the fishing ecoregion considered. Exchange and discussion with
stakeholders could be also useful at this stage.
4. Data availability issues and small-scale fleets
As mentioned above, the application exercise of the alternative segmentation tool is only
applicable to fleets where vessel (individual) landings data (landings per species and stocks) are
available. However, the lack and incompleteness of reliable data at vessel level has been reported
in many contexts especially for small-scale fleet. This situation may jeopardise the capacity to carry
out alternative segmentation approaches but there is no valid reason not to apply an alternative
segmentation as these fleets are economically and socially important and may also be affected by
management measures or more broadly by management plans. Because small-scale fleets present
regularly a greater diversity in term of fishing gears used than the large-scale fleets, it considered
as inadequate to allocate them into one unique heterogeneous PGP (Vessels using polyvalent
“passive” gears only) Fleet segment. For these fleet, complementary data collection scheme as
the vessel fishing activity calendar census survey (VFACCS) is considered as an appropriate
approach as soon as declarative data are assessed as incomplete or insufficient to meet the end-
users needs.
40 See the fisheries overviews as a first step to follow: https://www.ices.dk/advice/Fisheries-
overviews/Pages/fisheries-overviews.aspx
- 66 -
Segmentation of the Demersal Trawlers (DTS) operating in Iberian Waters
(ICES Divisions 27.9.a and 27.8)
Authors: Ana Cláudia Fernandes, Suzana Faria Cano
Description of analyzed population
The Portuguese Demersal Trawlers (DTS) have annual fixed licenses to operate with bottom otter
trawl and, considering the mesh sizes used, they can target mainly demersal species (65-69mm or
≥70mm) or crustacean species (55-59mm and ≥70mm). The areas considered for this group reflect
their fishing behavior: it is known that some vessels may operate in ICES Div 27.8.c (Spanish
waters) part of the year. The data used in the analysis is from 2021 and includes 90 vessels with
vessel length range between 9-34m (mean=23 m; median=24m).
In Portugal there are fisheries targeting species that are not assessed by ICES but are very
important at national level (e.g. Atlantic chub mackerel). In this work we removed the ICES stock
assignment limitation to enable the classification of all ‘species versus area’ as a new ‘stock’. The
high number of species*area present in the data (276) reflects the multispecificity of this fleet and
the possible difficulty in assigning the target assemblages to some of the vessels (or fishing trips).
The objective of this case study is to analyze the differences observed in the results, if we consider
different number of clusters in the data.
Analysis of the data
The analysis was performed using the ‘FleetSegmentation’ package.
The clustering diagnosis used to analyze the optimal number of clusters were the average
silhouettes, the Mantel test, the Davis-Bouldin index, the SD index and the Calinski-Harabasz index
(Figure 1). The optimal number of clusters varies between 2 and 15, depending on the indices
considered. Also the clustering dendrograms presented in Figure 2 support that 1 to 24 distinct
clusters are present in the data. These results reflect well the diversity of species assemblages
observed in this fleet. In our study we will start to consider the optimal cluster number obtained
from the average silhouette (6 clusters).
- 67 -
- 68 -
- 69 -
- 70 -
- 71 -
- 72 -
- 73 -
