Spectral–Spatial Transformer-based Semantic Segmentation for Large-scale Mapping of Individual Date Palm Trees using Very High-resolution Satellite Data
Abstract
Date palm plantations in the United Arab Emirates (UAE) are under threat from soil salinity, drought, and date palm weevils. Accordingly, monitoring and conserving date palms are crucial to preserving a vital component of the country’s agricultural heritage, economy, food security, and ecological balance. Previous studies have effectively identified date palm trees using RGB-based aerial and UAV imagery utilizing diverse deep learning methods. However, the utilization of very high-resolution satellite data for delineating individual date palm crowns remains unexplored due to the limited spatial resolution capabilities of existing satellite systems. This study primarily aimed to achieve precise and comprehensive mapping of date palm trees using WorldView-3 (WV-3) satellite data by leveraging the high representational power of the state-of-the-art vision transformers (ViT) in capturing global information from the input data. First, an in-depth analysis assessment of the various transformer-based semantic segmentation architectures, including UperNet with vision transformer and Swin transformer, SegFormer, Mask2Former, and UniFormer, was conducted. Second, the integration of spectral data on the performance of ViTs was evaluated. Moreover, the models’ generalizability and complexity effect on the segmentation effectiveness were assessed. Accordingly, a postprocessing strategy was developed to aid in delineating and counting date palm trees from semantic segmentation outputs. Results demonstrated that integration of WV-3 spectral data into the analysis resulted in a marked improvement in segmentation quality. The UniFormer, UperNet-Swin, and Mask2Former models demonstrated considerable improvements in multispectral data analysis, with increases in mean intersection over union (mIoU) of 2.17% (77.88% mIoU, 86.01% mean F-score [mF-score]), 2% (78.10% mIoU, 86.18% mF-score), and 1.15% (77.36% mIoU, 85.59% mF-score), respectively, compared with their RGB-based results. Evaluations of model transferability also indicated that Mask2Former, UniFormer, and UperNet-Swin transformers efficiently adapted to multispectral data in the Dibba region. These models achieved mIoU scores of 84.36%, 84.25%, and 83.17% and mF-scores of 90.95%, 90.87%, and 90.13%, highlighting their effectiveness and potential for broader regional application. This research highlights the efficacy and feasibility of using ViTs with WV-3 multispectral data for accurate and comprehensive surveying of date palm plantations, enabling the development of palm tree inventories and continuously updating geospatial databases.
Full text
Ecological Indica o s 163 (2024) 112110
A ailable online 8 May 2024
1470-160X/© 2024 The Au ho s. Published by Else ie L d. This is an open access a icle unde he CC BY-NC-ND license (h p://c ea i ecommons.o g/licenses/by-
nc-nd/4.0/).
O iginal A icles
Spec al–Spa ial ans o me -based seman ic segmen a ion o la ge-scale
mapping o indi idual da e palm ees using e y high- esolu ion
sa elli e da a
Rami Al-Ruzouq
a
, Mohamed Ba aka A. Gib il
a
,
*
, Abdallah Shanableh
a
, Jan Bolcek
a
,
b
,
Fouad Lamgha i
c
, Neza A alla Hammou
d
,
e
, Ali El-Keblawy
, Ra i anjan Jena
a
a
GIS and Remo e Sensing Cen e , Resea ch Ins i u e o Sciences and Enginee ing, Uni e si y o Sha jah, Sha jah 27272, Uni ed A ab Emi a es
b
Depa men o Radio Elec onics, Facul y o Elec ical Enginee ing and Communica ion, B no Uni e si y o Technology, B no-K alo o pole 61600, Czech Republic
c
Fujai ah Resea ch Cen e, Al-Hilal Towe , 3003, P.O. Box 666 Fujai ah, Uni ed A ab Emi a es
d
Depa men o Applied Physics and As onomy, Facul y o Science, Uni e si y o Sha jah, Sha jah 27272, Uni ed A ab Emi a es
e
Depa men o Ea h and En i onmen al Sciences, P ince El-Hassan Bin Talal Facul y o Na u al Resou ces & En i onmen , The Hashemi e Uni e si y, Za qa 13133,
Jo dan
Depa men o Applied Biology, College o Sciences, Uni e si y o Sha jah, Sha jah P.O. Box 2727, Uni ed A ab Emi a es
ARTICLE INFO
Keywo ds:
T ee c own delinea ion
Seman ic segmen a ion
Vision ans o me s
Deep lea ning
ABSTRACT
Da e palm plan a ions in he Uni ed A ab Emi a es (UAE) a e unde h ea om soil salini y, d ough , and da e
palm wee ils. Acco dingly, moni o ing and conse ing da e palms a e c ucial o p ese ing a i al componen o
he coun y’s ag icul u al he i age, economy, ood secu i y, and ecological balance. P e ious s udies ha e
e ec i ely iden i ied da e palm ees using RGB-based ae ial and UAV image y u ilizing di e se deep lea ning
me hods. Howe e , he u iliza ion o e y high- esolu ion sa elli e da a o delinea ing indi idual da e palm
c owns emains unexplo ed due o he limi ed spa ial esolu ion capabili ies o exis ing sa elli e sys ems. This
s udy p ima ily aimed o achie e p ecise and comp ehensi e mapping o da e palm ees using Wo ldView-3
(WV-3) sa elli e da a by le e aging he high ep esen a ional powe o he s a e-o - he-a ision ans o me s
(ViT) in cap u ing global in o ma ion om he inpu da a. Fi s , an in-dep h analysis assessmen o he a ious
ans o me -based seman ic segmen a ion a chi ec u es, including Upe Ne wi h ision ans o me and Swin
ans o me , SegFo me , Mask2Fo me , and UniFo me , was conduc ed. Second, he in eg a ion o spec al da a
on he pe o mance o ViTs was e alua ed. Mo eo e , he models’ gene alizabili y and complexi y e ec on he
segmen a ion e ec i eness we e assessed. Acco dingly, a pos p ocessing s a egy was de eloped o aid in
delinea ing and coun ing da e palm ees om seman ic segmen a ion ou pu s. Resul s demons a ed ha in e-
g a ion o WV-3 spec al da a in o he analysis esul ed in a ma ked imp o emen in segmen a ion quali y. The
UniFo me , Upe Ne -Swin, and Mask2Fo me models demons a ed conside able imp o emen s in mul ispec al
da a analysis, wi h inc eases in mean in e sec ion o e union (mIoU) o 2.17% (77.88% mIoU, 86.01% mean F-
sco e [mF-sco e]), 2% (78.10% mIoU, 86.18% mF-sco e), and 1.15% (77.36% mIoU, 85.59% mF-sco e),
espec i ely, compa ed wi h hei RGB-based esul s. E alua ions o model ans e abili y also indica ed ha
Mask2Fo me , UniFo me , and Upe Ne -Swin ans o me s e icien ly adap ed o mul ispec al da a in he Dibba
egion. These models achie ed mIoU sco es o 84.36%, 84.25%, and 83.17% and mF-sco es o 90.95%, 90.87%,
and 90.13%, highligh ing hei e ec i eness and po en ial o b oade egional applica ion. This esea ch
highligh s he e icacy and easibili y o using ViTs wi h WV-3 mul ispec al da a o accu a e and comp ehensi e
su eying o da e palm plan a ions, enabling he de elopmen o palm ee in en o ies and con inuously
upda ing geospa ial da abases.
* Co esponding au ho .
E-mail add ess: [email p o ec ed] (M.B.A. Gib il).
Con en s lis s a ailable a ScienceDi ec
Ecological Indica o s
jou nal homepage: www.else ie .com/loca e/ecolind
h ps://doi.o g/10.1016/j.ecolind.2024.112110
Recei ed 8 Feb ua y 2024; Recei ed in e ised o m 15 Ap il 2024; Accep ed 2 May 2024
Ecological Indica o s 163 (2024) 112110
2
1. In oduc ion
1.1. Backg ound
The 2021 s a is ics o he Food and Ag icul u e O ganiza ion (FAO)
indica ed ha app oxima ely 9.66 million ons o da e palms we e
p oduced, co e ing an a ea o 1.3 million hec a es (FAO, 2023). The
Uni ed A ab Emi a es (UAE) is among he op en p oducing coun ies,
wi h almos 40 million ees dis ibu ed ac oss he UAE (El-Juhany,
2010). Da e palm ees a e an impo an pa o UAE’s cul u al he i age
and a i al ag icul u al esou ce. Thus, accu a e mapping o da e palm
plan a ions ac oss he coun y is impe a i e o e ec i ely and sus ain-
ably moni o and manage his pi o al ag icul u al asse .
T adi ional me hods o da e palm ee in en o y de elopmen , such
as on-si e su eys and isual examina ion o ae ial pho og aphs, a e
cos ly, labo ious, and ime-consuming. Remo e sensing echniques
p o ide a mo e cos -e icien app oach, o e ing de ailed da a wi h
e sa ile empo al and spa ial esolu ions (Ha ling e al., 2019). Remo e
sensing using unmanned ae ial ehicles (UAVs), inc easingly adop ed in
nume ous s udies ocusing on da e palms o i s supe io spa ial and
empo al esolu ions and p ecise posi ional accu acy, aids in he p ecise
de ec ion and mapping o da e palm ees (Amma e al., 2021; Gib il
e al., 2022; Jin asu isak e al., 2022; Gib il e al., 2023). Unlike ea h-
obse ing sa elli es, UAVs ha e limi ed a ea co e age and a e con-
s ained by wea he condi ions and lying es ic ions (Ozda ici-ok and
Ok, 2023). Sa elli e emo e sensing p o ides a comp ehensi e iew,
enabling epea ed obse a ions and ex ensi e co e age ac oss as e-
gions. Combining sa elli e and UAV da a could esul in a mo e
comp ehensi e and in-dep h assessmen o da e palm ees.
A di e se ange o machine lea ning algo i hms has been es ablished
and u ilized o localize and cha ee c owns using da a de i ed om
emo e sensing in au oma ic and semi-au oma ic manne s (Al-Ruzouq
e al., 2018; Ghasemi e al., 2022; Ji e al., 2022; Qin e al., 2022; Zhang
e al., 2023; H. Zhao e al., 2023). In ecen yea s, compu e ision
echniques based on deep lea ning, p ima ily con olu ional neu al
ne wo ks (CNNs), ha e become inc easingly p e alen o indi idual
ee c own (ITC) delinea ing om e y high- esolu ion (VHR) emo ely
sensed images. This p e alence can be a ibu ed o he ema kable
capabili y o hese me hods o au oma ically ex ac in ica e high-le el
ea u es and complex pa e ns om he inpu image da ase s, subs an-
ially enhancing he e icacy and obus ness o he models (Gib il e al.,
2022; Zheng e al., 2023). CNNs ha e demons a ed ou s anding pe -
o mance in a ious ITC s udies (H. Zhao e al., 2023) due o hei
dis inc i e a chi ec u e, which includes localized ecep i e ields,
weigh sha ing, and he p ocess o subsampling (Ka enbo n e al.,
2021). CNN-based models a e ypically used o unde ake di e en asks
o de ec and delinea e ee species. Some widely used asks include
objec de ec ion (Zheng e al., 2021; Zheng and Wu, 2022; Cai e al.,
2023; Velasquez-camacho and E xega ai, 2023), ins ance (B aga e al.,
2020; Yang e al., 2022b; Ball e al., 2023; Hao e al., 2023), and se-
man ic segmen a ion (F eudenbe g e al., 2022; Gib il e al., 2023).
Nume ous seman ic segmen a ion models based on CNNs, a se o
encode –decode deep lea ning a chi ec u es ha a e used o pe o m
pixel-wise classi ica ion, has been de eloped and u ilized o ou line ee
canopies om di e en ypes o emo ely sensed da a (Bha naga e al.,
2020; Cao and Zhang, 2020; Ken sch e al., 2020a; Ken sch e al., 2020b;
To es e al., 2020; Wagne e al., 2020; Wu and Mis a, 2020; Xiao e al.,
2020; Akca and Pola , 2022; Sun e al., 2023). F eudenbe g e al. (2022)
u ilized a wo-s ep app oach based on U-Ne a chi ec u e o ITC om
Wo ldView-3 (WV-3) sa elli e da a (acqui ed in Bengalu u, India) and
ae ial image y (acqui ed in a densely o es ed a ea nea Ga ow, Ge -
many). Thei app oach in ol es ex ac ing ee c owns wi h an
in e sec ion-o e -union (IoU) o 71.2 % on he sa elli e da a and an IoU
o 81.9 ±2.2% on he ae ial image y. Seman ic segmen a ion a chi-
ec u es, wi h di e en backbone ne wo ks, commonly u ilized o ee
c own delinea ion in ecen yea s include he U-Ne (F eudenbe g e al.,
2019; Ken sch e al., 2020a; Wagne e al., 2019,2020; Liu,Wang,and
Wang, 2019; Ken sch e al., 2020b; Schie e e al., 2020) and Deep-
LabV3+(Ayhan and Kwan, 2020; Fe ei a e al., 2020; Cheng,Qi,and
Cheng, 2021).
Gi en ha con olu ions p ocess images by examining local a eas,
hei design inhe en ly es ic s hem om g asping he global con ex o
an en i e image in a single ope a ion, po en ially diminishing accu acy,
especially in scena ios whe e he inpu images exhibi in ica e e-
la ionships be ween pixels (Fayad e al., 2024). Recen ly, he ield o
emo e sensing has expe ienced an inc easing adop ion o di e se ision
ans o me s (ViT) o a a ie y o asks (Abozeid e al., 2022; Fan e al.,
2022; Mekhal i e al., 2022; Sun e al., 2022; L. Yang e al., 2022; Zhou
e al., 2022; Li e al., 2023; Yi e al., 2023; Zhao e al., 2023b; Jamali
e al., 2024). In con as o CNNs, ope a ions wi hin ViTs a e inhe en ly
pa allel and sequence-independen , allowing ViTs o e icien ly ga he
comp ehensi e con ex ual in o ma ion and a ain a supe io le el o
ep esen a ional capabili y (Xia e al., 2022; Zhou e al., 2022). Addi-
ionally, encode s cons uc ed wi h ViTs main ain a uni o m dimen-
sional ep esen a ion h oughou he p ocessing phases, hus
ci cum en ing he necessi y o explici down-sampling ope a ions
(Fayad e al., 2024). Recen ly, an inc easing numbe o s udies ha e
u ilized a ious ViTs in he ields o plan disease de ec ion (Chang e al.,
2024; Li e al., 2024; Rezaei e al., 2024), ui iden i ica ion (Bai e al.,
2024; Liu e al., 2024), canopy heigh mapping (Fayad e al., 2024;
Tolan e al., 2024), and c own delinea ion s udies.
1.2. Rela ed wo k
T ans o me -based models a e inc easingly u ilized in ITC esea ch
o a a ie y o isual asks, including ins ance segmen a ion (Gib il
e al., 2022; De sch e al., 2023; Fi oze e al., 2023), objec de ec ion
(De sch e al., 2022; Zhang e al., 2022), and ee coun ing (Chen and
Shang, 2022; Ami kolaee e al., 2023). De sch e al. (2023) u ilized he
ans o me -based ins ance segmen a ion app oach de ec ion ans-
o me (DETR) o delinea e ITCs in mixed o es a eas u ilizing mul i-
spec al UAV images and UAV-de i ed LiDAR da a. The au ho s
conduc ed expe imen s wi h wo-channel selec ions – RGB and Colo -
In a ed (CIR) – using DETR and achie ed he highes mean F-sco es
(mF-sco es) on he CIR da ase , a aining 90 % in coni e ous o es s, 81
% in deciduous, and 84 % in mixed o es s. Zhang e al. (2022) used a
Fas e R-CNN wi h a Swin ans o me and ResNe -50 backbone o
iden i ying indi idual ees om VHR ae ial images. The s udy exam-
ined he pe o mance o he models in di e se en i onmen s and
obse ed hei highes pe o mance in s ee a eas. Chen and Shang
(2022) in oduced a densi y ans o me designed o he au oma ed
coun ing o ees in ae ial pho og aphs and e alua ed i s e icacy agains
se e al leading-edge algo i hms. Thei p oposed ans o me -based
me hod showed p omising esul s, su passing he pe o mance o
a ious exis ing s a e-o - he-a echniques. Ami kolaee e al. (2023)
in oduced a semi-supe ised me hodology based on a ans o me a -
chi ec u e (encode –decode design) o ee coun ing. Thei ne wo k’s
encode , designed o ex ac ea u es a mul iple scales, u ilized a py -
amid ision ans o me (Wang e al., 2021). Despi e u ilizing he same
quan i y o labeled images, his no el app oach ou pe o med exis ing
semi-supe ised and ully-supe ised echniques in pe o mance.
Fu e al. (2023) examined he pe o mance o wo algo i hms,
namely, adap i e s acking ensemble lea ning (ASEL) and mul i-pa h
ision ans o me o dense p edic ion (MPViT), o mang o e species
mapping om mul ispec al UAV-based images. The s udy showed ha
MPViT ou pe o med ASEL and achie ed 95.5 %–97.3 % o he o e all
classi ica ion accu acy. Zhou e al. (2022) p oposed a semi-au oma ic
wo k low-based T ansUNe o de ec and cha ac e ize b occoli canopy
and heads om he RGB image y and LiDAR da a. Thei esul s showed
ha he p oposed ans o me -based amewo k ou pe o med h ee
CNN-based and wo shallow lea ning-based echniques. Zhang and Liu
(2024) in oduced a dual-b anch ne wo k designed o segmen s ee
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
3
ees using high- esolu ion sa elli e image y. They in eg a ed dila ed
con olu ional laye s in pa allel wi h ans o me blocks o imp o e he
ne wo k’s capaci y o ex ac ing ea u es ac oss mul iple scales. He
e al. (2024) le e aged he CNN-based ne wo k (imp o ed E icicen Ne -
V2) and ans o me -based ne wo k (CSwin ans o me ) o cap u e
local and global seman ic in o ma ion o ci us ee canopy segmen a-
ion om 3D da a ob ained om UAV. Thei app oach showed supe io
pe o mance han using some CNN only and some ans o me -based
only.
A b oad spec um o s udies has ocused on de ec ing palm ees om
mul ipla o m ae ial o hopho os using CNNs (Culman e al., 2020;
Gib il e al., 2021; Amma e al., 2021; Gib il e al., 2022; Jin asu isak
e al., 2022). Gib il e al., (2022,2023, & 2024) demons a ed ha ViTs
ou pe o med a ious CNN-based a chi ec u es o he seman ic and
ins ance segmen a ion o indi idual da e palms om mul ipla o m
ae ial images. Gib il e al. (2022) comp ehensi ely e alua ed a ious
ins ance segmen a ion a chi ec u es o iden i y and ou line indi idual
palm ees h ough mul iscale d one-based images. The au ho s e alu-
a ed a ange o ne wo ks, including Mask R-CNN wi h di e en CNN-
based and Swin ans o me backbones, SOLO, SOLO 2, YOLACT, and
Mask Sco ing R-CNN. The ans o me -based model, Mask R-CNN wi h
Swin ans o me backbones, exhibi ed supe io pe o mance, su pass-
ing CNN-based models. Gib il e al. (2023) examined he ealiabili y o
a ious ision ans o me o ex ac da e palm ees om mul iscale
UAV and ae ial images. Thei esul s showed ha ans o me -based
models achi ed an mF-sco e anging om 91.62 % o 92.44 %.
Among he e alua ed models, he SegFo me model ou pe o med all o
he e alua ed CNN-based models in he mul iscale and he independen
es ing da ase s, ollowed by he Upe Ne -Swin ans o me . Gib il e al.
(2024) p esen ed a ans o me -based me hod o delinea e, coun i y,
and assess he heal h o indi idual da e palm ees om la ge-scale UAV
da a. The p oposed me hod is based on he syne gy be ween an
imp o ed mul iscale ViT (MViT 2), a ea u e py amid ne wo k, Mask R-
CNN, and an imp o ed slicing-aided hype in e ence module. The
model, which was ini ially de eloped o delinea e da e palm ees, was
hen ine- uned o assess he heal h o he da e palm ees. This a chi-
ec u e ou pe o med a ious ins ance segmen a ion models, achie ing
an F-sco e o 94.2 % o segmen a ion and 88.4 % o heal h assessmen .
P e ious esea ch on mapping and moni o ing palm ees based on
deep lea ning has p edominan ly u ilized d one-based and ae ial im-
age y wi h spa ial esolu ions anging om 2 cm o 20 cm.
Dis inguishing he c owns o palm ees in he e ogeneous landscapes
in ensi ies as he da a’s g ound space dis ance (GSD) becomes coa se .
This inc eased di icul y can ad e sely a ec he pe o mance o se-
man ic segmen a ion models, po en ially hinde ing hei abili y o
accu a ely ecognize indi idual da e palm ees. The easibili y o u i-
lizing VHR sa elli e image y o ex ac da e palm ees om sa elli e
images has ye o be ho oughly in es iga ed. This wo k mainly aims o
in oduce an e ec i e deep lea ning me hodology o ex ensi e, de ailed
and accu a e egional su eys o da e palm dis ibu ion, u ilizing he
VHR WV-3 sa elli e image y ac oss mul iple ci ies. Accu a e mapping
and quan i ica ion o indi idual da e palms ac oss egions enables he
au oma ed de elopmen o da e palm ee in en o ies and can p o ide
insigh s in o hei heal h and gene ic di e si y. These da a a e c ucial o
de eloping conse a ion measu es and ensu ing sus ainable cul i a ion
in hei na u al habi a s. The speci ic objec i es o his s udy a e o (1)
e alua e he pe o mance o di e en deep ViTs in egional su eying o
da e palms using WV-3 sa elli e da a, (2) in es iga e he in luence o
spec al da a in eg a ion on he e icacy o he e alua ed ViTs, (3) assess
model gene alizabili y ac oss a ied geog aphic egions, and (4)
de elop a no el pos p ocessing echnique o indi idual da e palm ee
delinea ion and quan i ica ion.
2. Ma e ials and me hods
2.1. O e iew
The esea ch me hodology encompasses i e main s ages (Fig. 1).
Ini ially, WV-3 sa elli e da a iles we e p ep ocessed, including a mo-
sphe ic co ec ion, image pansha pening, and da a no maliza ion. The
second s age in ol ed manually anno a ing da e palm ees; di iding he
da a in o aining, alida ion, and es ing egions; and gene a ing image-
mask pai s. The hi d s age consis ed o comp ehensi e expe imen s o
assess he e icacy o se e al ad anced ViTs in delinea ing da e palm
ees in WV-3 sa elli e image y, u ilizing spa ial and spec al in o ma-
ion. The e alua ed ViTs in his s udy a e Upe Ne (Xiao e al., 2018)
wi h ision ans o me (Kolesniko e al., 2021) and Swin ans o me
(Liu e al., 2021), SegFo me (Xie e al., 2021), Mask2Fo me (Cheng
e al., 2022), and UniFo me (Li e al., 2022). T aining and e alua ion o
hese models we e conduc ed using di e se da a inpu s, including RGB,
a combina ion o RGB wi h NIR1 and NIR2, eigh mul ispec al chan-
nels, and an amalgama ion o hese channels wi h NDVI. In he ou h
Fig. 1. Me hodological amewo k o his wo k.
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
4
s age, he adap abili y o hese models o a ious WV-3 da ase s om
di e en loca ions in Dibba, collec ed on di e en da es, was examined
o de e mine hei gene aliza ion capabili ies. The inal s ep p oposes a
da a pos p ocessing echnique o coun ing and delinea ing indi idual
da e palm ees.
Fig. 2. Loca ion o he expe imen al si es: (a) UAE emi a es, (b) de ailed zoom o he s udy si es, and (c–e) enhanced zoom le els highligh ing he da e palm ees as
seen in mul iple WV-3 images.
Fig. 3. Va ia ions in da e palm ees in e ms o c own size, shape, heigh , and he su ounding en i onmen s.
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
5
2.2. S udy a eas
The expe imen al si es a e si ua ed ac oss a ious ci ies in he UAE,
including Fujai ah, Sha jah, and Ajman (Fig. 2a and 2b). Speci ically,
he a eas a e loca ed in eas Ajman, Al Zubai , Kalba, Dibba, and Masa i.
Collec i ely, hese a eas encompass an a ea o 275 km
2
, dis ibu ed as
ollows: Ajman and Sha jah co e an a ea o 38.6 km
2
, Kalba encom-
passes an a ea o 93 km
2
, Dibba is 77.1 km
2
, and Masa i spans an a ea o
66.3 km
2
. The s udy si es hos a di e se a ay o ee and ege a ion
species, including he da e palm (Phoenix dac yli e a L.), gha (P osopis
cine a ia), Mesqui e (P osopis juli lo a), Sid o Ch is ’s ho n jujube
(Ziziphus spina-ch is i), umb ella ho n acacia (Acacia o ilis), and neem
(Azadi ach a indica), among a ious sh ubs and o he lo a. Fig. 2c–e
display close-up iews o he egions ma ked by cyan ci cles, high-
ligh ing di e en da e palm ees in WV-3 sa elli e da a.
In he p elimina y analysis s age, a comp ehensi e ield campaign
was conduc ed o ho oughly unde s and he loca ions, dis ibu ions,
and isual cha ac e is ics o da e palm ees. This campaign was also
c ucial o gaining insigh in o he s udy a ea’s b oade geog aphic and
en i onmen al con ex s, such as ege a ion and ee species’ e ain,
a ie y, and su oundings. The A cGIS Field Map mobile applica ion was
u ilized o eco d and manage obse a ions h oughou he ield isi s.
The coo dina es o he da e palm plan a ions wi hin he s udy a eas
we e documen ed. The da e palm ees displayed a conside able a ia-
ion in c own shape and size, age, heigh , densi y, and he su ounding
landscape cha ac e is ics (Fig. 3).
2.3. Da a acquisi ion and p ep ocessing
In his esea ch, a o al o 20 mul ida e VHR WV-3 sa elli e image
iles we e u ilized. The WV-3 da ase ea u es eigh spec al bands wi h a
GSD o 1.24 m alongside a panch oma ic band wi h a 0.31 m GSD. The
WV-3 image y o he di e en si es was acqui ed in Ajman and Sha jah
ci y in Ma ch 2021, Kalba in Augus 2022, Dibba in No embe 2019, and
Masa i in Sep embe 2019. Table 1 lis s he gene al cha ac e is ics o he
WV-3 da a. The p ocessing o he WV-3 da a comp ised se e al s ages.
Fi s , a mosphe ic co ec ion was conduc ed using he FLAASH me hod.
Second, he spa ial esolu ion o he mul ispec al da a was imp o ed
h ough he applica ion o he G am–Schmid pansha pening echnique.
The ea e , he bi dep h o he da a unde wen con e sion om 11 bi o
8 bi , s anda dizing he image alues wi hin a 0–255 ange.
Conduc ing p ecise, la ge-scale mapping o da e palm ees u ilizing
WV-3 sa elli e da a wi h classical machine lea ning echniques, which
p ima ily depend on he spec al in o ma ion o he da a, encoun e
no able sho comings. Some expec ed challenges include misclassi ica-
ion and a dec ease in gene alizabili y, which may esul om sub-
s an ial a ia ions in acquisi ion da es and illumina ion condi ions o he
acqui ed sa elli e da a. Fig. 4 illus a es he a e age pixel alues o
Table 1
Cha ac e is ics o he WV-3 da a.
Channel Wa eleng h ange (nm) GSD (m)
Panch oma ic 450–800 0.31
Coas al 400–450 1.24
Blue 450–510 1.24
G een 510–580 1.24
Yellow 585–625 1.24
Red 630–690 1.24
Red Edge 705–745 1.24
NIR-1 770–895 1.24
NIR-2 860–1,040 1.24
Fig. 4. A e age pixel alues o comp ehensi e da e palm ees de i ed om a ious channels o he WV-3 da ase .
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
6
di e se da e palm ees, de i ed om a ious channels o he p e-
p ocessed WV-3 da ase , based on an analysis o 76,950 indi idual da e
palm ees. Fu he mo e, he da e palms wi hin he s udy a ea exhibi ed
a wide ange o c own sizes, s uc u es, ages, and heal h s a es, adding
ano he laye o complexi y o he mapping p ocess. The e o e, his
s udy a emp s o le e age spec al, spa ial, and global con ex ual in-
o ma ion h ough he u iliza ion o s a e-o - he-a ViTs o ensu e ac-
cu a e la ge-scale mapping o da e palm ees om di e en WV-3
da ase s.
2.4. Da a p epa a ion o seman ic segmen a ion
The WV-3 images ea u ing da e palm ees om Ajman and Kalba
we e manually ou lined using g ound- u h da a and isual in e p e a-
ion o VHR ae ial images o ain and assess di e en ans o me -based
deep lea ning models. The p epa a ion o g ound u h da a was
Fig. 5. Examples o (a), (b), and (c) image iles, and he co esponding anno a ions (A, B, C) we e selec ed om he aining da ase s.
Fig. 6. S eps in ol ed in pos p ocessing o delinea e and enume a e sepa a e da e palm ees.
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
7
comp ehensi e and aimed o include la ge a eas wi h di e se a ia ions
in da e palm ees, such as palm ype, canopy size, ele a ion, age, and
adjacen en i onmen s. Delinea ing smalle da e palm ees was chal-
lenging due o he sa elli e da a’s limi ed spa ial esolu ion. The manual
e i ica ion o delinea ed palms, se ing as g ound- u h da a, was
ho oughly conduc ed by analys s using VHR ae ial and UAV image y
om a ious loca ions in he s udy a ea and e ised by mul iple anno-
a o s o quali y assu ance. The ec o da a encompassed a o al o
76,950 da e palm ees sca e ed ac oss ag icul u al and u ban land-
scapes. The WV-3 da a and i s anno a ions we e c opped in o consis en
512 ×512 image iles. A o al o 16,260 image iles we e dedica ed o
de eloping deep lea ning models. Meanwhile, app oxima ely 1800
images we e kep o e alua ing and es ing he models. Fig. 5 displays a
se o image iles and hei co esponding labels (masks) chosen om he
aining da ase .
2.5. Seman ic segmen a ion a chi ec u es
Va ious ans o me -based seman ic segmen a ion a chi ec u es
we e subjec ed o aining and e alua ion, u ilizing a di e se a ay o
spec al image composi es (i.e., s anda d RGB, an in eg a ion o RGB
wi h NIR1 and NIR2 bands, o eigh dis inc mul ispec al channels, and
a usion o hese channels wi h he NDVI). The e alua ed seman ic
segmen a ion models include Upe Ne (Xiao e al., 2018) based on ViT
backbone, Upe Ne based on Swin ans o me (Liu e al., 2021),
Mask2Fo me (Cheng e al., 2022), SegFo me (Xie e al., 2021), and
UniFo me (Li e al., 2022). The models we e implemen ed u ilizing he
PyTo ch (Paszke e al., 2019) and MMsegmen a ion (MMsegmen a ion,
2020) amewo k. The ollowing subsec ions o e a b ie desc ip ion o
he a o emen ioned models.
2.5.1. Upe Ne -ViT
The Upe Ne , an encode –decode ne wo k o seman ic segmen a-
ion de eloped by Xiao e al. (2018), is based on a ea u e py amid
ne wo k (FPN) amewo k (Lin e al., 2017). This amewo k in eg a es a
py amid pooling module (PPM) om he PSPNe (Zhao e al., 2017). The
FPN acili a es a dual p ocess in ol ing he downsampling o ea u e
maps h ough a bo om-up pa hway and subsequen upsampling ia a
op-down mechanism connec ed h ough la e al linkages. This bo om-
up pa hway gene a es ea u e maps ac oss mul iple scales, con ingen
on he a chi ec u e o he chosen backbone a chi ec u e. The spa ial
dimensions o he ex ac ed ea u es a e e ec i ely doubled by he op-
down pa h. This con igu a ion acili a es he amalgama ion o ea u es
ha a e low in esolu ion ye ich in seman ic con en wi h hose ha
Table 2
Lis o he u ilized hype pa ame e s.
Upe Ne based on
ViT and Swin
ans o me and
SegFo me
UniFo me Mask2Fo me
Op imize SGD AdamW AdamW
Loss C oss en opy C oss en opy C oss en opy
P e ained
weigh s
AED20K AED20K AED20K
Ini ial
lea ning
a e
0.01 0.0001 0.0001
Momen um 0.9 0.9 0.9
Weigh
decay
0005 0.0001 0.05
Ba ch size 2 2 2
L.R.
schedule
PolyLR (powe =
0.9, begin =1500,
and end =160,000)
Linea LR (powe
=0.9, begin =
1500, and end =
160,000)
PolyLR (powe =
0.9, begin =1500,
and end =160,000)
Fig. 7. (a) T aining ime, (b) loss g aph, (c) mIoU, and (d) mF-sco e.
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
8
possess high esolu ion bu compa a i ely weake seman ic s eng h.
The e ec i eness o Upe Ne , u ilizing ViT and Swin ans o me back-
bones, was assessed in his s udy.
The s uc u e o ViT comp ises an embedding laye , a ans o me
encode wi h se e al iden ical laye s, and a head classi ie . Fi s , ViT
spli s an image in o uni o m squa e pa ches (i.e., 16 ×16 o 32 ×32
sizes), wi h each pa ch p ocessed as a oken. These pa ches a e hen
con e ed in o 1D ec o sequences ia a lea nable linea p ojec ion o
dimension E, and posi ional embeddings a e added o e ain hei
loca ion in o ma ion. This sequence is subsequen ly inpu ed in o he
ans o me encode . E e y ans o me encode con ains a mul i-head
sel -a en ion (MSA) block and a dense eed o wa d block. A esidual
connec ion suppo s each sublaye , ollowed by laye no maliza ion o
he ou pu . The MSA block comp ises ou componen s: an ini ial linea
laye , a sel -a en ion laye , a conca ena ion laye ha me ges he ou -
pu s om a ious a en ion heads, and a inal linea laye . Sel -a en ion
in ol es calcula ing a en ion sco es e lec ing he pa ch embeddings’
ela ionships. A posi ion-wise eed o wa d ne wo k is applied o each
pa ch embedding by le e aging he sel -a en ion mechanism. The
ans o med encode ’s ou pu is hen ed in o he classi ica ion head,
which pe o ms he classi ica ion o he inpu image, u ilizing he
ea u e ep esen a ions de i ed om he encoded pa ch embeddings.
2.5.2. Upe Ne -Swin ans o me
The Swin ans o me , a hie a chical ans o me s uc u ed in
mul iple s ages, u ilizes a “shi ed window” echnique o de i a e i s
ep esen a ions. This app oach allows he ex ac ion o ea u es a
a ious le els, he eby e icien ly cap u ing he ex ensi e dependencies
inhe en in he da ase (Liu e al. 2021). The “shi ed windowing”
echnique enhances compu a ional e iciency by con ining he sel -
a en ion mechanism o dis inc , non-o e lapping local windows. This
me hod p omo es he es ablishmen o in e connec ions be ween hese
windows. The Swin ans o me ’s ou -s age design encompasses
a ious p ocesses, including mul iscale ea u e map ex ac ion, pa ch
pa i ioning, me ging o pa ches, linea embedding, and implemen ing
Swin ans o me blocks. The pa ch pa i ion module ini ially di ides
he o iginal image in o dis inc , non-o e lapping pa ches, wi h each
pa ch being p ocessed as an indi idual oken. These okens combine aw
pixel da a om mul iple channels o o m hei ea u e ep esen a ion.
This aw ea u e is hen ans o med h ough a linea embedding laye
in o a ea u e ec o o a speci ied dimension (C). Finally, hese pa ch
okens unde go p ocessing by a se ies o ans o me blocks, aiding he
de elopmen o ea u e ep esen a ions. The Swin ans o me block
comp ises h ee main componen s: window MSA, shi ed window MSA,
and a mul ilaye pe cep on (MLP). A e each s age, he a chi ec u e
uses a pa ch me ging laye o educe he oken coun and c ea e a hi-
e a chical ep esen a ion. This model’s “ iny” a ian was chosen as he
ounda ional a chi ec u e o he Upe Ne amewo k, which consis s o
{2, 2, 6, 2} laye s ac oss i s ou s ages, and i ope a es wi h a ea u e
dimension (C) o 96. The eade s may e e o Liu e al. (2021) o
addi ional de ails abou he design and a ia ions o he Swin
ans o me .
2.5.3. Mask2Fo me
The Mask2Fo me (Cheng e al., 2022) is a uni e sal a chi ec u e o
e sa ile asks, such as seman ic segmen a ion, panop ic segmen a ion,
and ins ance segmen a ion. Mask2Fo me is cons uc ed using a simple
me a-a chi ec u e encompassing a backbone ne wo k, a pixel decode ,
and a ans o me decode . The backbone’s a chi ec u e can be designed
using ei he CNN (i.e., esidual lea ning ne wo ks) o ans o me -based
amewo ks. In his wo k, he adop ed backbone ne wo k was based on
he Swin ans o me (Liu e al., 2021), conside ing i s e iciency in
cap u ing global and local ea u es. The pixel decode o Mask2Fo me
u ilizes a mul iscale de o mable a en ion ans o me (MSDe o mA n)
(Zhu e al. 2020) o e ec i ely inco po a e low- and high- esolu ion
ea u es, op imizing he compu a ional e iciency. The ans o me
decode in Mask2 o me uses a masked a en ion mechanism, ocusing
en i ely on local ea u es su ounding he p edic ed segmen s a he
han p ocessing he en i e ea u e map. In his s udy, he “ iny” e sion
o he Swin ans o me was selec ed o se e as he backbone a chi-
ec u e o he Mask2 o me . Reade s can e e o he wo k by Cheng
e al. (2022) o mo e insigh s in o he design and a ia ions o he
Mask2Fo me .
2.5.4. SegFo me
SegFo me (Xie e al., 2021), which is a obus and e ec i e
ans o me -based seman ic segmen a ion a chi ec u e, has demon-
s a ed ema kable pe o mance in a ious emo e sensing applica ions
(Tang e al., 2022; Gib il e al., 2023; Gonçal es e al., 2023; Jiang e al.,
2023). SegFo me is designed using encode –decode design, uniquely
in eg a ing T ans o me s wi h a ligh weigh MLP-based decode . The
encode ex ac s mul iscale ea u es using ou dis inc ans o me
blocks. Each block en ails h ee in eg al modules: an e icien sel -
a en ion mechanism, a mixed eed o wa d ne wo k (FFN) (known as
Mix-FFN), and me ging blocks o o e lapping pa ches. Con a y o he
con en ional ViT ha u ilizes ixed esolu ion posi ion encodings (P.E.s)
o posi ional in o ma ion in eg a ion, he SegFo me adop s con olu-
ional laye s in i s FFN o a da a-dependen app oach o posi ional
encoding. The decode uses mul iscale in o ma ion, encompassing local
and global de ails, o accu a ely p edic he inal segmen a ion ou -
comes. This s udy selec ed a mix ans o me (MiT) encode based on B3
(MiT-B3). Fu he in o ma ion can be ound in he s udy by Xie e al.
(2021).
2.5.5. UniFo me
The uni ied ans o me (Li e al., 2022) le e ages he capabili ies o
CNNs and ans o me s o mi iga e each app oach’s limi a ions,
achie ing an op imal balance be ween compu a ional e iciency and
accu acy. A basic ans o me o ma (Vaswani e al., 2017) is u ilized
and cus omized o e icien and e ec i e lea ning o spa io empo al
ep esen a ions. The UniFo me block consis s o h ee main modules:
dynamic posi ion embedding (DPE), mul i-head ela ion agg ega o
Table 3
Quan i a i e e alua ion o he ans o me -based seman ic segmen a ion
models.
Me ics RGB
(%)
RGB þ
NIR1 þ
NIR2
(%)
Eigh MS
bands
(%)
Eigh MS
bands þ
NDVI (%)
SegFo me -MiT-
B3
mAcc 85.29 85.69 86.89 85.60
mIoU 76.16 76.54 76.62 76.47
mP ecision 83.92 84.16 83.27 84.15
mRecall 85.29 85.69 86.89 85.60
mF-sco e 84.59 84.91 84.98 84.86
Upe Ne -ViT mAcc 83.41 74.29 78.06 78.13
mIoU 75.04 68.99 71.57 71.64
mP ecision 83.87 83.09 83.46 83.51
mRecall 83.41 74.29 78.06 78.13
mF-sco e 83.64 77.98 80.51 80.57
Upe Ne -Swin-
iny
mAcc 85.84 86.81 86.93 85.64
mIoU 76.12 76.78 78.10 76.65
mP ecision 83.38 83.57 85.46 84.38
mRecall 85.84 86.81 86.93 85.64
mF-sco e 84.56 85.12 86.18 85.0
Mask2Fo me -
Swin- iny
mAcc 85.11 84.23 87.54 85.41
mIoU 76.21 75.28 77.36 75.73
mP ecision 84.17 83.48 83.85 83.14
mRecall 85.11 84.23 87.54 85.41
mF-sco e 84.63 83.85 85.59 84.24
UniFo me mAcc 83.58 86.86 88.30 86.33
mIoU 75.71 77.51 77.88 77.34
mP ecision 84.86 84.62 84.00 84.83
mRecall 83.58 86.86 88.3 86.33
mF-sco e 84.21 85.70 86.01 85.56
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
9
(MHRA), and eed o wa d ne wo k (FFN). The DPE ac i ely in-
co po a es 3D posi ional da a in o each oken, e icien ly le e aging he
spa io empo al sequence o he okens o ideo modeling pu poses. The
MHRA combines each oken wi h i s con ex ual coun e pa s, which
handles local edundancy and global dependency in ideos h ough i s
adap able app oach o oken a ini y lea ning ac oss shallow and deep
laye s. Al hough he local MHRA in he shallow laye s signi ican ly
lessens he compu a ional bu den, he global MHRA in he deepe laye s
ocuses on comp ehensi ely lea ning global oken ela ionships. The
eade may e e o Li e al. (2022) o he de ailed in o ma ion abou he
UniFo me a chi ec u e.
2.6. Accu acy me ics
This s udy e alua ed he e icacy o a ious seman ic segmen a ion
a chi ec u es h ough mul iple accu acy me ics. These me ics de e -
mine he cong uence be ween he e e ence ee c owns and he p e-
dic ed delinea ed ee c owns by ViTs. The combina ion o he a ious
seman ic segmen a ion e alua ion me ics p o ides a ho ough unde -
s anding o he esul s o da e palm ee mapping, highligh ing he e o s
and limi a ions o he di e en models. The u ilized me ics include
mIoU, p ecision, ecall, and mF-sco e and a e de ailed in Equa ions (1)–
(6). The IoU and F-sco e a e commonly u ilized me ics in ITC s udies o
assess he le el o ag eemen be ween deep lea ning model esul s and
g ound- u h da a. These me ics ange om ze o o one o hund ed,
Fig. 8. Cu a ed selec ion o se en mul ispec al images (a–g) andomly chosen om he es ing da a, accompanied by hei co esponding e e ence da a and he
expe imen al esul s.
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
16
managemen and moni o ing da e palm ees.
CRediT au ho ship con ibu ion s a emen
Rami Al-Ruzouq: W i ing – e iew & edi ing, W i ing – o iginal
d a , Supe ision, P ojec adminis a ion, Me hodology, In es iga ion,
Fo mal analysis, Concep ualiza ion. Mohamed Ba aka A. Gib il:
W i ing – e iew & edi ing, W i ing – o iginal d a , Visualiza ion,
Valida ion, So wa e, Me hodology, Fo mal analysis, Da a cu a ion,
Concep ualiza ion. Abdallah Shanableh: W i ing – e iew & edi ing,
Supe ision, Resou ces, Concep ualiza ion. Jan Bolcek: W i ing –
o iginal d a , Valida ion, So wa e, Me hodology, Fo mal analysis.
Fouad Lamgha i: W i ing – e iew & edi ing, Supe ision, P ojec
adminis a ion, Funding acquisi ion, Concep ualiza ion. Neza A alla
Hammou : W i ing – e iew & edi ing, Supe ision, Concep ualiza ion.
Ali El-Keblawy: W i ing – e iew & edi ing, W i ing – o iginal d a ,
Supe ision, Concep ualiza ion. Ra i anjan Jena: W i ing – e iew &
edi ing, Me hodology, Fo mal analysis.
Decla a ion o compe ing in e es
The au ho s decla e ha hey ha e no known compe ing inancial
in e es s o pe sonal ela ionships ha could ha e appea ed o in luence
he wo k epo ed in his pape .
Da a a ailabili y
The like o he code is gi en in he manusc ip
Acknowledgmen s
The au ho s exp ess hei app ecia ion o he Fujai ah Resea ch
Cen e (FRC) o i s inancial suppo (Funded P ojec Numbe : 133049)
and o he Uni e si y o Sha jah o o e ing esea ch acili ies.
Re e ences
“Food and Ag icul u e O ganiza ion.” FAOSTAT [In e ne ]. [accessed 2021 Ma 9].
h p://www. ao.o g/ aos a /en/#da a/QC.
Abozeid, A., Alanazi, R., Elhadad, A., Taloba, A.I., Abd El-Aziz, R.M., 2022. A la ge-scale
da ase and deep lea ning model o de ec ing and coun ing oli e ees in sa elli e
image y. Compu In ell Neu osci. 2022 h ps://doi.o g/10.1155/2022/1549842.
Akca S, Pola N. 2022. Seman ic segmen a ion and quan i ica ion o ees in an o cha d
using UAV o hopho o. Ea h Sci In o ma ics 2022 154 [In e ne ]. [accessed 2022
No 28] 15(4):2265–2274. h ps://doi.o g/10.1007/S12145-022-00871-Y.
Al-Ruzouq, R., Shanableh, A., Ba aka , A., Gib il, M., AL-Mansoo i, S., 2018. Image
segmen a ion pa ame e selec ion and an colony op imiza ion o da e palm ee
de ec ion and mapping om e y-high-spa ial- esolu ion ae ial image y. Remo e
Sens [in e ne ]. 10 (9), 1413. h ps://doi.o g/10.3390/ s10091413.
Ami kolaee, H.A., Shi, M., Membe , S., Mulligan, M., 2023. T eeFo me : a semi-
supe ised ans o me -based amewo k o ee coun ing om a single. IEEE T ans
Geosci Remo e Sens. 61, 1–15. h ps://doi.o g/10.1109/TGRS.2023.3295802.
Amma , A., Koubaa, A., Benjdi a, B., 2021. Deep-lea ning-based au oma ed palm ee
coun ing and geoloca ion in la ge a ms om ae ial geo agged images. Ag onomy
[in e ne ]. 11 (8), 1458. h ps://doi.o g/10.3390/ag onomy11081458.
Ayhan, B., Kwan, C., 2020. T ee, sh ub, and g ass classi ica ion using only RGB images.
Remo e Sens. 12 (8) h ps://doi.o g/10.3390/RS12081333.
Bai, Y., Yu, J., Yang, S., Ning, J., 2024. An imp o ed YOLO algo i hm o de ec ing
lowe s and ui s on s awbe y seedlings. Biosys Eng [in e ne ]. 237(June 2023):
1–12 h ps://doi.o g/10.1016/j.biosys emseng.2023.11.008.
Ball JGC, Hickman SHM, Jackson TD, Koay XJ, Hi s J, Jay W, A che M, Coomes DA.
2023. Accu a e delinea ion o indi idual ee c owns in opical o es s om ae ial
RGB image y using Mask R-CNN. :1–14. h ps://doi.o g/10.1002/ se2.332.
Bha naga , S., Gill, L., Ghosh, B., 2020. D one image segmen a ion using machine and
deep lea ning o mapping aised bog ege a ion communi ies. Remo e Sens. 12 (16)
h ps://doi.o g/10.3390/RS12162602.
B aga, J.R.G., Pe ipa o, V., Dalagnol, R., Fe ei a, M.P., Ta abalka, Y., A ag˜
ao, L.E.O.C.,
de Campos Velho, H.F., Shiguemo i, E.H., Wagne , F.H., 2020. T ee c own
delinea ion algo i hm based on a con olu ional neu al ne wo k. Remo e Sens. 12 (8),
1–27. h ps://doi.o g/10.3390/RS12081288.
Cai C, Xu H, Chen S, Yang L, Weng Y, Huang S, Dong C, Lou X. 2023. T ee Recogni ion
and C own Wid h Ex ac ion Based on No el Fas e -RCNN in a Dense Loblolly Pine
En i onmen .
Cao, K., Zhang, X., 2020. An imp o ed Res-UNe model o ee species classi ica ion
using ai bo ne high- esolu ion images. Remo e Sens. 12 (7) h ps://doi.o g/
10.3390/ s12071128.
Chang, B., Wang, Y., Zhao, X., Li, G., Yuan, P., 2024. A gene al-pu pose edge- ea u e
guidance module o enhance ision ans o me s o plan disease iden i ica ion.
Expe Sys Appl [in e ne ]. 237, 121638 h ps://doi.o g/10.1016/j.
eswa.2023.121638.
Chen, G., Shang, Y., 2022. T ans o me o ee coun ing in ae ial images. Remo e Sens.
14 (3), 476. h ps://doi.o g/10.3390/ s14030476.
Cheng B, Mis a I, Schwing AG, Ki illo A, Gi dha R. 2022. Masked-a en ion Mask
T ans o me o Uni e sal Image Segmen a ion. P oc IEEE Compu Soc Con Compu
Vis Pa e n Recogni . 2022-June:1280–1289. h ps://doi.o g/10.1109/
CVPR52688.2022.00135.
Cheng, Z., Qi, L., Cheng, Y., 2021. Che y ee c own ex ac ion om na u al o cha d
images wi h complex backg ounds. Ag icul u e [in e ne ]. 11 (5), 431. h ps://doi.
o g/10.3390/ag icul u e11050431.
Culman, M., Delalieux, S., Van T ich , K., 2020. Indi idual palm ee de ec ion using
deep lea ning on RGB image y o suppo ee in en o y. Remo e Sens. 12 (21),
1–31. h ps://doi.o g/10.3390/ s12213476.
De sch, S., K zys ek, P., Heu ich, M., Resou ces, N., Fo es , B., Pa k, N., Moni o ing, N.P.,
Technology, I., Sys ems, I., 2022. NOVEL SINGLE TREE DETECTION BY
TRANSFORMERS USING UAV-BASED MULTISPECTRAL IMAGERY, XLIII(June):
6–11.
De sch, S., Sch¨
o l, A., K zys ek, P., Heu ich, M., 2023. Towa ds comple e ee c own
delinea ion by ins ance segmen a ion wi h Mask R-CNN and DETR using UAV-based
mul ispec al image y and lida da a. ISPRS Open J Pho og amm Remo e Sens. 8,
100037.
El-Juhany, L., 2010. Deg ada ion o da e palm ees and da e p oduc ion in A ab
coun ies: causes and po en ial ehabili a ion. Aus J Basic Appl Sci [in e ne ]. 4 (8),
3998–4010. h ps://doi.o g/10.1016/j. c .2011.04.017.
Fan, F., Zeng, X., Wei, S., Zhang, H., Tang, D., Shi, J., Zhang, X., 2022. E icien ins ance
segmen a ion pa adigm o in e p e ing SAR and op ical images. Remo e Sens. 14
(3), 1–22. h ps://doi.o g/10.3390/ s14030531.
Fayad, I., Ciais, P., Schwa z, M., Wigne on, J.P., Baghdadi, N., de T uchis, A.,
d’Asp emon , A., F appa , F., Saa chi, S., Sean, E., e al., 2024. Hy-TeC: a hyb id
ision ans o me model o high- esolu ion and la ge-scale mapping o canopy
heigh . Remo e Sens En i on. 302 (Decembe 2023) h ps://doi.o g/10.1016/j.
se.2023.113945.
Fe ei a MP, Almeida DRA de, Papa D de A, Mine ino JBS, Ve as HFP, Fo mighie i A,
San os CAN, Fe ei a MAD, Figuei edo EO, Fe ei a EJL. 2020. Indi idual ee
de ec ion and species classi ica ion o Amazonian palms using UAV images and deep
lea ning. Fo Ecol Manage [In e ne ]. 475(Ap il):118397. h ps://doi.o g/10.1016/
j. o eco.2020.118397.
Fi oze A, Wing en C, Yeh RA, Benes B, Aliaga D. 2023. T ee Ins ance Segmen a ion Wi h
Tempo al Con ou G aph. In: P oc IEEE/CVF Con Compu Vis Pa e n Recogni .
[place unknown]; p. 2193–2202.
F eudenbe g, M., N¨
olke, N., Agos ini, A., U ban, K., W¨
o g¨
o e , F., Kleinn, C., 2019. La ge
scale palm ee de ec ion in high esolu ion sa elli e images using U-Ne . Remo e
Sens. 11 (3), 1–18. h ps://doi.o g/10.3390/ s11030312.
F eudenbe g, M., Magdon, P., N¨
olke, N., 2022. Indi idual ee c own delinea ion in high-
esolu ion emo e sensing images based on U-Ne . Neu al Compu Appl. 34 (24),
22197–22207. h ps://doi.o g/10.1007/s00521-022-07640-4.
Fu, B., He, X., Liang, Y., Deng, T., Li, H., He, H., Jia, M., Fan, D., Wang, F., 2023.
Examina ion o he pe o mance o ASEL and MPViT algo i hms o classi ying
mang o e species o mul iple na u al ese es o Beibu Gul , sou h China. Ecol Indic.
154 (Augus ) h ps://doi.o g/10.1016/j.ecolind.2023.110870.
Ghasemi, M., La i i, H., Pou hashemi, M., 2022. A no el me hod o de ec ing and
delinea ing coppice ees in UAV images o moni o ee decline. Remo e Sens. 14
(23), 5910.
Gib il MBA, Sha i HZM, Shanableh A, Al-Ruzouq R, Wayayok A, Hashim SJ bin, Sachi
MS. 2022. Deep con olu ional neu al ne wo ks and Swin ans o me -based
amewo ks o indi idual da e palm ee de ec ion and mapping om la ge-scale
UAV images. Geoca o In [In e ne ]. 37(27):18569–18599. h ps://doi.o g/
10.1080/10106049.2022.2142966.
Gib il MBA, Sha i HZM, Shanableh A, Al-Ruzouq R, bin Hashim SJ, Wayayok A, Sachi
MS. 2024. La ge-scale assessmen o da e palm plan a ions based on UAV emo e
sensing and mul iscale ision ans o me . Remo e Sens Appl Soc En i on
[In e ne ].:101195. h ps://doi.o g/h ps://doi.o g/10.1016/j. sase.2024.101195.
Gib il, M.B.A., Sha i, H.Z.M., Shanableh, A., Al-Ruzouq, R., Wayayok, A., Hashim, S.J.,
2021. Deep con olu ional neu al ne wo k o la ge-scale da e palm ee mapping
om ua -based images. Remo e Sens. 13 (14), 1–24. h ps://doi.o g/10.3390/
s13142787.
Gib il, M.B.A., Sha i, H.Z.M., Al-Ruzouq, R., Shanableh, A., Nahas, F., Al, M.S., 2023.
La ge-scale da e palm ee segmen a ion om mul iscale UAV-based and ae ial
images using deep ision ans o me s. D ones. 7 (2) h ps://doi.o g/10.3390/
d ones7020093.
Gonçal es DN, Ma ca o J, Ca ilho AC, Acos a PR, Ramos APM, Gomes FDG, Osco LP, da
Rosa Oli ei a M, Ma ins JAC, Damasceno GA, e al. 2023. T ans o me s o mapping
bu ned a eas in B azilian Pan anal and Amazon wi h Plane Scope image y. In J
Appl Ea h Obs Geoin [In e ne ]. 116(Decembe 2022):103151. h ps://doi.o g/
10.1016/j.jag.2022.103151.
Hao, Z., Lin, L., Pos , C.J., Mikhailo a, E.A., Yu, K., 2023. The co-e ec o image
esolu ion and c own size on deep lea ning o indi idual ee de ec ion and
delinea ion. In J Digi Ea h [In e ne ].:3753–3771. h ps://doi.o g/10.1080/
17538947.2023.2257636.
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
17
Ha ling, S., Sagan, V., Sidike, P., Maimai ijiang, M., Ca on, J., 2019. U ban ee species
classi ica ion using a wo ld iew-2/3 and liDAR da a usion app oach and deep
lea ning. Senso s (swi ze land). 19 (6), 1–23. h ps://doi.o g/10.3390/s19061284.
He, H., Zhou, F., Xia, Y., Chen, M., Chen, T., 2024. Pa allel usion neu al ne wo k
conside ing local and global seman ic in o ma ion o ci us ee canopy
segmen a ion. IEEE J Sel Top Appl Ea h Obs Remo e Sens. 17, 1535–1549. h ps://
doi.o g/10.1109/JSTARS.2023.3339290.
Jamali, A., Roy, S.K., Hashemi Beni, L., P adhan, B., Li, J., Ghamisi, P., 2024. Residual
wa e ision U-Ne o lood mapping using dual pola iza ion Sen inel-1 SAR
image y. In J Appl Ea h Obs Geoin [in e ne ]. 127, 103662 h ps://doi.o g/
10.1016/j.jag.2024.103662.
Ji, Y., Yan, E., Yin, X., Song, Y., Wei, W., Mo, D., 2022. Au oma ed ex ac ion o Camellia
olei e a c own using unmanned ae ial ehicle isible images and he ResU-Ne deep
lea ning model. F on Plan Sci. 13 (Augus ), 1–13. h ps://doi.o g/10.3389/
pls.2022.958940.
Jiang, K., A zaal, U., Lee, J., 2023. T ans o me -based weed segmen a ion o g ass
managemen . Senso s. 23 (1), 1–15. h ps://doi.o g/10.3390/s23010065.
Jin asu isak T, Edi isinghe E, Elba ay A. 2022. Deep neu al ne wo k based da e palm
ee de ec ion in d one image y. Compu Elec on Ag ic [In e ne ]. 192(Ap il 2021):
106560. h ps://doi.o g/10.1016/j.compag.2021.106560.
Ka enbo n, T., Lei lo , J., Schie e , F., Hinz, S., 2021. Re iew on con olu ional neu al
ne wo ks (CNN) in ege a ion emo e sensing. ISPRS J Pho og amm Remo e Sens.
173, 24–49.
Ken sch, S., Cace es, M.L.L., Se ano, D., Rou e, F., Diez, Y., 2020a. Compu e ision and
deep lea ning echniques o he analysis o d one-acqui ed o es images, a ans e
lea ning s udy. Remo e Sens. 12 (8), 1–19. h ps://doi.o g/10.3390/RS12081287.
Ken sch S, Ka a siolis S, Kamila is A, Tomha e L, Lopez Cace es ML. 2020. Iden i ica ion
o T ee Species in Japanese Fo es s based on Ae ial Pho og aphy and Deep Lea ning.
a Xi . h ps://doi.o g/10.1007/978-3-030-61969-5_18.
Kolesniko , A., Doso i skiy, A., Weissenbo n, D., Heigold, G., Uszko ei , J., Beye , L.,
Minde e , M., Dehghani, M., Houlsby, N., Gelly, S., e al., 2021. An image is wo h
16x16 wo ds: T ans o me s o image ecogni ion a scale. In: [place Unknown].
Li, R., Chen, T., Liu, Y., Jiang, H., 2023. CoupleUNe : Swin T ans o me coupling CNNs
makes s ong con ex ual encode s o VHR image oad ex ac ion. In J Remo e Sens.
44 (18), 5788–5813.
Li K, Wang Y, Gao P, Song G, Liu Y, Li H, Qiao Y. 2022. Uni o me : Uni ied ans o me
o e icien spa io empo al ep esen a ion lea ning. a Xi P ep a Xi 220104676.
Li, Y., Ma, L., Sun, N., 2024. A bilinea ans o me in e ac i e neu al ne wo ks-based
app oach o ine-g ained ecogni ion and p o ec ion o plan diseases o ga dening
design. C op P o [in e ne ]. 180, 106660 h ps://doi.o g/10.1016/j.
c op o.2024.106660.
Lin, T.-Y., Doll´
a , P., Gi shick, R., He, K., Ha iha an, B., Belongie, S., 2017. Fea u e
py amid ne wo ks o objec de ec ion. In: P oc IEEE Con Compu is Pa e n
Recogni . Honolulu, HI, USA, pp. 2117–2125.
Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S, Guo B. 2021. Swin T ans o me :
Hie a chical Vision T ans o me using Shi ed Windows. In: 2021 IEEE/CVF In Con
Compu Vis. Mon eal, QC, Canada: IEEE; p. 9992–10002. h ps://doi.o g/10.1109/
ICCV48922.2021.00986.
Liu, J., Wang, X., Wang, T., 2019. Classi ica ion o ee species and s ock olume
es ima ion in g ound o es images using Deep Lea ning. Compu Elec on Ag ic
[in e ne ]. 166 (May), 105012 h ps://doi.o g/10.1016/j.compag.2019.105012.
Liu, H., Wang, X., Zhao, F., Yu, F., Lin, P., Gan, Y., Ren, X., Chen, Y., Tu, J., 2024.
Upg ading swin-B ans o me -based model o accu a ely iden i ying ipe
s awbe ies by coupling ask-aligned one-s age objec de ec ion mechanism.
Compu Elec on Ag ic [in e ne ]. 218, 108674 h ps://doi.o g/10.1016/j.
compag.2024.108674.
Mekhal i, M.L., Nicolo, C., Bazi, Y., Al, R.MM., Alsha i , N.A., Al, M.E., 2022. Con as ing
YOLO 5, ans o me , and e icien de de ec o s o c op ci cle de ec ion in dese .
IEEE Geosci Remo e Sens Le . 19, 19–23. h ps://doi.o g/10.1109/
LGRS.2021.3085139.
Con ibu o s Mms. 2020. {MMSegmen a ion}: OpenMMLab Seman ic Segmen a ion
Toolbox and Benchma k.
Ozda ici-ok, A., Ok, A.O., 2023. Scien ia Ho icul u ae Using emo e sensing o iden i y
indi idual ee species in o cha ds : Sci Ho ic (Ams e dam). [in e ne ]. 321 (10),
112333 h ps://doi.o g/10.1016/j.scien a.2023.112333.
Paszke, A., G oss, S., Massa, F., Le e , A., B adbu y, J., Chanan, G., Killeen, T., Lin, Z.,
Gimelshein, N., An iga, L., e al., 2019. Py o ch: An impe a i e s yle, high-
pe o mance deep lea ning lib a y. Ad Neu al In P ocess Sys . 32.
Qin, H., Zhou, W., Yao, Y., Wang, W., 2022. Indi idual ee segmen a ion and ee
species classi ica ion in sub opical b oadlea o es s using UAV-based LiDAR,
hype spec al, and ul ahigh- esolu ion RGB da a. Remo e Sens En i on. 280,
113143.
Rezaei, M., Diepe een, D., Laga, H., Jones, M.G.K., Sohel, F., 2024. Plan disease
ecogni ion in a low da a scena io using ew-sho lea ning. Compu Elec on Ag ic
[in e ne ]. 219, 108812 h ps://doi.o g/10.1016/j.compag.2024.108812.
Schie e , F., Ka enbo n, T., F ick, A., F ey, J., Schall, P., Koch, B., Schmid lein, S., 2020.
Mapping o es ee species in high esolu ion UAV-based RGB-image y by means o
con olu ional neu al ne wo ks. ISPRS J Pho og amm Remo e Sens [in e ne ]. 170
(No embe ), 205–215. h ps://doi.o g/10.1016/j.isp sjp s.2020.10.015.
Sel a aju RR, Cogswell M, Das A, Vedan am R, Pa ikh D, Ba a D. 2017. G ad-cam:
Visual explana ions om deep ne wo ks ia g adien -based localiza ion. In: P oc
IEEE In Con Compu Vis. [place unknown]; p. 618–626.
Sun, Y., Li, Z., He, H., Guo, L., Zhang, X., Xin, Q., 2022. Coun ing ees in a sub opical
mega ci y using he ins ance segmen a ion me hod. In J Appl Ea h Obs Geoin . 106,
102662 h ps://doi.o g/10.1016/j.jag.2021.102662.
Sun, Y., Hao, Z., Guo, Z., Liu, Z., Jh., 2023. De ec ion and mapping o ches nu using
deep lea ning om high- esolu ion UAV-Based RGB image . Remo e Sens.:1–18.
Suzuki, S., 1985. Topological s uc u al analysis o digi ized bina y images by bo de
ollowing. Compu Vision, G aph Image P ocess. 30 (1), 32–46.
Tang, X., Tu, Z., Wang, Y., Liu, M., Li, D., Fan, X., 2022. Au oma ic de ec ion o coseismic
landslides using a new ans o me me hod. Remo e Sens. 14 (12), 1–19. h ps://doi.
o g/10.3390/ s14122884.
Tolan J, Yang HI, Nosa zewski B, Couai on G, Vo H V., B and J, Spo e J, Majumda S,
Haziza D, Vama aju J, e al. 2024. Ve y high esolu ion canopy heigh maps om
RGB image y using sel -supe ised ision ans o me and con olu ional decode
ained on ae ial lida . Remo e Sens En i on [In e ne ]. 300(Ap il 2023):113888.
h ps://doi.o g/10.1016/j. se.2023.113888.
To es, D.L., Fei osa, R.Q., Happ, P.N., La Rosa, L.E.C., Junio , J.M., Ma ins, J.,
B essan, P.O., Gonçal es, W.N., Liesenbe g, V., 2020. Applying ully con olu ional
a chi ec u es o seman ic segmen a ion o a single ee species in u ban
en i onmen on high esolu ion UAV op ical image y. Senso s (swi ze land). 20 (2),
1–20. h ps://doi.o g/10.3390/s20020563.
Ulku, I., Akagündüz, E., Ghamisi, P., Membe , S., 2022. Deep Seman ic Segmen a ion o
T ees Using Mul ispec al Images. 15, 7589–7604. h ps://doi.o g/10.1109/
JSTARS.2022.3203145.
Vaswani A, Shazee N, Pa ma N, Uszko ei J, Jones L, Gomez AN, Kaise Ł, Polosukhin I.
2017. A en ion is all you need. In: Ad Neu al In P ocess Sys . Vol. 30. [place
unknown]; p. 5998–6008.
Velasquez-camacho, L., E xega ai, M., 2023. Deep Lea ning algo i hms o u ban ee
de ec ion and geoloca ion wi h high- esolu ion ae ial, sa elli e, and g ound-le el
images. Compu En i on U ban Sys [in e ne ]. 105 (Augus ), 102025 h ps://doi.
o g/10.1016/j.compen u bsys.2023.102025.
Wagne , F.H., Sanchez, A., Ta abalka, Y., Lo e, R.G., Fe ei a, M.P., Aida , M.P.M.,
Gloo , E., Phillips, O.L., A ag˜
ao, L.E.O.C., 2019. Using he U-ne con olu ional
ne wo k o map o es ypes and dis u bance in he A lan ic ain o es wi h e y high
esolu ion images. Remo e Sens Ecol Conse . 5 (4), 360–375. h ps://doi.o g/
10.1002/ se2.111.
Wagne , F.H., Sanchez, A., Aida , M.P.M., Rochelle, A.L.C., Ta abalka, Y., Fonseca, M.G.,
Phillips, O.L., Gloo , E., A ag˜
ao, L.E.O.C., 2020. Mapping A lan ic ain o es
deg ada ion and egene a ion his o y wi h indica o species using con olu ional
ne wo k. PLoS One. 15 (2), e0229448.
Wang W, Xie E, Li X, Fan D-P, Song K, Liang D, Lu T, Luo P, Shao L. 2021. Py amid Vision
T ans o me : A Ve sa ile Backbone o Dense P edic ion wi hou Con olu ions
[In e ne ]. :568–578. h p://a xi .o g/abs/2102.12122.
Wu, Y., Mis a, S., 2020. In elligen image segmen a ion o o ganic- ich shales using
andom o es , wa ele ans o m, and hessian ma ix. IEEE Geosci Remo e Sens Le .
17 (7), 1144–1147. h ps://doi.o g/10.1109/LGRS.2019.2943849.
Xia Z, Pan X, Song S, Li LE, Huang G. 2022. Vision T ans o me wi h De o mable
A en ion [In e ne ]. h p://a xi .o g/abs/2201.00520.
Xiao T, Liu Y, Zhou B, Jiang Y, Sun J. 2018. Uni ied Pe cep ual Pa sing o Scene
Unde s anding. Lec No es Compu Sci (including Subse Lec No es A i In ell Lec
No es Bioin o ma ics). 11209 LNCS:432–448. h ps://doi.o g/10.1007/978-3-030-
01228-1_26.
Xiao, C., Qin, R., Huang, X., 2020. T ee op de ec ion using con olu ional neu al
ne wo ks ained h ough au oma ically gene a ed pseudo labels. In J Remo e Sens
[in e ne ]. 41 (8), 3010–3030. h ps://doi.o g/10.1080/01431161.2019.1698075.
Xie, E., Wang, W., Yu, Z., Anandkuma , A., Al a ez, J.M., Luo, P., 2021. SegFo me :
simple and e icien design o seman ic segmen a ion wi h ans o me s. Ad Neu al
In P ocess Sys . 15, 12077–12090.
Yang, M., Mou, Y., Liu, S., Meng, Y., Liu, Z., Li, P., Xiang, W., Zhou, X., Peng, C., 2022b.
De ec ing and mapping ee c owns based on con olu ional neu al ne wo k and
Google Ea h images. In J Appl Ea h Obs Geoin . 108(Augus , 2021). h ps://doi.
o g/10.1016/j.jag.2022.102764.
Yang, L., Wang, X., Zhai, J., 2022a. Wa e line ex ac ion o a i icial coas wi h ision
ans o me s. F on En i on Sci. 10 (Feb ua y), 1–12. h ps://doi.o g/10.3389/
en s.2022.799250.
Yi, S., Liu, X., Li, J., Chen, L., 2023. UAV o me : a composi e ans o me ne wo k o
u ban scene segmen a ion o UAV images. Pa e n Recogni . 133 h ps://doi.o g/
10.1016/j.pa cog.2022.109019.
Zhang, T., Hu, D., Wu, C., Liu, Y., Yang, J., Tang, K., 2023. La ge-scale apple o cha d
mapping om mul i-sou ce da a using he seman ic segmen a ion model wi h image-
o-image ansla ion and ans e lea ning. Compu Elec on Ag ic. 213, 108204.
Zhang, L., Lin, H., Wang, F., 2022. Indi idual ee de ec ion based on high- esolu ion
RGB images o u ban o es y applica ions. IEEE Access. 10, 46589–46598. h ps://
doi.o g/10.1109/ACCESS.2022.3171585.
Zhang, H., Liu, S., 2024. Double-b anch mul i-scale con ex ual ne wo k: a model o
mul i-scale s ee ee segmen a ion in high- esolu ion emo e sensing images.
Senso s. 24 (4), 1110. h ps://doi.o g/10.3390/s24041110.
Zhao, J., Be ge, T.W., Geipel, J., 2023b. T ans o me in UAV image-based weed
mapping. Remo e Sens. 15 (21), 5165.
Zhao H, Shi J, Qi X, Wang X, Jia J. 2017. Py amid scene pa sing ne wo k. In: P oc - 30 h
IEEE Con Compu Vis Pa e n Recogni ion, CVPR 2017. Vol. 2017-Janua. [place
unknown]; p. 6230–6239. h ps://doi.o g/10.1109/CVPR.2017.660.
Zhao, H., Mo gen o h, J., Pea se, G., Schindle , J., 2023a. A Sys ema ic e iew o
indi idual ee c own de ec ion and delinea ion wi h con olu ional neu al ne wo ks
(CNN). Cu o Repo s [in e ne ]. 9 (3), 149–170. h ps://doi.o g/10.1007/
s40725-023-00184-3.
Zheng, J., Wu, W., Yu, L., Fu, H., 2021. COCONUT TREES DETECTION ON THE
TENARUNGA USING HIGH-RESOLUTION SATELLITE IMAGES AND DEEP
LEARNING Minis y o Educa ion Key Labo a o y o Ea h Sys em Modeling, and
R. Al-Ruzouq e al.
Ecological Indica o s 163 (2024) 112110
18
Depa men o Ea h Sys em Science, Tsinghua Uni e si y, Beijing 100084, China
Na ion. In Geosci Remo e Sens. Symp.:6512–6515.
Zheng, Y., Wu, G., 2022. YOLO 4-Li e – Based U ban Plan a ion T ee De ec ion and
Posi ioning wi h High-Resolu ion Remo e Sensing Image y. 9(Janua y):1–12
h ps://doi.o g/10.3389/ en s.2021.756227.
Zheng, J., Yuan, S., Wu, W., Li, W., Yu, L., Fu, H., Coomes, D., 2023. Su eying coconu
ees using high- esolu ion sa elli e image y in emo e a olls o he Paci ic Ocean.
Remo e Sens En i on. 287 h ps://doi.o g/10.1016/j. se.2023.113485.
Zhou, C., Ye, H., Sun, D., Yue, J., Yang, G., Hu, J., 2022. An au oma ed, high-
pe o mance app oach o de ec ing and cha ac e izing b occoli based on UAV
emo e-sensing and ans o me s: A case s udy om Haining, China. In J Appl Ea h
Obs Geoin [in e ne ]. 114 (July), 103055 h ps://doi.o g/10.1016/j.
jag.2022.103055.
Zhou, B., Zhao, H., Puig, X., Fidle , S., Ba iuso, A., To alba, A., 2017. Scene Pa sing
h ough ADE20K Da ase . In: In: P oc IEEE Con Compu is Pa e n Recogni . [place
Unknown], p. p. page 4..
Zhu X, Su W, Lu L, Li B, Wang X, Dai J. 2020. De o mable DETR: De o mable
T ans o me s o End- o-End Objec De ec ion [In e ne ]. :1–16. h p://a xi .o g/
abs/2010.04159.
R. Al-Ruzouq e al.