CNRS report .pdf



Nom original: CNRS report.pdfTitre: Microsoft Word - CNRS_report REACTUALISEAuteur: FORMATECO 7

Ce document au format PDF 1.4 a été généré par PScript5.dll Version 5.2.2 / GPL Ghostscript 8.15, et a été envoyé sur fichier-pdf.fr le 12/10/2017 à 15:21, depuis l'adresse IP 86.205.x.x. La présente page de téléchargement du fichier a été vue 222 fois.
Taille du document: 1.2 Mo (21 pages).
Confidentialité: fichier public


Aperçu du document


Rousselot Cyrielle
Zoccola Emmanuelle

Elaboration of training sets for the
phytoplankton with the FlowCam method and
test on other samples.

1

Index
Abstract :

3

General presentation of the LOV
Presentation of the team
Studies lead by the team
Presentation of our mission

Introduction :

5

Description of the site End-to-End Point B
Phytoplankton

Material and Method :

6

Sampling of the phytoplankton
Description of the FlowCam process
Elaboration of a set

Results :

12

Training set 50µm
Training set 20µm

Our Experience :

21

2

Abstract :
General presentation of the LOV :
The Laboratoire d’Océanographie de Villefranche (LOV) is located in France, between
Monaco and Nice. It is recognised as one of the world most famous oceanographic
laboratory, thanks to its research domains and the skills of its searchers. The research
themes are various, from the chemical, physical and biological oceanography to the
marine optics and remote sensing including the geochemistry, the biochemistry and the
biodiversity and ecology of marine plankton. The LOV is under the guardianship of the
CNRS (Centre National de la Recherche Scientifique) and the UPMC (Université Pierre
et Marie Curie). It is part of the Observatoire Oceanologique de Villefranche (OOV).
Presentation of the team :
Lars STEMMANN
Research professor and lecturer specialized in the
dynamics of the plankton, physical and chimical
processes.

Jean-Baptiste ROMAGNAN
PhD student.

Franck PREJGER
Technician specialized in zooplancton and responsible of
the zooplankton collections.

Corinne DESNOS
Laboratory technician specialized in the analysis of
zooplankton (use of the ZOOSCAN – Imaging particle
analyzer.)

3

Studies lead by the team :
• At small and medium spatial and temporal scales, interactions between the
chemical and physical process, the primary biological production and the
exploitation of this production by secondary producers.
• The effect of the biological cycles on the chemical properties of the mixed layer
and the exchanges between atmosphere and oceans.
• The diversity and the spatial structuring of the zooplankton community in
relationship with the trophic and ecological environment under the influence of
climatic, hydrological, biological and anthropic factors at seasonal and pluri
annual scales.
• The role of zooplankton keystone species in energy and matter flows into the
pelagic ecosystem.
• The diagnostic and/or statistic modelling of the trophic web, from mineral matter to
macroplankton.
• The effect of the natural variability (micro-scale) on the dynamics of the biological
process of the plankton. It is realised by experimentations in monitored
environment and by phenomenological modelling.
• Role of small and medium scale flows (jets, fronts, eddies,) on the formation and
circulation of oceanic water masses. The impacts on the vertical advections and
the ocean-atmosphere exchanges.
Presentation of our mission :
We are undergraduate students in Marine Biology in second year at the SKEMA
Bachelors EAI in Sophia Antipolis and we were volunteers to have a research experience
in the laboratory of Villefranche-sur-mer. This internship allowed us to have a scientific
experience in a real situation and to get into a laboratory team.
We spent our first month sorting out the samples of point B before hand analysed with
the FlowCam. We also learned to use the FlowCam and to get the different vignettes of
the samples, pre-sorted thanks to a training set. As we noticed that the pre-sorting was
very approximate, we decided as an aim to our internship, to develop a better training set
for the prediction of the FlowCam’s samples sorting of the smallest particles (minimum of
20µm), thanks to the manual validation that we did at first.

4

Introduction :
Description of the site End-to-End Point B :
The roads of Villefranche-sur-Mer is located in the North occidental Mediterranean Sea.
It measures 1.5 km large and 3.5 km long with, at East, the peninsula of the Cap Ferrat,
mountains at North and the Cape of Nice at West. The lack of continental shelf and the
presence of the Canyon of Nice make that the roads presents deep founds : until 80
meters in the roads and more than 500 meters at the end of it. As it is opened to the
South, the dominant winds come from the East. The region is not really exposed to the
Mistral but bear a North wind more or less strong during the winter.
The hydrodynamics of the roads of Villefranche-sur-Mer is impacted by the Ligure current
which circulates from East to West. It enters at the surface (North-East current) and goes
through the roads before exiting in depths near the Cape of Nice (South-South East
current).
The sampling took place at the “Point B” (43° 41’ 00’’ N and 7° 18’ 90’’ E) located at the
beginning of the roads with a depth of 95 meters.
Phytoplankton :
Phytoplankton are all the organism of plankton which belong to Plantae kingdom. They
are autotrophs. These organisms are very tiny and live in suspension in the sea water.
An autotroph (autos = self and trophe = nutrition), also called a primary producer, is an
organism that produces complex organic compounds (such as carbohydrates, fats and
proteins) from simple inorganic molecules using energy taken from the sun or from
chemicals. In the case of phytoplankton, energy is taken from light (by photosynthesis =
primary production). They are thus called photo-autotrophs (photo = light, auto = self,
troph = nutrition).

5

Diatomophyceae, which are usually called diatoms, are characterized by their siliceous
shell called frustules. The frustules are surrounding the diatom and they have a shape
just like a box with its bottom and top. They form a big family of micro seaweeds and they
represent 90 % of the plant plankton. They play an essential role in the marine life
ecosystems because they are primary producer and a unique food source for numerous
species. They are separated in two big groups that are centric diatoms (or biddulphiales
diatoms) and pennate diatoms (or bacillariophyceae diatoms). The first ones appeared
150 million years ago , have a centric or elliptic shape, whereas the second ones,
appeared on Earth 70 million years ago and have often stretched out valves. Numerous
pennates diatoms present a crack which crosses them throughout and which is called the
raphe.
Material and method :
1- Sampling of the phytoplankton
• Net sampling:
Phytoplankton nets are conical in shape, with the wide mouth opening attached to a
metal ring and the narrow tapered end fastened to a collecting jar known as the cod end
(= gatherer) (Figure 1). Nets used in the Laboratory of Villefranche sur Mer have meshes
comprised between 20 and 100 micrometers. At fishing time, the nets are attached to the
hydrowire of the boat and towed behind the ship (Figure 2).

Figure 1

Figure 2
6

• Niskin bottle sampling:
Plankton tows can also be done at any depth or time of day, and can be used with
opening/ closing mechanisms thanks to the Niskin Bottle (Figure 3). It enables them to
collect phytoplankton at a desired depth.
The Niskin bottle has stoppers on both ends, which are held by springs (Figure 4). The
bottle is prepared by opening both ends of the bottle. The Niskin bottle is then attached
to a winch line and lowered to the desired depth. At that point, a small weight, called a
"messenger", is attached to the line and released (Figure 5). When the messenger
reaches the Niskin bottle, the ‘cocking’ mechanism releases the two stoppers, and a
sample of the water from that depth is captured in the bottle.

Figure 3

Figure 4

Figure 5

• Treatment of the different types of sampling before Flowcam analysis:
When nets and Neskin are towed back, some of the planktons collected in the gatherers/
Niskin bottles are put into bottles and fixed with lugol to be analysed by the Flowcam
imaging system (Figure 6).

Figure 6
7

Treatment of the Flowcam samples.
Live sample 20 micrometers

Fixed sample 20 micrometers

Fixed sample 50 micrometers

Fixed sample 100 micrometers

Fixed sample - Niskin bottle












Net (mesh size: 20 micrometers)
No fixation
Depth: 0-75 m.
Net (mesh size: 20 micrometers)
Fixation
Depth: 0-75 m
WP2 net (mesh size: 100 micrometers)
Fraction between 50 and 300 micrometers
Fixation with lugol (2.5%)
Depth: 0-75 m

WP2 net (mesh size: 100 micrometers)
Fraction between 100 and 300 micrometers
Fixation with lugol (2.5%)
Depth: 0-75m
Filtration on a net (mesh size: 20 micrometers)
and concentration on sieve (1.5 micrometers)
• Fixation with lugol (1-2 %)
• Sampling at 55-45-35-25-15-5 m






2- Description of the FlowCam process
Principle of the FlowCam :
This measurement device is used in the analysis and characterisation of phytoplankton
and their repartition in liquid samples. The FlowCam has the ability to examine quite
important volumes of samples in a short time and it takes pictures in flow.
Experimental protocol :
- Measure the volume of the flask to analyse in the FlowCam thanks to a graduated tube.
- Wash the cell with fresh water and dry it with “normal” paper (do not dry it with optical
paper used especially for microscopes because there is a need for some dust particles to
do the focusing).
- Assemble the cell on the cell-holder.
- Wash the funnel with fresh water.
- Place the funnel in the high tube.
- Place the low tube in the device tube.
- Place water in the funnel.
- Switch on the peristaltic pump.
8

- Launch the software “VisualSpreadSheet” for the acquisition and the gestation of the
data.
> Set up
> Set up and Focus
- Focus on debris found at each extremity of the cell (with the help of a toothed wheel of
25 graduations) and then place the camera in the middle of the cell (here 12.5
graduations).
- Verification of the focus.
> Set up
> Auto image Mode (no Save)
Analysis of the data :
- Put the software in automatic mode.
- Parameter the “STOP”.
> Auto image Mode : Stop at 10000 particles
Stop at 45 minutes
- Launch the acquisition.
- Put the end of the tube in a graduated test tube.
- Homogenize each sample to analyse. Place some of the volume of the flask to analyse
in the funnel (there should be as few water as possible).
- During the acquisition, make sure that there is no plug. If there are some, press gently
on the plastic tube to make them go. Think about adding more sample during the process
and to mix the volume in the funnel to avoid deposits and lacks of liquid.
- Calculations of the concentration in phytoplankton with the graphs obtain thanks to the
VisualSpreadSheet.
End of the experimental protocol :
- Read and note the passed volume measured thanks to the graduated test tube.
- Add water in the funnel and place the end tube in the trash beaker.
- Stop the pump once the totality of the volume is discarded in the beaker.
- Free the tube from the pump.
- Rinse the material with fresh water.
Precisions :
- The speed of the picture capture is measured in FPS (frames per second). FPS is a
video display measure mode. Each picture taken is a fixed picture.
- The settings of the lens done from the software are numerous :
capture duration
intensity of the flash
- The volume of each flask contains phytoplankton. These phytoplankton are treated with
potassium iodide and acid which fix the membranes.

9

3- Elaboration of a set
As a training to elaborate learning sets, we first had an initial training for the 50µm set
with vignettes already sorted in 22 large categories.
Then, we wanted to know haw many pictures we needed in each category to obtain an
accurate training set. We decided to make one sub set with the minimum number of
picture to have all the categories homogenous, one with at maximum 100 pictures and
the least with at maximum 200 pictures.
For this, the software used is the ImageJ:
> Launch Zooprocess
> Create Sub-learning Set
> Un-tick the “Use default project architecture ?”
> Browse the initial training set
> Save the sub-learning set at the good place (in the file)
> Enter the number of pictures you want in each category
If there are more pictures than the number wanted, a random selection is made by the
zooprocess, in the ImageJ software.
If there is the same number of pictures than the number wanted, all the pictures in the
category are taken.
If there are less pictures than the number wanted, all the pictures in the category are
taken, but there won’t be the same number of pictures in all the categories.
Then, to test the new learning set, the software used is Plankton identifier :
> Copy the file at the right place (training_set_XXµm)
> Learning
- Browse the file
- Save the “Learn_XXX.pid” in the associated file
- Create learning file
> Data analysis
- Browse the .pid in the good file
- Select the “Cross-validation 4(Random Forest)” method
- Start Analysis
- Save at the right place (PID_process -> Prediction)
When the analysis is over, there is a file named “Analysis_00XX_frame_Cross-validation
1.html” which represents a table of results.
Then, we tried to elaborate a training set for the 20µm nets. The first step in the
elaboration of a set is to sort out vignettes in different categories. These categories have
10

to be very precise because they will determine algorithms which will associate data to
names of organisms. This system is used to sort out the vignettes directly into files
labelled with the names of the organisms they are supposed to contain.
First of all, we made a pre sorting with “same looking” organisms for different
samples : 20110103, 20110110, 20110117, 20110124, 20110131, 20110207, 20110214.
Then, we tried to determine the different species contained in these categories
more precisely and we put together the 7 samples. The categories we determined were
the following :
- aggregates
- badfocus
- ciliates_naked
- ciliates_tiarina_fusus
- ciliates_tintinnids_amphorella
- ciliates_tintinnids_codonellopsis
- ciliates_tintinnids_cyttarocylis
- ciliates_tintinnids_dadayiella_ganymedes
- ciliates_tintinnids_dictyocysta
- ciliates_tintinnids_eutintinnus
- ciliates_tintinnids_proplectella
- ciliates_tintinnids_rhabdonellopsis
- ciliates_tintinnids_salpingella_long
- ciliates_tintinnids_salpingella_short
- ciliates_tintinnids_steenstrupsiella
- ciliates_tintinnids_stenosomella
- crustacean_copepods
- crustacean_copepods_nauplii
- crustacean_exuvies
- diatom_asterionellopsis
- diatom_bacteriastrum
- diatom_centric
- diatom_chaetoceros
- diatom_chaetoceros_danicus
- diatom_chaetoceros_peruvianus
- diatom_chains
- diatom_chains_thin
- diatom_coscinodiscus
- diatom_cyclotella
- diatom_cylindrotheca
- diatom_dytilum
- diatom_guinardia
- diatom_hemiaulus
- diatom_licmophora
- diatom_navicula
11

- diatom_nitzschia
- diatom_odontella
- diatom_pennate
- diatom_pleurosigma
- diatom_rhizosolenia
- diatom_thalassionema
- diatom_thalassiosira
- dino
- dino_ceratium
- dino_ceratium_candelabrum
- dino_ceratium_furca
- dino_ceratium_fusus
- dino_ceratium_geniculatum
- dino_ceratium_limulus
- dino_ceratium_pentagonum
- dino_ceratium_trichoceros
- dino_2
- dino_oxytoxum
- dino_podolampas
- dino_prorocentrum
- dino_protoperidinium
- dino_pyrocystis
- fecal_pellets
- meroplanktonic_larvae
- multiples
- other_small_protists
- others
- pollen
- radiolarians
- rods
- silicoflagellates_dictyocha
- silks
- tubes
In the end, the initial training set was composed of 68 files containing vignettes (Cf.
annexe). Some categories contained very few vignettes, so we decided to delete some of
them. We deleted 14 categories which had less than 40 images. Then, with the ImageJ,
we created 3 sub leaning sets, one of 50 vignettes, one of 100 vignettes and the last of
200 vignettes.

12

Results :
Training set 50µm :
To determine if a training set (TS) is accurate, the first data to be analysed is the
global error rate. In the first TS of 50µm with only 36 images, the global error
rate (GER) is of 52.18 %, the one with 100 images has a GER of 48.91% and
the one with 200 images has a GER of 45.28%.

Global error rate in function of the number of pictures
0,53
0,525
0,52
0,515
0,51
0,505
0,5
0,495
0,49
0,485
0,48
0,475
0,47
0,465
0,46
0,455
0,45
0,445

Graph 1

0

25

50

75

100

125

150

175

200

225

We can notice in the graph 1 that there is a lower error rate with a greater
number of pictures, so we decided to work on the TS with 200 pictures
maximum. As some of the categories had much less pictures than 200, the
decision was taken to find other pictures to complete the categories.
However, looking at the graph showing the recall and the 1-precision (pollution
from other categories in each category), we noticed that there was a great
percentage of 1-precision and that the recall was not very precise except for
some categories (cf graph 2). To optimise the TS, we had to eliminate the
categories with a too high percentage of pollution or to combine two categories
in which the objects have similar shapes.

13

te

fa
ct
ci s_
li a g h a
g
ci tes ost gr
lia _ s e g
t e t i n _b a
s_ tin ad te
tin n fo s
t in id s c u
cr
n i _d s
us
d a
ta c s_l rk
ce o ig
an pe ht
_e p o
xu ds
di
v
di ato dia ies
at m t
di
om _c om
no
_p ha
_c
e n in s
er
at di
na
i n
te
di um o_
no _ ce di
di _c can ra no
no er d tiu
_c at el_ m
er ium pe
at _ n
di
iu f u t
no
m rc
_p
_f a
ro
u
to d s u
p e in s
r id o _
2
m in iu
u
m
lt i
ot
p
he
n a le s
r_
sm
up
al ot lii
l_ h e
pr r
ot s
i
r a p s ts
di ol
ol le
ar n
ia
ns

ar
ac
t
ci s_g
lia h ag
g
t o
ci es sts re
lia _t _ ga
te int ba te
s _ i n df s
tin ni oc
tin ds_ us
ni da
cr
ds r k
us
ta c _li
c e o p gh
an e p t
_e od
xu s
v
di
at d ies
di o i a
at m to
om _c m
di
no
_p hai
_c
en ns
er
na
a t di
te
iu no
m
_
di
di
c
_
no c e n
di
o
a r
no di _c nd ati
_c no era ela um
er _c tiu b
at er m ru
di
no
i u at _ m
m iu fu
_p
_p m rc
ro
en _fu a
ro
di
t a su
pe
no
go s
rid
d
_p in i
n
ro o_ niu d um
to p m in
pe ro _t o_
rid top yp 2
in er iqu
iu id
m in e
_l iu
os m
m an g
ul e
tip
ot
le
he
n
au s
r_
sm
pl
al oth ii
l_
pr ers
ot
is
ra p ts
di oll
ol en
ar
ia
ns

ar
te
f

1,2
1

0,8

0,6
Recall

0,4
1-precision

0,2

0

0,6

Graph 2

Global Error Rate : 0.4528

The categories of dino_ceratium_candelabrum and dino_ceratium_pentagonum
had a recall too weak and a lot of pollution but the organisms were similar so we
combined the 2 categories. The same problem was noticed for the three
categories of protoperidinium, so we also combined them (graph 3).
1,2

1

0,8

Recall
1-precision

0,4

0,2

0

Graph 3

Global Error Rate : 0.3279

To obtain the next training set, we worked principally on the polluted categories.
We combined the nauplii with the copepods as there were polluted by each
other. We did the same with the dino and the dino_2. Moreover, we deleted

14

te
fa
c

ra

po
lle
di
n
ol
ar
ia
ns

ts
_g
ho
ci
lia
st
s_
te
s_
ba
tin
ci
df
lia
oc
tin
te
us
ni
s_
ds
tin
_d
tin
ar
ni
k
ds
_l
ig
cr
ht
us
co
ta
pe
ce
po
an
ds
_e
xu
di
at
vi
om
es
di
_c
at
om ha
in
_p
s
en
na
di
te
no
_c
d
in
er
di
o_
at
no
iu
ce
m
r
at
_c
di
iu
an
no
m
de
_c
l_
e
ra
pe
di
no
tiu
nt
_c
m
_
e
fu
di
ra
no
rc
tiu
a
_p
m
_f
ro
us
to
pe
us
rid
in
iu
m

ar

ra
di

le
n
ol
ar
ia
ns

po
l

ct
s_
ag
gh
gr
ci
eg
os
lia
at
ts
te
_
es
s_
ba
ci
t
d
i
nt
lia
fo
in
te
cu
ni
s_
s
ds
tin
tin _da
rk
ni
ds
_l
cr
ig
us
ht
co
ta
pe
ce
an pod
s
_e
di
xu
at
om vie
s
di
_c
at
h
om
a
_p ins
en
na
di
no
te
_c
d
er
in
di
o_
at
no
iu
ce
m
r
at
_
di
iu
no can
m
de
_c
l
e
_
di
ra
pe
no
tiu
nt
_c
m
e
_
di
fu
no rat
rc
iu
_p
a
m
ro
_f
to
us
pe
us
rid
in
iu
m

ar
te
fa

categories with a recall too weak. The categories deleted were the diatom, the
multiples, the others and the other_small_protists (graph 4).

1,2

1

0,8

0,6
Recall
1-precision

0,4

0,2

0

0,6

Graph 4

Global Error Rate : 0.3128

As the Global Error Rate was not getting lower significantly, we decided to
delete the aggregates category (graph 5).
1,2

1

0,8

Recall

1-precision

0,4

0,2

0

Graph 5

Global Error Rate : 0.2653

15

Analysi
s

Number of
categories

Error
rate

Recall (<0.5)

Changes to bring for the next analysis.

Finally, according to the pollution rate of dino_protoperidinium, we decided to
combine them with the dino to obtain the final 50µm training set (graph 6) with
a GER divided by 2 according to the one we started with.

1,2
1
0,8
Recall
1-precision

0,6
0,4
0,2

ar
te
fa
ct
ci s_g
lia h
t o
ci es sts
lia _t _
te in ba
s_ tin df
tin ni oc
tin ds_ us
ni da
cr
ds rk
us
ta c _li
ce op gh
a e t
di n_ po
a e d
di tom xu s
vi
at
e
_
o
di
m ch s
no
_p ai
_c
en ns
er
na
at din
te
i
di um o_c d
no _c e in
o
r
di _c an ati
no er de um
_c ati l_
er um pe
at _ nt
iu fu
m rc
_f a
us
ra p us
di oll
o l en
ar
ia
ns

0

Graph 6

Global Error Rate : 0.2538

Training set 20µm :
For this second training set, we elaborated a new protocol which allowed us to
put the results in the following table.

16

Analysis

69

54

0.4754 Aggregates(0.1020)*
Ciliates_tintinnids_eut
intinnus (0.1674)
Ciliates_tintinnids_sal
pingella_short
(0.4821)
Ciliates_tintinnidssteenstrupsiella
(0.2667)
Diatom_bacteriastrum
(0.1055)
Diatom_centric(0.0833
)
Diatom_chaetoceros(0
.3450)
Diatom_chaetoceros_
peruvianus (0.1929)
Diatom_chains
(0.3700)
Diatom_chains_thin
(0.2830)
Diatom_coscinodiscus
(0.4571)
Diatom_hemiaulus
(0.0850)
Diatom_navicula
(0.1100)
Diatom_pennate
(0.1469)
Diatom_thalassiosira
(0.2786)
Dino_ceratium_candel
abrum (0.4559)
Dino_ceratium_penta
gonum (0.2483)
Dino_oxy_podo_proro
(0.2375)
Multiples (0.1350)
Others (0.0781)
Other_small_protists

1) To group together
ciliates_tintinnids_salpingella_short
and
ciliates_tintinnids_salpingella_long
in ciliates_tintinnids salpingella.
2) To suppress
ciliates_tintinnids_eutintinnus. This
category mixes with a lot of other
categories creating a muddle in the
training set and distorting the global
results.
3) To fuse
diatom_chaetoceros_peruvianus
with diatom_chaetoceros_danicus
to try to limit the confusion.
4) To group together
dino_ceratium_candelabrum and
dino_ceratium_pentagonum as
there is nearly ¼ chance that they
mix together in the confusion
matrix.
5) To suppress the categories: other
small protists, others, and multiples
as they contain different kinds of
organisms with different shapes
creating confusion in the training
set.

17

(0.1410)
Radiolarians (0.4910)
Rods (0.4890)
Silks (0.3550)
Tubes (0.3590)
NUMBERS OF
CATEGORIES: 26

Analysis

70

46

0.4183 Aggregates(0.0770)*
Ciliates_tintinnidssteenstrupsiella
(0.2711)
Diatom_bacteriastrum
(0.1055)
Diatom_centric(0.0875
)
Diatom_chaetoceros(0
.4040)
Diatom_chains
(0.4170)
Diatom_chains_thin
(0.2200)
Diatom_coscinodiscus
(0.4857)
Diatom_ditylum
(0.4780)
Diatom_hemiaulus
(0.0400)
Diatom_pennate

6) To group together
diatom_chains_thin and and
diatom_chains as these two
categories mix with the second
most important and significative
percentage (nearly 10% of
confusion)
7) To suppress diatom_hemiaulus,
diatom_coscinodiscus, and
diatom_pleurosigma. These
categories mix with a lot of other
categories creating a muddle in the
training set and distorting the global
results.

18

Analysis

71

41

(0.1907)
Diatom_pleurosigma
(0.4914)
Diatom_thalassiosira
(0.3286)
Dino_ceratium_candel
abrum (0.4559)
Dino_ceratium_pentag
onum (0.2483)
Dino_oxy_podo_proro
(0.2333)
Other_small_protists
(0.1410)
Radiolarians (0.4430)
Rods (0.4840)
Silks (0.3360)
Tubes (0.4240)
NUMBERS OF
CATEGORIES: 21
0.3895 Aggregates (0.0820)*
Ciliates_tintinnids_dad
ayiella_ganymedes
(0.4947)
Ciliates_tintinnids_ste
enstrupsiella (0.2711)
Diatom_centric
(0.4176)
Diatom_chaetoceros
(0.3280)
Diatom_chains
(0.3600)
Diatom_thalassiosira
(0.3036)
Dino_oxy_podo_proro
(0.2042)
Radiolarians (0.4960)
Silks (0.4200)
Tubes (0.4450)
NUMBERS OF
CATEGORIES: 11

8) To group together
ciliates_tintinnids_steenstrupsiella
and ciliates_ tintinnids_amphorella
because there is 20% chance that
they mix together.
9) To group together
diatom_chaetoceros and
diatom_chains as there shapes are
quite similar.
10) To suppress the category
diatom_thalassiosira as there is only
30% chance that there is no
confusion with other categories (not
enough)

19

Analysis

72

Analysis

73

Analysis

74

38

0.3725 Aggregates (0.1825)*
Diatom_centric
(0.4904)
Diatom_chains_chaeto
ceros (0.2286)
Silks (0.3810)
Tubes (0.4498)

32

32

Recall (≈0.5)
Ciliates_tintinnids_da
dayiella_ganymedes
(0.5066)
Diatom_asterionellopsi
s (0.5110)
Diatom_guinardia
(0.5523)
Diatom_pennate
(0.5460)
Pollen (0.5920)
Radiolarians (0.5510)
NUMBERS OF
CATEGORIES: 11
0.3326 Aggregates (0.1290)*
Diatom_chains_chaet
oceros (0.4010)
NUMBERS OF
CATEGORIES : 2

0.3246 Aggregates (0.1390)*
Recall (≈0.5)
Diatom_asterionellops
is (0.5820)
Diatom_guinardia
(0.5640)
Fecal pellets ( 0.6300)
Pollen (0.5300)

11) To group together
diatom_chaetoceros and
diatom_asterionellopsis as there is
20% chance that they mix together
as well (as in the precedent
analysis.)
12) To group together diatom_centric
and diatom_cyclotella as their
shapes are very similar and there is
10% chance that they mix together.
13) To suppress silks, tubes,
ciliates_tintinnids_dadayiella_gany
medes as their recall is either
inferior to 0.5 or close to 0.5 from
analysis 69 to analysis 72.

14) To get back the
diatom_asterionellopsis in the
category
diatom_chains_chaetoceros and
suppress this last category as this
category mix with all other
categories and doesn’t appear to be
significant while
diatom_asterionellopsis have a
significant shape (match) and tend
not to mix with other categories.
15) To group together
diatom_chaetoceros_danicus_peruv
ianus, diatom_thalassionema and
diatom asterionnelopsis as there is
10% chance that they mix together
(the moszt important and significant
percentage of the confusion matrix)
16) To group together
diatom_pennate, diatom_dytilum
and diatom guinardia as there is

20

Radiolarians (0.5280)
NUMBERS OF
CATEGORIES : 6

Analysis

75

22

0.2807 Aggregates (0.1330)*
Recall (≈0.5)
Diatom_centric
(0.5320)
Radiolarians (0.5240)

Analysis

76

20

0.2547

NO CATEGORIES.

10% chance that they mix together
(the most important and significant
percentage of the confusion matrix
as well)
17) To group together
ciliates_tintinnids_cyttaroclis,
ciliates_tintinnids_dictyocysta and
ciliates_tintinnids_codonellopsis as
there isnearly 10% chance that they
mix together and because they have
a similar shape (elongated and
amphora-like shape)
18) To group together dino_ceratiu_
furca and
dino_ceratium_candelabrum_penta
gonum, as there is nearly 10%
chance that they mix together and
because they have a similar shape
as well (elongated with kinds of
horns)
19) To suppress pollen and fecal pellets
as these categories are not
significant (mixing with the others.)
20 To suppress radiolarians and
diatom_centric as these two
categories have the least recall just
to try

Error rate considered as correct. STOP OF
THE ANALYSIS.

Our experience :
During 3 months, we participate to the life of a laboratory. We met a lot of
scientist and had the opportunity to discuss about scientific subjects with
specialists. We learned to use the imaging processes and analysis in the study of
plankton, but also we learnt and identified a great number of phytoplankton. We
had an overview of some sampling methods related to the marine studies and
had the opportunity to participate at each level of the studies made on the point
B end-to-end phytoplankton, from the sampling to the imaging and analysis.

21


CNRS report.pdf - page 1/21
 
CNRS report.pdf - page 2/21
CNRS report.pdf - page 3/21
CNRS report.pdf - page 4/21
CNRS report.pdf - page 5/21
CNRS report.pdf - page 6/21
 




Télécharger le fichier (PDF)


CNRS report.pdf (PDF, 1.2 Mo)

Télécharger
Formats alternatifs: ZIP



Documents similaires


summary note dehar hamdaoui lecocq sitbon
resuscitation
algorithme et pensee
workmayenolry
nieuwsbrief2
cnrs report

Sur le même sujet..