Chap e 3 | Da a and Me hodology
OBJECT DETECTION FOR SINGLE TREE SPECIES
IDENTIFICATION WITH HIGH RESOLUTION AERIAL IMAGES
Ekanayaka Mudiyanse Ralahamilage Chamodi Lakmali Boyagoda
ii
OBJECT DETECTION FOR SINGLE TREE SPECIES
IDENTIFICATION WITH HIGH RESOLUTION AERIAL IMAGES
Disse a ion supe ised by
Joel Dinis Bap is a Fe ei a da Sil a, PhD
Ins i u o Supe io de Es a ís ica e Ges ão de In o mação,
Uni e sidade No a de Lisboa
Lisbon, Po ugal
Disse a ion co-supe ised by
Ped o da Cos a B i o Cab al, PhD
Ins i u o Supe io de Es a ís ica e Ges ão de In o mação,
Uni e sidade No a de Lisboa
Lisbon, Po ugal
Disse a ion co-supe ised by
Filibe o Pla Bañón, PhD
Ins i u e o New Imaging Technologies,
Uni e si a Jaume I
Cas ellón de la Plana, Spain
Feb ua y 2020
iii
DECLARATION OF ORIGINALITY
I decla e ha he wo k desc ibed in his documen is my own and no om
someone else. All he assis ance I ha e ecei ed om o he people is duly
acknowledged and all he sou ces (published o no published) a e e e enced.
This wo k has no been p e iously e alua ed o submi ed o NOVA In o ma ion
Managemen School o elsewhe e.
Lisbon, 17.02.2020
Ekanayaka Mudiyanse Ralahamilage Chamodi Lakmali Boyagoda
[ he signed o iginal has been a chi ed by he NOVA IMS se ices]
i
ACKNOWLEDGMENTS
I dedica e his pa ag aph o exp ess my g a i ude o pe sons who helped
me o comple e my mas e hesis success ully. Fi s ly, I would like o exp ess my
since e g a i ude o my supe iso D . Joel Sil a o his con inuous knowledge
sha ing and guidance du ing mas e hesis. Secondly, I would like o hank my
co-supe iso s P o . D . Ped o Cab al and P o . D . Filibe o Bañón o hei
aluable eedback. Special hank goes o P o . Ma co Painho and D . Ch is oph
B ox o hei suppo and guidance h oughou he mas e p og am. I would like
o ex end my g a i ude o all p o esso s o helping me imp o e my knowledge
on he subjec a ea. Finally, I exp ess my g a i ude, lo e, and espec o my amily
and iends o hei encou agemen and suppo .
OBJECT DETECTION FOR SINGLE TREE SPECIES
IDENTIFICATION WITH HIGH RESOLUTION AERIAL IMAGES
ABSTRACT
Objec ecogni ion is one o he compu e ision asks de eloping apidly
wi h he in en ion o Region-based Con olu ional Neu al Ne wo k (RCNN). This
hesis con ains a s udy conduc ed using RCNN base objec de ec ion echnique o
iden i y palm ees in h ee da ase s ha ing RGB images aken by Unnamed Ae ial
Vehicles (UAVs). The me hod was en i ely implemen ed using Tenso Flow objec
de ec ion API o compa e he pe o mance o p e- ained as e RCNN objec
de ec ion models. Acco ding o he esul s, bes pe o mance was eco ded wi h
he highes o e all accu acy o 93.1 ± 4.5 % and he highes speed o 9m 57s om
as e RCNN model which was ha ing incep ion 2 as ea u e ex ac o . The
poo es pe o mance was eco ded wi h he lowes o e all accu acy o 65.2 ±
10.9% and he lowes speed o 5h 39m 15s om as e RCNN model which was
ha ing incep ion_ esne 2 as ea u e ex ac o .

i
KEYWORDS
Con olu ional Neu al Ne wo k
High Resolu ion Ae ial Images
Image Classi ica ion
Objec De ec ion
Region-based Con olu ional Neu al Ne wo k
Remo e Sensing
Unnamed Ae ial Vehicle
ii
ACRONYMS
AI – A i icial In elligence
CNN – Con olu ional Neu al Ne wo k
DEM – Digi al Ele a ion Model
IoU – In e sec ion o e Union
ML – Machine Lea ning
NDVI – No malized Di e en ial Vege a ion Index
NMS – Non-Maximum Supp ession
OBIA – Objec Based Image Analysis
RCNN – Region-based Con olu ional Neu al Ne wo k
RF – Random Fo es
RGB – Red G een Blue
RoI – Region o In e es
RPN – Region P oposal Ne wo k
SSD – Single Sho De ec o
SVM – Suppo Vec o Machine
UAV – Unmanned Ae ial Vehicles
YOLO – You Only Look Once
iii
INDEX OF THE TEXT
DECLARATION OF ORIGINALITY ................................................................... III
ACKNOWLEDGMENTS ...................................................................................... IV
ABSTRACT ............................................................................................................. V
KEYWORDS ......................................................................................................... VI
ACRONYMS ........................................................................................................ VII
INDEX OF THE TEXT ...................................................................................... VIII
INDEX OF TABLES ............................................................................................... X
INDEX OF FIGURES ........................................................................................... XI
1. INTRODUCTION .......................................................................................... 1
1.1 An o e iew o he wo k ............................................................................ 1
1.2 Objec i es ................................................................................................... 3
1.3 Disse a ion O ganiza ion ......................................................................... 3
2. LITERATURE REVIEW ............................................................................... 4
2.1 Iden i ica ion o T ee Species wi h High Resolu ion Ae ial Images ......... 4
2.1.1 Applica ions in Ag icul u e and C op Managemen ................................................................ 4
2.1.2 Applica ions in Fo es y Managemen .......................................................................................... 5
2.2 Applica ions o Objec De ec ion ................................................................ 6
2.3 Objec De ec ion wi h Con olu ional Neu al Ne wo k (CNN) ................. 7
2.4 Objec de ec ion wi h R-CNN .................................................................... 9
2.4.1 Region p oposals .................................................................................................................................... 9
2.4.2 Fea u e ex ac ion ............................................................................................................................... 10
2.4.3 Classi ie .................................................................................................................................................. 10
2.4.4 E olu ion o RCNN Family ............................................................................................................... 11
3. THEORETICAL BACKGROUND .............................................................. 12
3.1 Fas e RCNN A chi ec u e ...................................................................... 12
3.2 P e- ained CNN (Fea u e Ex ac o ) ..................................................... 12
3.2.1 Incep ion V2 ........................................................................................................................................... 13
3.2.2 Resne 50 and Resne 101 ............................................................................................................... 14
3.2.3 Incep ion esne V2 ............................................................................................................................ 15
ix
3.3 Region P oposal Ne wo k (RPN) ............................................................. 16
3.3.1 Ancho s .................................................................................................................................................... 17
3.4 Region o In e es (RoI) Pooling .............................................................. 17
3.5 RCNN ....................................................................................................... 18
3.6 T aining .................................................................................................... 18
4. DATA AND METHOD ................................................................................ 20
4.1 Da a Desc ip ion ...................................................................................... 20
4.2 Me hod ..................................................................................................... 21
4.2.1 Tools .......................................................................................................................................................... 22
4.3 P e-p ocessing & labelling ....................................................................... 23
4.4 T aining objec de ec ion models ............................................................. 24
4.5 Tes ing ained models ............................................................................ 24
4.6 Accu acy assessmen ............................................................................... 25
5. RESULTS .................................................................................................... 26
5.1 Image p e-p ocessing & labelling ............................................................ 26
5.2 T aining objec de ec ion models ............................................................. 28
5.3 Tes ing ained models ............................................................................ 28
5.4 Accu acy assessmen ............................................................................... 30
6. DISCUSSION .............................................................................................. 33
7. CONCLUSIONS .......................................................................................... 37
BIBLIOGRAPHIC REFERENCES ...................................................................... 38
ANNEX I ............................................................................................................... 42
ANNEX II ............................................................................................................. 46
Chap e 2 | Li e a u e Re iew
5
images, high accu acy may be achie ed by inco po a ing all bands in he
classi ica ion (Nomu a, 2018).
Mo e c owded o e lapping oil plam ee c owns can be de ec ed and
coun ed wi h high esolu ion mul ispec al sa elli e images by aining a CNN
using housands o manually labelled samples (Li e al., 2017). In Li e al. (2017),
a ound 9000 samples we e used o achie e 96% accu acy o de ec ion wi h
op imum pa ame e se ings as numbe o ke nal in wo con olu ion laye s as 30
and 55 and numbe o hidden uni s in ully-connec ed laye as 600. In Csillik e
al. (2018), simila accu acy was achie ed in iden i ying ci us ees by aining
CNN using housands o aining samples de i ed om UAV mul ispec al images.
Classi ica ion accu acy can be imp o ed by applying objec -based pos
p ocessing wi h esul s om CNN (Csillik e al., 2018).
In Oli a es (2019), i is p oposed a me hod o classi y UAV images o
de ec palm ees wi h educed numbe o aining samples using p e- ained
CNN. Howe e , due o he inpu size cons ain in CNN based ea u e ex ac o he
de ec ion o indu ial palm ees we e no possible (Oli a es, 2019). The model
showed less pe o mance in iden i ying a ge ee when i is su ounded by
o he ee ypes.
2.1.2 Applica ions in Fo es y Managemen
Remo e sensing akes ad an age when disc imina ing ee species om
ae ial images om he ac ha ees sha e unique spec al signa u e depending
on he ype o species hey belong o (Lisein e al., 2015). Howe e , spec al
signa u e wi hin and be ween species changes empo ally. In Lisein e al. (2015),
i was disco e ed ha he op imal phenological ime window o disc imina ing
b oadlea ed ees lies be ween la e sp ing and ea ly summe ( he end o lea
lushing). Du ing his pe iod, he in a-spec al a ia ion wi hin ee species
minimizes while maximizing in e -spec al a ia ion be ween species (Lisein e
al., 2015). In Lisein e al. (2015), ee c owns we e manually delinea ed wi h help

Chap e 2 | Li e a u e Re iew
6
o ield mapping o inpu hem in au oma ic objec -based supe ised RF classi ie
o iden i y i e ee species based on spec al a ia ion.
In Onishi and Ise (2018), i has been exposed in hei s udy ha deep
lea ning can dis inguish se en ee species wi h 89.0 % accu acy using indi idual
ee c owns segmen ed om basic RGB UAV image o o es . This me hod has
bene i s in bo h pe o mance-wise and cos -wise o e p e ious me hods which
used expensi e mul ispec al images (Onishi and Ise, 2018). Howe e , his
model’s accu acy depends highly on he numbe o aining samples pe each
class and he quali y o ee c own segmen a ion. Al hough DEM and slope
models ha e been inco po a ed wi h inc eased numbe o aining samples,
misclassi ica ions ha e esul ed due o impe ec segmen a ion.
In Sa ono a e al. (2019), i has also been p o ed ha CNN based
classi ica ion models ha e po en ial o iden i y ou damage s ages o Fi ees
wi h an accu acy o 99.7% based on he shape, ex u e and colou o ee c own
in UAV RGB images o mixed o es . Inc easing he olume o samples by da a
augmen a ion echniques can subs an ially imp o e he accu acy o classi ica ion
when he e is ela i ely small aining da ase (Sa ono a e al., 2019). Al hough
candida e selec ion echnique was adop ed o ind egions o po en ial c owns in
he image be o e eeding hem in CNN, some misclassi ica ions we e esul ed.
In ac , iden i ying8 indi idual ee c owns using only RGB bands is a
di icul asks specially in dense o ess s which may be equi ing mo e
in o ma ion such as mul ispec al bands, NDVI o o he spec al indices and 3D
LIDAR da a (Sa ono a e al., 2019). Combining small amoun o hand anno a ed
high quali y ee c won aining da a wi h a la ge amoun o ee c owns au o-
gene a ed om LIDAR da a by unsupe ised algo i hm can be used o ain CNN
o iden i y indi idual ees (Weins ein e al., 2019).
2.2 Applica ions o Objec De ec ion
Objec de ec ion is a combina ion o compu e ision asks image
classi ica ion and objec localiza ion o classi y and loca e he p esence o objec s
Chap e 2 | Li e a u e Re iew
7
in an image (B ownlee, 2019). Image classi ica ion comp ises o algo i hms o
p edic he class o one objec in an image. Objec localiza ion comp ises o
algo i hms o loca e one o mo e objec s p esen in an image by d awing
bounding boxes a ound hem. Acco dingly, objec de ec ion comp ises o
algo i hms o loca e he p esence o one o mo e objec s in an image by d awing
bounding boxes wi h labels indica ing hei class o ype o ca ego y. These h ee
compu e ision asks a e gene ally e e ed o as objec ecogni ion (B ownlee,
2019).
Applica ions o objec de ec ion echniques a e isible in a ious indus ies
such as ehicle de ec ion and coun ing in anspo a ion indus y, building
de ec ion in u ban planning, ace de ec ion and people coun ing in secu i y
pu poses and animal moni o ing and coun ing in li es ock managemen .
Al hough ML is widely using in ag icul u e and o es y indus y, a e y ew
s udies ha e been ocused on using objec de ec ion. ML applica ions in c op
managemen a e ound in yield p edic ion, disease de ec ion, weed de ec ion and
species ecogni ion (Liakos e al., 2018).
In A seno ic e al. (2019), objec de ec ion was applied o iden i y diseased
lea es in high esolu ion images p o ing i s capabili y o he ask o e adi ional
ML echniques. P e- ained objec de ec ion models such as Fas e RCNN, SSD
and YOLO 3 can success ully de ec diseased lea es e en in complex
backg ounds a high accu acy (A seno ic e al., 2019).
2.3 Objec De ec ion wi h Con olu ional Neu al Ne wo k
(CNN)
The mos s aigh o wa d app oach is de ec ing objec s using CNN which is
widely used deep lea ning algo i hm in image classi ica ion (Sha ma, 2018). The
ollowing sec ion explains he CNN based on he S an o d Uni e si y lec u e
se ies on CNNs o Visual Recogni ion (Ka pa hy, 2019).
CNN di e s om egula neu al ne wo ks as i explici ly assumes ha he
inpu s a e images. In his way laye s in CNN ha e neu ons a anged in h ee
Chap e 2 | Li e a u e Re iew
8
dimensions (wid h, heigh , dep h) which a e ela ed o he dimensions o he
image (wid h, heigh , numbe o bands). Simila ly, he inpu laye o he CNN has
he same dimensions as he inpu image whe e wid h and heigh a e equal o he
size o he image and dep h is he numbe o bands in he image (Figu e 1).
CNN a chi ec u e consis s o h ee ypes o laye s; con olu ional laye s,
pooling laye and ully connec ed laye . Con olu ional laye s ha e small spa ial
il e s which con ol e ac oss he wid h and heigh o inpu olume compu ing he
do p oduc be ween pixel alues o he il e and inpu image a each posi ion
du ing he o wa d pass esul ing a wo-dimensional ac i a ion map o each
il e . This il e shi s o e he o iginal image by ce ain numbe o pixels called
s ide. When he s ide is one, hen he il e shi s by one pixel a a ime. A e
e e y con olu ional ope a ion, non-linea ope a ion called Rec i ied Linea Uni
(ReLU) is in oduced o eplace nega i es alues in he ea u e map o ze o.
Pooling laye s a e inse ed in be ween con olu ional laye s o educe he spa ial
size and compu a ional complexi y in he ne wo k. This is gene ally done by
sliding a il e ac oss he wid h and heigh o he inpu while aking he maximum
wi hin he il e which is called as Max pooling. Fully connec ed laye has all
neu ons ully connec ed o he p e ious laye . I is a Mul i-Laye Pe cep on using
a so max ac i a ion unc ion which keeps he ou pu alue anging om 0 o 1.
Once an image is passed h ough con olu ion and pooling laye s o CNN, i
p edic s he class o he objec in he image as ou pu . O iginal image is di ided
in o small ile and each ile is ed o CNN so ha i p edic s he class o each inpu
ile. La e , classi ied iles a e combined o ob ain he classes o all objec s in he
Figu e 1: Laye a angemen o neu al ne wo k (Ka pa hy, 2019)
Le : laye a angemen o egula neu al ne wo k. Righ : laye a angemen o CNN.
Chap e 2 | Li e a u e Re iew
9
o iginal image (Sha ma, 2018). This kind o s aigh o wa d applica ion o CNN
equi es all objec s o sha e a common aspec a io (Gi shick e al., 2016). Hence,
his app oach con ains p oblems as objec s in an image has di e en aspec a ios
and spa ial loca ions. I is equi ed o u ilize housands o such iles o o e come
ha p oblem esul ing mo e compu a ional ime (Sha ma, 2018). In Gi shick e
al. (2016), i is p oposed a me hod called RCNN o add ess he abo e p oblems.
2.4 Objec de ec ion wi h R-CNN
Following sec ion desc ibes he a chi ec u e o objec de ec ion sys em
p oposed by Gi shick e al., 2016. I has been ecognized as he i s and mos
success ul CNN app oach add essing p oblems ela ed o objec localiza ion,
de ec ion and segmen a ion (B ownlee, 2019). RCNN consis s o h ee modules.
They a e egion p oposal, ea u e ex ac o and classi ie . Region p oposals
gene a e ca ego y independen egions o conside o ea u e ex ac ion.
Fea u e ex ac o ex ac ixed leng h ec o om each egion using CNN. Thi d
module classi ies each egion in o one o he known classes using linea SVM
model. The Figu e 2 illus a es he o e iew o objec de ec ion sys em p oposed
by Gi shick e al., 2016.
2.4.1 Region p oposals
In RCNN, ca ego y independen egion p oposals a e gene a ed using
selec i e sea ch me hod (Gi shick e al., 2016). Region p oposals a e bounding
Figu e 2: O e iew o objec de ec ion sys em (Gi shick e al., 2016)
(a) I akes image as inpu . (b) a ound 2000 egion p oposals a e gene a ed. (c) ea u es a e compu ed o
each p oposal egion using CNN. (d) Each egion is classi ied using class speci ic linea SVM model
(a)
(b)
(c)
(d)
Chap e 2 | Li e a u e Re iew
10
boxes ep esen ing po en ial objec s in an image. Selec i e sea ch echnique
p oposes Region o In e es (ROI) by ecognizing pa e ns based on he a ying
scales, colou s, ex u es, and enclosu e in he image (Sha ma, 2018). Image is
ini ially di ided in o small segmen s and combined hem o ob ain la ge segmen s
aking in o conside a ion o simila i ies o scale, colou and ex u e e c (Sha ma,
2018).
2.4.2 Fea u e ex ac ion
Fea u es a e he c ucial ac o in ML modelling echniques which a e
esponsible o gene a ing e ec i e solu ion when as much as in o ma ion is
ex ac ed om a ge ed da ase (Dey, 2018). Gene ally, CNN ha ing a chi ec u e
o con olu ional and pooling laye s ac s as ea u e ex ac o s which can be used
o ex ac ea u es be o e in oducing hem in o ML models such as SVM, RF, e c.
(Pe one, 2015). Mos cases, p e- ained CNNs a e u ilized as ea u e ex ac o s
by emo ing las ou pu laye whe e he p ocess is called ans e lea ning
(Pe one, 2015). The ea u es ex ac ed by p e- ained CNN ha e po en ial o ain
ML models o pe o m classi ica ion on e y high esolu ion images (Oli a es,
2019).
In RCNN, ixed leng h ea u e ec o s a e ex ac ed om each egion
p oposal using CNN de eloped by K izhe sky, Su ske e and Hin on, 2012. CNN
accep s ixed size o images as inpu (Gi shick e al., 2016). The e o e, pixels
su ounded by bounding box o egion p oposal a e wa ped o ge 227 x 227 pixel
size image be o e eeding hem in o CNN. Inpu images a e passed h ough i e
con olu ional laye s and wo ully connec ed laye s.
2.4.3 Classi ie
A e ex ac ing ea u es, egions a e classi ied in o known classes using
linea SVM. One SVM is ained o each known class by applying ex ac ed
ea u es wi h aining labels du ing he aining phase. The pe o mance o he
model becomes highe when in ol ing SVM o classi ica ion a he han ge ing
he ou pu om las laye o ine- uned CNN (Gi shick e al., 2016).

Chap e 2 | Li e a u e Re iew
11
2.4.4 E olu ion o RCNN Family
Howe e , RCNN possesses ew d awbacks such as mul i-s age aining
p ocess equi ing ope a ion o sepa a e models, equi ing a s o age o hund eds
o gigaby es and slow objec de ec ion due o he ex ac ion o ea u es om each
egion p oposal in image (Gi shick, 2015). Fas RCNN was in oduced o add ess
hose issues by educing aining s ages o single pipeline. Al hough, i was as
han RCNN i s ill equi es CNN o pass h ough egion p oposals o each image
(Gi shick, 2015). Fas e RCNN which is desc ibed in nex chap e , is a u he
imp o ed e sion o ob ain as and accu a e de ec ion model.
Chap e 3 | Theo e ical Backg ound
12
3. THEORETICAL BACKGROUND
P e ious chap e p o ides an in oduc ion o RCNN. This chap e ex ensi ely
explains he as e RCNN a chi ec u e and ea u e ex ac o s used in he me hod
p oposed in his hesis.
3.1 Fas e RCNN A chi ec u e
Fas e RCNN is an in eg a ed ne wo k which sha es a deep ully
con olu ional ne wo k known as RPN wi h s a e-o - he-a objec de ec ion
ne wo k known as Fas RCNN (Ren e al., 2017). I comp ises o ou modules
namely P e- ained CNN, RPN, ROI pooling and RCNN as illus a ed in Figu e 3.
Figu e 3: Fas e RCNN a chi ec u e (Rey, 2018)
Fi s s ep is o gene a e a con olu ional ea u e map om a enso
(mul idimensional a ay) o inpu image which is ed in o a p e- ained CNN. In
RPN, p ede ined numbe o egions which con ain objec s a e p oposed by using
ixed sized e e ence bounding boxes called ancho s which a e placed on ea u e
map. RoI pooling is applied o ex ac ea u es o ele an objec s (bounding
boxes p oposed by RPN) om ea u e map compu ed by p e- ained CNN. RCNN
classi ies he objec in o a class and adjus he bounding box coo dina es (Rey,
2018).
3.2 P e- ained CNN (Fea u e Ex ac o )
Fas e RCNN o iginally used ou pu o an in e media e laye o VGG which
is a CNN ained o classi y ImageNe da ase o ex ac ea u es om inpu image.
I akes he ou pu o con olu ional laye s by lea ning edges, pa e ns and shapes
o objec s esul ing a con olu ional ea u e map o inpu image which has smalle
Chap e 3 | Theo e ical Backg ound
13
spa ial dimensions wi h g ea e dep h han o iginal image (Rey, 2018). Al hough
ZF and VGG we e conside ed as deep ne wo ks, much deepe ne wo ks ha e been
in en ed a e hem. Such ne wo ks known as Incep ion 2, Resne 50, Resne
101 and Incep ion esne 2 which we e used in he s udy a e explained below.
3.2.1 Incep ion V2
Incep ion 2 was p esen ed by imp o ing p e ious e sion o inc ease
accu acy and educe compu a ional complexi y. Accu acy has been imp o ed by
a oiding dimension educ ion o inpu which may lead o loss o in o ma ion and
compu a ional complexi y has been educed by ac o izing 5x5 con olu ion laye
o wo 3x3 con olu ion laye s (Szegedy e al., 2016). The laye a angemen o
incep ion module is illus a ed in Figu e 4.
Figu e 4: Con olu ional laye a angemen o incep ion module (Szegedy e al., 2016)
Incep ion 2 consis s o se e al con olu ional and pooling laye s wi h a
so max laye o he inal classi ica ion (Io e and Szegedy, 2015). In as e RCNN
incep ion 2 model, end poin has been se o incep ion (4e) laye o incep ion
2 o ac as a ea u e ex ac o . A chi ec u e o incep ion 2 ea u e ex ac o
used in he expe imen is illus a ed in Table 1 assuming inpu is ha ing spa ial
dimensions o 224x224.
Chap e 3 | Theo e ical Backg ound
14
Type
Pa ch Size/ S ide
Ou pu Size
Con olu ion
7×7/2
112×112×64
Max Pool
3×3/2
56×56×64
Con olu ion
3×3/1
56×56×192
Max Pool
3×3/2
28×28×192
Incep ion (3a)
28×28×256
Incep ion (3b)
28×28×320
Incep ion (3c)
S ide 2
28×28×576
Incep ion (4a)
14×14×576
Incep ion (4b)
14×14×576
Incep ion (4c)
14×14×576
Incep ion (4d)
14×14×576
Incep ion (4e)
S ide 2
14×14×1024
Table 1: A chi ec u e o Incep ion 2 ea u e ex ac o (Io e and Szegedy, 2015)
3.2.2 Resne 50 and Resne 101
Residual ne wo ks ha e been de eloped by inse ing sho cu connec ions
o he plain ne wo ks which we e inspi ed by VGG ne s (He, 2015). Sho cu
connec ions a e inse ed a each block as illus a ed in Figu e 5. Resne 50 and
Resne 101 a e buil by block o laye s consis ing o 1x1, 3x3 and 1x1
con olu ional laye s. De ail a chi ec u e o Resne 50 and Resne 101 wi h
numbe o blocks a e summa ized in Table 2.
Figu e 5: A block o 3 laye s wi h sho cu connec ion (He, 2015).
Chap e 4 | Da a and Me hod
21
Da ase 3: The da ase consis s o one RGB image co e ing a ea be ween
Mangapwani and Bumbwini illages on Tanzanian island o Unguja, he main
island o Zanziba . I has been cap u ed using UAV pla o m on 17 No embe
2016. The image is ha ing dimensions o 38107 x 42858 pixels a a spa ial
esolu ion o 7 cm. The image comp ises o a eas whe e he palm ees a e
spa sely loca ed as well as densely loca ed wi h he p esence o o he ees,
g asslands, ba e land, oads and houses.
4.2 Me hod
This hesis p oposed a me hod o iden i y and loca e palm ees by use o
RCNN. The en i e me hod consis s o ou majo s eps; image p e-p ocessing and
labelling, aining objec de ec ion models o hesis da ase s, es ing ained
models and accu acy assessmen . Figu e 10 illus a es comple e p ocess in s ep
by s ep.
Figu e 10: Me hodological Flowcha

Chap e 4 | Da a and Me hod
22
Du ing p e-p ocessing, o iginal images we e subjec ed o o ma
con e sion, c opping and objec s labelling. Fou p e- ained objec de ec ion
models which a e being ained o iden i y a ious objec s we e selec ed om
Tenso Flow objec de ec ion model zoo. They we e ained using p e-p ocessed
hesis da ase s. Accu acy assessmen was conduc ed by compa ing de ec ion
esul s om ained models. All s eps we e epea ed o da ase 1, da ase 2 and
da ase 3 sepa a ely. De ail explana ion o each s ep is w i en in sec ion 4.3 o
sec ion 4.6.
4.2.1 Tools
This sec ion lis s ou all he so wa e and ha dwa e which we e u ilized o
implemen he me hod p oposed in sec ion 4.2. The en i e me hodology was
de eloped using open sou ce so wa e and packages.
• Anaconda is an open sou ce dis ibu ion o py hon aiming o p o ide
a ious packages and package managemen ools o scien i ic compu ing
in da a science and machine lea ning. Anaconda e sion 2019.07 was
ins alled o manage p og amming en i onmen o he p ojec .
• Py hon is high le el p og amming language which suppo s p ocedu al,
objec -o ien ed and unc ional p og amming allowing use s o w i e easy
and logical codes. Py hon e sion 3.7.4 was ins alled inside Anaconda
en i onmen .
• Tenso Flow is an end- o-end open sou ce lib a y o di e en iable
p og amming le ing use s o de elop machine lea ning applica ions. I
can be deployed in pla o ms like CPUs, GPUs, and TPUs. Tenso Flow CPU
e sion 1.14.0 was ins alled inside Anaconda en i onmen .
• Tenso Flow Objec de ec ion API is an open sou ce amewo k
p o iding explici ly w i en collec ion o codes o build, ain and deploy
objec de ec ion models.
Chap e 4 | Da a and Me hod
23
• LabelImg was used o label objec s on image iles. I is a g aphical image
anno a ion ool w i en in Py hon. Anno a ions a e w i en in o XML iles
in PASCAL VOC o ma used by Imagene .
• Jupy e No ebook is an open-sou ce web applica ion ha allows use s
in e ac i e p og amming and isualiza ion o scien i ic compu ing and
da a science applica ions. Jupy e No ebook e sion 6.0.1 was ins alled
inside Anaconda en i onmen .
• All p ocesses we e un on a lap op wi h an In el i7-8550U CPU @ 1.99 GHz
p ocesso , 8 GB RAM and windows 10 64-bi ope a ing sys em.
4.3 P e-p ocessing & labelling
The pu pose o his s ep was o con e he o iginal images in o a o m ha
is compa ible wi h eeding hem as inpu o p e- ained objec de ec ion models.
Tenso Flow objec de ec ion models only accep images in PNG/JPG o ma wi h
h ee bands i.e. RGB (Huang e al., 2017). Since he o iginal images we e in TIF
o ma , hey we e con e ed o JPG o ma using a py hon sc ip which is a ached
in Annex II.
The e is no size limi o he inpu images o Tenso Flow objec de ec ion
models since only in e media e laye (con olu ion laye s) o p e- ained CNN
ea u e ex ac o a e being used (Rey, 2018). Howe e , inpu images a e esized
o 600 x 1024 pixels in as e RCNN models keeping he aspec a io o images
cons an o a oid he memo y issues may a ise du ing p ocess (Ren e al., 2017).
This size has been es ed o NVidia Keple GPU (Szegedy e al., 2016).
Ne e heless, as e RCNN expec s objec s g ea e han 30 x 30 pixels. Since he
a e age size o a ge ed objec s is a ound 100 x 100 pixels, o iginal images we e
c opped in o iles o 1000 x 1000 pixels o keep he size o objec s g ea e han
30 x 30 pixels a e esizing. Tiles which a e no ha ing palm ees we e
disca ded. 10% o da ase was sepa a ed as es images and emaining da ase
was spli in o 70% as ain images and 30% as alida ion images.
Chap e 4 | Da a and Me hod
24
P epa a ion o da ase s o aining: All objec s (Palm ees) in ain and
alida ion images we e labelled by d awing bounding boxes using LabelImg.
Coo dina es o bounding boxes and class we e eco ded in XML ile o each
image in he da ase s. XML iles oge he wi h co esponding images we e
con e ed o TFReco d o ma and gene a ed ain. eco d and alida ion. eco d
iles in o de o eed hem objec de ec ion models.
P epa a ion o da ase s o accu acy assessmen : All objec s (Palm ees) in
es images we e labelled by d awing bounding boxes using LabelImg and
coo dina es we e eco ded in o XML iles. XML iles we e con e ed o CSV iles
and sa ed as posi i e g ound u h boxes. Likewise, nega i e g ound u h boxes
we e c ea ed by d awing bounding boxes in which palm ees we e no
p esen ed.
4.4 T aining objec de ec ion models
Fou objec de ec ion models downloaded om Tenso Flow objec
de ec ion model zoo namely as e _ cnn_incep ion_ 2, as e _ cnn_ esne 50,
as e _ cnn_ esne 101, as e _ cnn_incep ion_ esne _ 2 we e ained p o iding
ain. eco d and alida ion. eco d p epa ed in sec ion 4.3 as inpu s. T aining
p ocess was moni o ed using Tenso boa d and e mina ed when he o al loss
eached alue a ound 0.1. In e ence g aph o he model was expo ed a e
aining was comple ed. This g aph con ains he weigh s o he ained model.
A chi ec u e o each model is desc ibed in chap e 3.
4.5 Tes ing ained models
Tes images we e ed in o ained model esul ed in sec ion 4.4. De ec ion
esul s we e ob ained in o images in JPG o ma and coo dina es o de ec ion
boxes we e eco ded in CSV ile pe each es image in da ase . The py hon sc ip
was un in Jupy e No ebook o es models and is a ached in Annex II.
Chap e 4 | Da a and Me hod
25
4.6 Accu acy assessmen
De ec ion boxes we e compa ed wi h co esponding posi i e g ound u h
boxes and nega i e g ound u h boxes c ea ed as explained in sec ion 4.3 o
de i e me ices; p ecision, Sensi i i y, speci ici y and accu acy. I was based on
he IoU calcula ed be ween de ec ion boxes and g ound u h boxes. Following
de ini ions we e de ined o he calcula ion o me ices;
T ue posi i es (TP): de ec ion boxes ha ing IoU wi h posi i e g ound u h
boxes > 0.5
False posi i es (FP): de ec ion boxes ha ing IoU ≤ 0.5 o no in e sec ion
wi h posi i e g ound u h boxes
False nega i es (FN): igno ed g ound u h boxes
T ue nega i es (TN): nega i e g ound u h boxes ha ing IoU ≤ 0.5 o no
in e sec ion wi h de ec ion boxes
Following o mulae we e conside ed when compu ing me ices.
P ecision = TP
TP + FP × 100 = all posi i es co ec ly de ec ed by model
all posi i es de ec ed by model
Sensi i i y = TP
TP + FN × 100 = all posi i es co ec ly de ec ed by model
all posi i es in ac ual
Speci ici y = TN
TN + FP × 100 = all nega i es co ec ly de ec ed by model
all nega i es in ac ual
Accu acy = TP + TN
TP + FP + TN + FN × 100 = all co ec ly de ec ed by model
all in ac ual
Chap e 5 | Resul s
26
5. RESULTS
This chap e consis s o he ou pu s esul ed om chap e 4. Only selec ed
images a e included in his chap e whe e app op ia e. All o he esul s a e
a ached in Annex I.
5.1 Image p e-p ocessing & labelling
Da ase 1 consis ed o 49 iles a e c opping in o 1000 x 1000 pixels o
ele en RGB images o 3000 x 4000 pixels. Da ase 2 consis ed o 283 iles a e
c opping in o 1000 x 1000 pixels o RGB image o 17761 x 25006 pixels. Da ase
3 consis ed o 432 iles a e c opping in o 1000 x 1000 pixels o RGB image o
38107 x 42858 pixels. Numbe o images a e spli ing da ase in o ain,
alida ion and es da a is summa ized in Table 3. Example o c opped image o
each da ase is shown in Figu e 11.
Da a
T ain
Valida ion
Tes
Da ase 1
31 images
12 images
6 images
Da ase 2
179 images
76 images
28 images
Da ase 3
276 images
118 images
38 images
Table 3: Numbe o images in ain, alida ion and es da a o each da ase
Figu e 11: Examples o c opped images.
(a) example o c opped images om da ase 1, (b) example o c opped images om da ase 2, (c) example
o c opped images om da ase 3
(a)
(b)
(c)

Chap e 5 | Resul s
27
Example o ec angula bounding boxes d awn using LabelImg is shown in Figu e
12. Coo dina es o each box eco ded in XML ile a e in o de o xmin, ymin, xmax
and ymax.
Figu e 12: Example o labelled image om da ase 1
(a) Example o labelled image om da ase 1, (b) showing a bounding box wi h coo dina es
Labels we e sa ed in XML ile which con ains coo dina es o boxes wi h
class. XML iles we e con e ed o CSV ile o ma . Numbe o bounding boxes
ex ac ed a e indica ed in Table 4. The gene a ed TFReco d iles om CSV iles
and co esponding images comp ise o numpy a ay o image wi h co esponding
bounding box coo dina es.
Da a
T ain
Valida ion
Tes
Palm ee
Non- palm ee
Da ase 1
1996
765
441
101
Da ase 2
6121
2743
1240
517
Da ase 3
7158
3056
1433
678
Table 4: Numbe o bounding boxes ex ac ed om c opped images
(xmin, ymin)
(xmax, ymax)
(a)
(b)
Chap e 5 | Resul s
28
5.2 T aining objec de ec ion models
T aining p ocess was e mina ed when he o al loss becomes close o 0.1.
To al loss s s eps g aph o as e _ cnn_incep ion_ 2 model o Da ase 1 is
shown in Figu e 13. To al ime du a ions including aining and alida ion o
each model wi h h ee da ase s we e eco ded in Tenso boa d and a e
summa ized in Table 5. Fo da ase 1, cnn_incep ion 2 model was he as es
and cnn_ esne 50 was sligh ly slowe . Fo da ase 2, cnn_ esne 50 model was
as e han o he models. Fo da ase 3, cnn_incep ion 2 model was he as es .
The cnn_incep ion_ esne 2 model had he lowes speed o all h ee da ase s.
O e all, cnn_incep ion 2 was he as es model among o he s.
Model
Da ase 1
Da ase 2
Da ase 3
cnn_incep ion 2
19m 57s
29m 44s
9m 57s
cnn_ esne 50
20m 1s
19m 49s
19m 51s
cnn_ esne 101
50m 0s
59m 38s
29m 40s
cnn_incep ion_ esne 2
5h 39m 15s
5h 0m 19s
5h 22m 51s
Table 5: To al ime du a ion o aining and alida ion o each model
5.3 Tes ing ained models
T ained model ou pu s each inpu image in which all de ec ions a e ma ked
by ec angles wi h he name o class ha objec belongs o and con idence ha
S eps
To al loss
Figu e 13: To al loss s s eps g aph o as e _ cnn_incep ion_ 2 model o Da ase 1
Chap e 5 | Resul s
29
objec belongs o iden i ied class. Images in Tes olde o each da ase we e used
o isualize he pe o mance o each model. Such ou pu images om
as e _ cnn_incep ion_ 2 model pe da ase a e shown in Figu e 14. I can be seen
ha in Figu e 14 (a), model was no capable o iden i ying all he objec s (Palm
ees) in he image. Figu e 14 (b) is an example image whe e shadows o ees
ha e been ma ked as Palm ees.
Figu e 14: Classi ied images om as e _ cnn_incep ion_ 2 model
(a) One o he classi ied images o Da ase 1; (b) One o he classi ied images o Da ase 2; (c) One o he
classi ied images o Da ase 3
(a)
(b)
(c)
Chap e 5 | Resul s
30
5.4 Accu acy assessmen
Images in he Tes olde o each da ase we e used o e alua e each model
by compu ing p ecision, Sensi i i y, speci ici y and accu acy. To al numbe o
bounding boxes manually ex ac ed o u ilize as posi i e (palm ees) and
nega i e (Non-palm ees) g ound u h boxes a e summa ized in Table 4. Figu e
15 (a) is an example image showing e alua ion esul s o one o he images in Tes
olde o da ase 1. Simila ly, Figu e 15 (b) and (c) illus a e examples o da ase
2 and 3 espec i ely.
Figu e 15: E alua ion esul s o classi ied images om as e _ cnn_incep ion_ 2 model
(a) E alua ion esul s o one o he classi ied images o Da ase 1; (b) E alua ion esul s o one o he
classi ied images o Da ase 2; (c) E alua ion esul s o one o he classi ied images o Da ase 3
(a)
(b)
(c)
T ue posi i es
False posi i es
False nega i es
T ue nega i es
Legend
Chap e 7 | Conclusion
37
7. CONCLUSIONS
C op and o es y managemen ields o en ge bene i s om a as ange
o emo e sensing applica ions o main ain sus ainable managemen . Image
classi ica ion is one o he impo an asks which e eals mo e in o ma ion o
decision making. Image classi ica ion echniques a e de eloping apidly wi h he
ad ancemen o deep lea ning. Al hough, a conside able numbe o esea ches
ha e been ca ied ou using CNN in his ield and p o en i s capabili ies, e y ew
esea ches ha e been conduc ed o assess he pe o mance o ecen ly de eloped
objec de ec ion echnique RCNN in ag icul u al ield. This hesis e alua es he
pe o mance o as e RCNN in iden i ying palm ees which may lead o many
applica ions in c op managemen ield. I is also an example o showing he
applica ion o RCNN o de ec indi idual ees wi h a smalle numbe o aining
samples.
I is e iden ha p e- ained as e RCNN objec de ec ion models can
achie e be e accu acies wi h only high esolu ion RGB images. Fas e RCNN
models can pe o m well in iden i ying a ge e en in complex backg ounds.
Objec de ec ion models beha e di e en ly wi h di e en da ase s. The speed and
pe o mance o models depend on he da ase . I also depends on he ype o
ea u e ex ac o . Acco ding he esea ch indings, he as e RCNN model which
was ha ing incep ion 2 as ea u e ex ac o pe o med bes by achie ing he
highes o e all accu acy and speed. I also well dis inguishes shadows om a ge
when compa ed o o he models. The as e RCNN model which was ha ing
incep ion esne 2 as ea u e ex ac o showed poo pe o mance wi h lowes
o e all accu acy and speed.

38
BIBLIOGRAPHIC REFERENCES
ARSENOVIC, M. e al. (2019) ‘Sol ing Cu en Limi a ions o Deep Lea ning Based
App oaches o Plan Disease De ec ion’, Symme y, 11(7), p. 939. doi:
10.3390/sym11070939.
BROWNLEE, J. (2019) A Gen le In oduc ion o Objec Recogni ion Wi h Deep
Lea ning, Deep Lea ning o Compu e Vision. A ailable a :
h ps://machinelea ningmas e y.com/objec - ecogni ion-wi h-deeplea ning
/ (Accessed: 17 Sep embe 2019).
CHIANG, S. H., VALDEZ, M. and CHEN, C. F. (2016) ‘Fo es ee species dis ibu ion
mapping using Landsa sa elli e image y and opog aphic a iables wi h he
Maximum En opy me hod in Mongolia’, In e na ional A chi es o he
Pho og amme y, Remo e Sensing and Spa ial In o ma ion Sciences - ISPRS
A chi es, 41(July), pp. 593–596. doi: 10.5194/isp sa chi es-XLI-B8-593-
2016.
CSILLIK, O. e al. (2018) ‘Iden i ica ion o Ci us T ees om Unmanned Ae ial
Vehicle Image y Using Con olu ional Neu al Ne wo ks’, D ones, 2(4), p. 39.
doi: 10.3390/d ones2040039.
DEY, S. (2018) CNN applica ion on s uc u ed da a-Au oma ed Fea u e Ex ac ion.
A ailable a : h ps:// owa dsda ascience.com/cnn-applica ion-ons uc u ed
-da a-au oma ed- ea u e-ex ac ion-8 2cd28d9a7e (Accessed: 6 Sep embe
2019).
GIRSHICK, R. (2015) ‘Fas R-CNN’, P oceedings o he IEEE In e na ional Con e ence
on Compu e Vision, 2015 In e , pp. 1440–1448. doi: 10.1109/ICCV.2015.169.
GIRSHICK, R. e al. (2016) ‘Region-Based Con olu ional Ne wo ks o Accu a e
Objec De ec ion and Segmen a ion’, IEEE T ansac ions on Pa e n Analysis
and Machine In elligence, 38(1), pp. 142–158. doi: 10.1109/TPAMI.2015
.2437384.
GUIRADO, E. e al. (2017) ‘Deep-lea ning Ve sus OBIA o sca e ed sh ub de ec ion
wi h Google Ea h Image y: Ziziphus lo us as case s udy’, Remo e Sensing,
39
9(12). doi: 10.3390/ s9121220.
HE, K. (2015) ‘Deep Residual Lea ning o Image Recogni ion’, ILSVRC2015.
HUANG, J. e al. (2017) ‘Speed/accu acy ade-o s o mode n con olu ional objec
de ec o s’, P oceedings - 30 h IEEE Con e ence on Compu e Vision and Pa e n
Recogni ion, CVPR 2017, 2017-Janua, pp. 3296–3305. doi: 10.1109/CVPR.
2017.351.
HUMBOLDT STATE UNIVERSITY (2019) Objec Based Classi ica ion, HSU Geospa ial
Online. A ailable a : h p://gsp.humbold .edu/OLM/Cou ses/GSP
_216_Online/lesson61/objec .h ml (Accessed: 14 Sep embe 2019).
ID, J. C., LENG, W. and LIU, K. (2018) ‘Objec -Based Mang o e Species Classi ica ion
Using Unmanned Ae ial Vehicle Hype spec al Images and Digi al Su ace
Models’, Remo e Sensing. doi: 10.3390/ s10010089.
IOFFE, S. and SZEGEDY, C. (2015) ‘Ba ch no maliza ion: Accele a ing deep ne wo k
aining by educing in e nal co a ia e shi ’, 32nd In e na ional Con e ence
on Machine Lea ning, ICML 2015, 1, pp. 448–456.
KARPATHY, A. (2019) CS231n Con olu ional Neu al Ne wo ks o Visual Recogni ion.
A ailable a : h p://cs231n.gi hub.io/con olu ional-ne wo ks/ (Accessed: 3
Sep embe 2019).
KRIZHEVSKY, A., SUTSKEVER, I. and HINTON, G. E. (2012) ‘ImageNe Classi ica ion wi h
Deep Con olu ional Neu al Ne wo ks’, ILSVRC-2012. doi: 10.1201/97814200
10749.
LI, W. e al. (2017) ‘Deep lea ning based oil palm ee de ec ion and coun ing o
high- esolu ion emo e sensing images’, Remo e Sensing, 9(1). doi: 10.33
90/ s9010022.
LIAKOS, K. G. e al. (2018) ‘Machine lea ning in ag icul u e: A e iew’, Senso s
(Swi ze land), 18(8), pp. 1–29. doi: 10.3390/s18082674.
LISEIN, J. e al. (2015) ‘Disc imina ion o deciduous ee species om ime se ies
o unmanned ae ial sys em image y’, PLoS ONE, 10(11), pp. 1–20. doi: 10.13
71/jou nal.pone.0141006.
OLIVARES, R. J. L. (2019) PALM TREE IMAGE CLASSIFICATION A con olu ional and
40
machine lea ning app oach.
ONISHI, M. and ISE, T. (2018) ‘Au oma ic classi ica ion o ees using a UAV onboa d
came a and deep lea ning’. A ailable a : h p://a xi .o g/abs/1804 .10390.
PERKINS, R. (2016) Enginee s Teach Machines o Recognize T ee Species |
www.cal ech.edu. A ailable a : h ps://www.cal ech.edu/abou /news/engine
e s- each-machines- ecognize- ee-species-52122 (Accessed: 19 Sep embe
2019).
PERONE, C. S. (2015) Deep lea ning – Con olu ional neu al ne wo ks and ea u e
ex ac ion wi h Py hon | Te a Incogni a. A ailable a : h p://blog.ch is ian
pe one.com/2015/08/con olu ional-neu al-ne wo ks-and- ea u e-ex ac io
n-wi h-py hon/ (Accessed: 3 Sep embe 2019).
REN, S. e al. (2017) ‘Fas e R-CNN: Towa ds Real-Time Objec De ec ion wi h
Region P oposal Ne wo ks’, IEEE T ansac ions on Pa e n Analysis and
Machine In elligence, 39(6), pp. 1137–1149. doi: 10.1109/TPAMI.2016.2577
031.
REY, J. (2018) Fas e R-CNN: Down he abbi hole o mode n objec de ec ion,
T yolabs Blog. A ailable a : h ps:// yolabs.com/blog/2018/01/18/ as e - -
cnn-down- he- abbi -hole-o -mode n-objec -de ec ion/ (Accessed: 12 Oc ob
e 2019).
SAFONOVA, A. e al. (2019) ‘De ec ion o Fi T ees (Abies sibi ica) Damaged by he
Ba k Bee le in Unmanned Ae ial Vehicle Images wi h Deep Lea ning’, Remo e
Sensing, 11(6), p. 643. doi: 10.3390/ s11060643.
SHARMA, P. (2018) A S ep-by-S ep In oduc ion o he Basic Objec De ec ion
Algo i hms (Pa 1). A ailable a : h ps://www.analy ics idhya.com/blog/20
18/10/a-s ep-by-s ep-in oduc ion- o- he-basic-objec -de ec ion-
algo i hms-pa -1/ (Accessed: 19 Sep embe 2019).
SHETTY, B. (2018) ‘Supe ised Machine Lea ning: Classi ica ion – Towa ds Da a
Science’, Towa d Da a Science. A ailable a : h ps:// owa dsda ascience.com/
supe ised-machine-lea ning-classi ica ion-5e685 e18a6d.
STEELE, M. (2016) ‘Fo es Fac s’, Vi ginia Depa men o Fo es y. A ailable a :
41
h p://www.do . i ginia.go /edu/ esou ces-educa o s.h m.
SZEGEDY, C. e al. (2016) ‘Re hinking he Incep ion A chi ec u e o Compu e
Vision’, P oceedings o he IEEE Compu e Socie y Con e ence on Compu e
Vision and Pa e n Recogni ion, 2016-Decem, pp. 2818–2826. doi: 10.1109/CV
PR.2016.308.
SZEGEDY, C. e al. (2017) ‘Incep ion- 4, incep ion-ResNe and he impac o
esidual connec ions on lea ning’, 31s AAAI Con e ence on A i icial
In elligence, AAAI 2017, pp. 4278–4284.
TORAHI, A. A. and RAI, S. C. (2011) ‘Land Co e Classi ica ion and Fo es Change
Analysis, Using Sa elli e Image y-A Case S udy in Dehdez A ea o Zag os
Moun ain in I an’, Jou nal o Geog aphic In o ma ion Sys em, 03(01), pp. 1–11.
doi: 10.4236/jgis.2011.31001.
WEINSTEIN, B. G. e al. (2019) ‘Indi idual T ee-C own De ec ion in RGB Image y
Using Semi-Supe ised Deep Lea ning Neu al Ne wo ks’, pp. 1–13.
42
ANNEX I
This sec ion con ains esul s ha we e no p esen ed in he main ex . Only
he images o de ec ion and e alua ion esul s om as e _ cnn_incep ion_ 2
model o da ase 1 a e a ached he e o he e e ence. Resul s o all h ee
da ase s can be iewed h ough his link;
h ps://gi hub.com/Chamodi88/Mas e Thesis .
S eps
To al loss
(a)
To al loss
S eps
(b)
To al loss
(c)
S eps
(d)
S eps
To al loss
To al loss
(e)
S eps
To al loss
( )
S eps

43
(a) To al loss s s eps g aph o as e _ cnn_incep ion_ 2 model o Da ase 1; (b) To al loss s s eps
g aph o as e _ cnn_ esne 50 model o Da ase 1; (c) To al loss s s eps g aph o as e _ cnn_ esne 101
model o Da ase 1; (d) To al loss s s eps g aph o as e _ cnn_incep ion_ esne 2 model o Da ase 1; (e)
To al loss s s eps g aph o as e _ cnn_incep ion_ 2 model o Da ase 2; ( ) To al loss s s eps g aph o
as e _ cnn_ esne 50 model o Da ase 2; (g) To al loss s s eps g aph o as e _ cnn_ esne 101 model o
Da ase 2; (h) To al loss s s eps g aph o as e _ cnn_incep ion_ esne 2 model o Da ase 2; (i) To al loss
s s eps g aph o as e _ cnn_incep ion_ 2 model o Da ase 3; (j) To al loss s s eps g aph o
as e _ cnn_ esne 50 model o Da ase 3; (k) To al loss s s eps g aph o as e _ cnn_ esne 101 model o
Da ase 3; (m) To al loss s s eps g aph o as e _ cnn_incep ion_ esne 2 model o Da ase 3
To al loss
(g)
S eps
To al loss
(h)
S eps
To al loss
(i)
S eps
To al loss
(j)
S eps
To al loss
(k)
S eps
To al loss
(m)
S eps
Figu e I.I To al loss s s eps g aphs o each model
44
Figu e I.II Classi ied images om as e _ cnn_incep ion_ 2 model o es images in Da ase 1
(a)
(b)
(c)
(d)
(e)
( )
(a) Image 1; (b) Image 2; (c) Image 3; (d) Image 4; (e) Image 5; ( ) Image 6;
45
Figu e I.III E alua ion esul s o classi ica ion om as e _ cnn_incep ion_ 2 model o es images in
Da ase 1
(a)
(b)
(c)
(d)
(e)
( )
(a) Image 1; (b) Image 2; (c) Image 3; (d) Image 4; (e) Image 5; ( ) Image 6;
46
ANNEX II
This sec ion con ains sc ip s used o p e-p ocess da a and he sc ip used o
e alua e he model. All sc ip s ha e been uploaded in he ollowing link;
h ps://gi hub.com/Chamodi88/Mas e Thesis .
Figu e II.I The py hon sc ip used o con e ing image o ma om TIF o JPG
53
Figu e II.III The py hon sc ip used o de ec palm ees and e alua e he de ec ions

Guia pa a a o ma ação de eses Ve são 4.0 Janei o 2006