1
Vol.:(0123456789)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s
Remo e sensing based o es
co e classi ica ion using machine
lea ning
Gouha Aziz
1, Nas u Minallah
3, Aami Saeed
1, Ja osla F nda
2,4* & Waleed Khan
3
Pakis an alls signi ican ly below he ecommended o es co e age le el o 20 o 30 pe cen o o al
a ea, wi h less han 6 pe cen o i s land unde o es co e . This de iciency is p ima ily a ibu ed o
illici de o es a ion o wood and cha coal, coupled wi h a ailu e o emb ace ad anced echniques
o o es es ima ion, moni o ing, and supe ision. Remo e sensing echniques le e aging Sen inel-2
sa elli e images we e employed. Bo h single-laye s acked images and empo al laye s acked images
om a ious da es we e u ilized o o es classi ica ion. The applica ion o an a i icial neu al ne wo k
(ANN) supe ised classi ica ion algo i hm yielded no able esul s. Using a single-laye s acked image
om Sen inel-2, an imp essi e 91.37% aining o e all accu acy and 0.865 kappa coe icien we e
achie ed, along wi h 93.77% es ing o e all accu acy and a 0.902 kappa coe icien . Fu he mo e,
he empo al laye s acked image app oach demons a ed e en be e esul s. This me hod yielded
98.07% o e all aining accu acy, 97.75% o e all es ing accu acy, and kappa coe icien s o 0.970
and 0.965, espec i ely. The andom o es (RF) algo i hm, when applied, achie ed 99.12% o e all
aining accu acy, 92.90% es ing accu acy, and kappa coe icien s o 0.986 and 0.882. No ably, wi h
he empo al laye s acked image o he Sen inel-2 sa elli e, he RF algo i hm eached excep ional
pe o mance wi h 99.79% aining accu acy, 96.98% alida ion accu acy, and kappa coe icien s o
0.996 and 0.954. In e ms o o es co e es ima ion, he ANN algo i hm iden i ied 31.07% o al o es
co e age in he Dis ic Abbo abad egion. In compa ison, he RF algo i hm eco ded a sligh ly highe
31.17% o he o al o es ed a ea. This esea ch highligh s he po en ial o ad anced emo e sensing
echniques and machine lea ning algo i hms in imp o ing o es co e assessmen and moni o ing
s a egies.
Backg ound
The o es ecosys em plays a i al ole in p ese ing en i onmen al equilib ium h ough pollu ion mi iga ion,
lood egula ion, and soil e osion p e en ion. The Food and Ag icul u al O ganiza ion ecommends a o es
co e o 20–30% o a coun y1. Pakis an has a limi ed o es co e , comp ising 5.1 pe cen o he o al land a ea,
equi alen o 4.478 million hec a es2. This ansla es o jus 0.021 hec a es pe pe son, signi ican ly below he
global a e age o 1 hec a e pe pe son. O e he pas hi y yea s, o e 60 pe cen o he Himalayan Fo es has
unde gone des uc ion3. The sca ci y o o es s in Pakis an can be a ibu ed o he apid g ow h o popula ion and
po e y, coupled wi h a lack o awa eness among he people. The p ima y d i e s o de o es a ion in he coun y
a e he ex ac ion o wood, uel, and cha coal by he local popula ion2. Howe e , Ce ain egions in Pakis an,
including Manseh a, Abbo abad and Swa , boas ich biodi e si y wi h o e 430 ee species. The con en ional
and manual app oaches o supe ising o es s p esen challenges in e ms o being ime-consuming, expensi e,
and labou -in ensi e. The ask o physically isi ing o es s o documen in o ma ion abou each ee is bo h
challenging and cos ly. Moni o ing becomes pa icula ly challenging in hilly a eas, especially du ing ha sh and
cold wea he condi ions when hese a eas a e co e ed in snow. In his age o echnological p og ess, i is essen ial
o he go e nmen o p io i ize he in eg a ion o ad anced and scien i ic echnologies, such as Remo e Sensing4,
o e ec i ely manage de o es a ion.
OPEN
1Depa men o Compu e Science and In o ma ion Technology, Uni e si y o Enginee ing and Technology,
Peshawa , Pakis an. 2Depa men o Quan i a i e Me hods and Economic In o ma ics, Facul y o Ope a ion
and Economics o T anspo and Communica ion, Uni e si y o Zilina, Zilina, Slo akia. 3Na ional Cen e o Big
Da a and Cloud Compu ing, Uni e si y o Enginee ing and Technology, Peshawa , Pakis an. 4Depa men o
Telecommunica ions, Facul y o Elec ical Enginee ing and Compu e Science, VSB Technical Uni e si y o Os a a,
70800Os a a,CzechRepublic. *email:ja osla [email p o ec ed]
2
Vol:.(1234567890)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
Remo e sensing echnologies
Remo e Sensing employs sa elli es and senso s o s udy he Ea h’s su ace, p o iding aluable in o ma ion om a
dis ance. Comme cial sa elli es like Sen inel-2, Modis, and Landsa o e enhanced spa ial, spec al, and empo al
esolu ion, p o iding open da a o emo e sensing. Sen inel-2, wi h i s 13 mul ispec al bands, including he
ege a ion ed edge bands, o e s 5days o empo al da a o egula analysis. Sen inel-2’s capabili ies make i
well-sui ed o de ailed o es analysis, change de ec ion, and comp ehensi e ea u e analysis. Va ious echniques
simpli y he es ima ion p ocess in emo e sensing, yielding no ably accu a e esul s in de ec ing and es ima ing
di e en o es ypes, o e ing insigh s in o hei heal h and ma u i y. In ou esea ch, we u ilized Sen inel-2’s
empo al da a o o es analysis, bene i ing om i s ea u es as an open da a sa elli e. Sen inel-2 p o es o be a
aluable esou ce, con ibu ing o an enhanced unde s anding o o es s, encompassing hei heal h and ma u i y
s a us.
Ou designa ed s udy a ea encompasses Dis ic Abbo abad, loca ed wi hin he Haza a Di ision o he
Khybe Pakh unkhwa p o ince in Pakis an. This dis ic alls unde he We Moun ains Ag i Ecozone, ea u ing
e dan hills and is widely ecognized as a popula summe eso des ina ion. Th ough he u iliza ion o empo al
da a om Sen inel-2, ou esea ch yielded no ewo hy esul s. Equa ion (1) p o ides he o mula o calcula ing
he a ea o he Sen inel-2 image.
Li e al. employed mul ispec al Sen inel-2 sa elli e image y5 o e alua e he e ec i eness o o es - ype
mapping in Shang i-La, he adminis a i e egion o Yunnan P o ince, China. They applied he Random Fo es
algo i hm wi hin he Google Ea h Engine (GEE)5, wi h a p ima y ocus on iden i ying and de ec ing a ious
o es ypes. The s udy aimed o assess he Random Fo es algo i hm’s e icacy wi hin he GEE pla o m and
dis inguish a ia ions in he main o es ypes ac oss an ex ensi e a ea. Fu he mo e, he esea ch aimed o
es ima e c i ical ea u es o o es classi ica ion. The analysis success ully iden i ied eigh dis inc o es co e
ypes, achie ing a 95.76% accu acy in dis inguishing be ween o es and non- o es a eas, along wi h a Kappa
coe icien o 91.34%. The u iliza ion o he Google Ea h Engine pla o m played a pi o al ole in e ec i ely
moni o ing he dynamic changes in o es co e .
Con en ional app oaches o moni o ing, classi ying, and es ima ing obacco c op yield a e expensi e and
ime-consuming. The absence o an ad anced sys em u ilizing s a e-o - he-a emo e sensing echnologies o
moni o ing, classi ica ion, and yield es ima ion o obacco c ops was e iden in Pakis an. To b idge his gap, Khan
e al. in collabo a ion4 wi h he Pakis an Tobacco Boa d (PTB), conduc ed esea ch o es ablish an inno a i e
machine lea ning mechanism. They employed empo ally laye -s acked Sen inel-2 sa elli e da a o es ima e
obacco c ops in Pakis an. Fo he de ec ion o obacco c ops, he esea che s de ised a model based on an
A i icial Neu al Ne wo k. Implemen ing he A i icial Neu al Ne wo k classi ie wi h a single image4 achie ed
an O e all accu acy o 88.49%, which was u he imp o ed o 90.45% h ough he applica ion o NDVI s acking.
No ably, h ough expe imen s wi h empo ally s acked image y, hey a ained an O e all accu acy o 95.81%,
ma king a signi ican 7.32% imp o emen o e he benchma k scheme.
Like many o he coun ies, China expe iences he e ec s o Land Use Land Co e (LULC) changes. In ackling
his challenge in he Ganan P e ec u e om 2000 o 2018, Liu e al.6 u ilized he dense ime s acking o mul i-
empo al Landsa images and implemen ed he andom o es algo i hm on he Google Ea h Engine (GEE)
pla o m o LULC mapping. The classi ica ion accu acy o he en i e da ase ell wi hin he ange o 89.14% o
91.41% and Kappa Coe icien 0.86. The p ima y land use and land co e (LULC) ca ego ies in he s udy a ea
we e g assland, making up 50% o he o al a ea, and o es , encompassing 25%.
Fo es dynamics esul om a ious ac o s, wi h seasonal in luences playing a signi ican ole. In esponse,
Jiang e al.7 in oduced Fo es -CD, a model ha u ilizes high- esolu ion images (VHR). This model employs
an encode –decode a chi ec u e, in eg a ing backg ound in o ma ion. The encode , d i en by he Swin
T ans o me , sys ema ically ex ac s change ea u es, e ec i ely mimicking global in o ma ion. Con e sely, he
Fo es Change De ec ion decode employs he ea u e py amid ne wo k o eco e used in o ma ion and ea u e
scales a di e en le els. Analysis o an ex ensi e o es da ase indica es ha he Fo es -CD ne wo k, u ilizing
VHR images, a ains a highe F1 sco e. Addi ionally, he ou comes om Fo es -CD demons a e a dec ease in
pseudo changes.
The Random Fo es machine lea ning algo i hms ind equen applica ions in da a classi ica ion8,9, objec
ecogni ion10,11, and image segmen a ion12. Feng e al.13 in oduced a no el aining sample selec ion me hod
speci ically designed o Random Fo es Modeling in g eenhouse iden i ica ion using Supe View-1 image y.
This inno a i e app oach enhances classi ica ion accu acy and gene aliza ion capabili ies. The new Random
Fo es Modeling allows o he au oma ic selec ion o high-quali y aining samples, esul ing in high-p ecision
classi ica ion. Fu he mo e, he esea che s an icipa e ha his imp o ed and ad anced Random Fo es model
can ex end i s u ili y o iden i y a ious g ound objec s such as oads and buildings.
To add ess challenges in ad ancing Remo e Sensing echnologies, Benson e al.14 in oduced mul imodal
emo e sensing model o o es pa ame e es ima ion. This app oach u ilizes Ligh De ec ion and Ranging
(LiDAR), pola ime ic ada , and nea -in a ed passi e op ical sensing pla o ms, coupled wi h physics-based
models. These models p o e bene icial in p ecisely es ima ing abo eg ound biomass and measu ing Canopy
Heigh in homogeneous a eas. The o es pa ame e es ima ion algo i hm employs a combina ion o geome ic
and elec omagne ic senso model me hods. Despi e ha ing minimal inpu in o ma ion, his in eg a ed me hod
yields accu a e esul s o es ima ing o es s uc u e, along wi h minimal oo mean squa e e o s.
In o de o p ecisely map and iden i y spa io empo al changes E ani a d e al.15 ca ied ou a h ee-decade
s udy in I an wi h a ocus on mang o e habi a s. The Subme ged Mang o e Recogni ion Index (SMRI), a ecen ly
(1)
A ea
=
(To alPixel ×100m
2
)
1000 ×1000
×100 =
Hec a es
3
Vol.:(0123456789)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
de eloped echnique, and Landsa da a om 1990 o 2020 we e used in he s udy. In he p ocess o Mang o e
mapping, he esea che s u ilized ou ege a ion indices in conjunc ion wi h eigh mang o e-speci ic indices. The
s udy ound SMRI o be a pa icula ly e ec i e index. U ilizing long- e m Landsa da a, he es ima ed mang o e
co e age in I an was app oxima ely 13,000 ha in he yea 2020.
Wallne e al.16 add essed he dynamic and impac ul changes occu ing in he Cen al Eu opean o es
ecosys em, d i en by clima e unce ain ies and shi s in wea he pa e ns. In esponse o his, hey employed
sa elli e da a om ZiYuan-3 (ZY-3) wi hin a Remo e Sensing-guided Fo es in en o y amewo k. The objec i e
was o educe he equi ed ield sample size while analysing he s anda d g id in en o y. The u iliza ion o
3D ZY-3 demons a ed i s sui abili y in suppo ing o es in en o y by e ec i ely minimizing sample size and
enhancing in en o y equencies.
Sunda ban a mang o e o es si ua ed a Nijhum Na ional Pa k17 aces challenges in he deg ada ion o he
o es co e . A s udy was conduc ed by Islam e al. in NDP o ind ou he decades’ changes in he mang o e o es
by using GIS ools and emo e sensing a ailable da a. They used maximum likelihood classi ica ion echniques
by using Landsa images o 3 decades om 1990 o 2020. SAVI and NDVI-based classi ica ion is pe o med o
o es co e changes in compa ison wi h supe ised classi ica ion. Wi h his wo k, hey ind ou ha in he i s
decade om 1990 o 2000 almos one- hi d o de o es a ion occu ed. Howe e , in he las decade inc ease o
310.32ha ha e eco ded in mang o e o es co e .
De o es a ion changes he o es s uc u e, unc ionali y, and ecosys em p ocess18. Challenges acing
de o es a ion a e he es ima ion o emissions and iden i ying he a ea a ec ed and he o al amoun o biomass
los . Till now no eliable me hod is es ablished o iden i y he causes o de o es a ion o moni o o es i e, ca le
g azing, and uelwood collec ion. High spa ial and empo al images a e used o de ec small-scale dis u bances.
Howe e , using high- esolu ion images is cos ly oo. Fo o es i es and de ec ing logging, Gao e al.18 sugges ed
he SMA Remo e sensing me hod o he de ec ion o de o es a ion. To measu e he in ensi y o de o es a ion
Lida and ada a e sui able because o hei capaci y o measu e he 3D o he o es s uc u e and biomass
measu emen .
In he las decades, de o es a ion and woodland is g ea ly a ec ed by na u al disas e s19. To de ec he ea ly
smoke and lame a ious emo e sensing echnologies sys ems and algo i hms a e used by Ba mpou is e al.19.
Te es ial, ai bo ne, and spacebo ne-based sys ems a e iden i ied. La ge Ea h Obse a ion Sa elli e p o ed o
be success ul in wide- ange b oadcas ing in ea ly smoke and lame de ec ion. CubeSa s is a low-Ea h-O bi ing
sa elli e ha has a signi ican ad an age o e adi ional sa elli es in smoke de ec ion and i e de ec ion, hey a e
also economical, be e in empo al esolu ion, ha e good esponse ime, and in be e co e age.
A Spa io- empo al s udy has been conduc ed by Negassa e al.20 on Ko mo o es which is si ua ed in he
Gu o Gi a Dis ic o he Eas Wollega zone o E hiopia o ind he s a us o he Fo es co e by using he GIS
and Remo e Sensing echniques. By using geospa ial echniques, i is eco ded ha he o al a ea o dense o es
in Ko mo o es was 32.73% in 1991. Which dec eased o 26.16% in 2002. The o es u he dec eased o 20.5%
in 2019. A dec ease in he open o es is also eco ded i.e., 18.19% and 16.14% in 1991 and 2019 simul aneously.
Howe e , a conside able amoun o inc ease in ag icul u al land is eco ded om 24.78% in 1991 o 29.21% and
33.50% in he yea s 2002 and 2019, espec i ely. This s udy sugges s policy in e en ions o p o ec he Ko mo
o es p io i y a ea om loss and deg ada ion.
Biomass mapping is a i al and p ac ical ool in he ealm o o es managemen , pa icula ly o moni o ing
o es s and e alua ing de o es a ion p ocesses. Fo his pu pose, Sha i i e al.21 conduc ed a s udy aimed o employ
Mul i a ia e Rele ance Vec o Reg ession (MVRVR) as a Bayesian model wi h a ke nel-based amewo k o
p edic ing abo e-g ound biomass (AGB) in he Hy canian o es s o I an. Using ield da a and mul i- empo al
PALSAR backsca e alues o T aining and Tes ing, he esea che s compa ed he esul s wi h al e na i e
me hods such as mul i a ia e linea eg ession (MLR), mul ilaye pe cep on neu al ne wo k (MLPNN), and
suppo ec o eg ession (SVR). The indings e ealed ha he SVR model ou pe o med o he s, especially
a he lowes sa u a ion poin . The MVRVR model signi ican ly enhanced AGB es ima ion p ecision, showing
excep ional pe o mance, pa icula ly insi ua ions in ol ing he maximum sa u a ion poin .
In he ealm o emo e sensing, hype spec al images (HSIs) dis inguish hemsel es as a aluable sou ce o
in o ma ion, owing o hei dis inc i e ea u es applicable in a ious con ex s. Bu s ill due o many easons he
hype spec al images pe o mance a e educing due o many easons especially o he limi ed numbe o samples.
In o de o imp o e he HIS accu acy Ghade izadeh e al.22 p oposed he c ea ion o a classi ica ion model o
hype spec al images (HSI) named MDBRSSN, an ac onym o Mul iscale Dual-B anch Residual Spec al–Spa ial
Ne wo k wi h A en ion. The p oposed model unde wen expe imen s on ou da ase s, showcasing i s excellence
compa ed o s a e-o - he-a me hods, pa icula ly in scena ios wi h a es ic ed numbe o T aining samples.
The p oposed model achie ed O e all accu acies o 99.64%, 98.93%, 98.17%, and 96.57% wi h only 1%, 1%,
5%, and 5% o labelled da a o T aining, espec i ely. These esul s su pass hose o s a e-o - he-a me hods.
Va ious ac o s, including looding, con ibu e o de o es a ion. The impo ance o implemen ing a eal- ime
moni o ing sys em o e alua ing lood isks and imp o ing disas e esponse imes canno be o e s a ed. In a
s udy conduc ed by A. Shi i i23, he classi ica ion o SAR da a in ol ed employing h esholding, machine lea ning
algo i hms, and an objec -based me hod. The h esholding p ocess played a c ucial ole in iden i ying looded
egions. Upon compa ing he esul s, he machine lea ning algo i hm exhibi ed signi ican success. These indings
highligh he impo ance o Sen inel-1 images as c ucial da a o e ining me hodological guides, indica ing hei
po en ial as a no el esou ce o moni o ing lood isks.
Remo e sensing p o es bene icial in lood con ol e o s, as loods can con ibu e o de o es a ion. Floods
pose a po en ial h ea in nume ous loca ions, wi h heigh ened suscep ibili y obse ed in o es s, he ag icul u al
indus y, and in as uc u e si ua ed nea i e s. This ulne abili y is a ibu ed o he widesp ead impac o loods
on o es s and ag icul u al land ac oss di e se a eas. Ta iq e al.24 implemen ed an expe imen al app oach o
e alua e he ulne abili y o lood mapping in he no he n a eas o Punjab, Pakis an, h ough he in eg a ion
4
Vol:.(1234567890)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
o FR and AHP echniques. Eigh pa ame e s we e delibe a ely selec ed o de e mine he weigh o ela i e
signi icance, employing pai wise ma ix co ela ion. Six pa ame e s a e om Remo e Sensing image y including
Sen inel-2 sa elli e. The lood haza d map was gene a ed using A cGIS algo i hms o iden i y he ex emely high,
mode a e, and low lood zones in he inal ou pu .
Se e al esea ch s udies ha e ecommended he u iliza ion o emo e sensing image y o en i onmen al
moni o ing. Fo his pu pose Mohammadi e al.25 employed Sen inel-1 SAR da a, along wi h he u iliza ion
o Sen inel-2 image y, o p omp ly iden i y oil spills in he Pe sian Gul . They employed VV-pola ized images
om Sen inel-1 SAR da a o illus a e he exis ence o oil pa ches. Sen inel-2 da a dis inguishes i sel as a highly
e ec i e senso o de ec ing oil slicks, hanks o i s excep ional spa ial, spec al, and empo al esolu ion. They
ecommended ha I use s lack access o ield da a, i is ad isable o employ he OBIA me hod o assess he
accu acy o esul s de i ed om SAR da a.
Zaman e al.26 conduc ed a s udy wi h he goal o de e mining he ideal zones o sa on cul i a ion in
Miyaneh. The esea ch u ilized Landsa 8 sa elli e images and applied he Weigh ed Linea Combina ion (WLC)
me hod. The s udy pe iod ex ended om No embe 2019 o May 2020. The esul s indica ed ha he p ime
loca ions o sa on cul i a ion in he examined a ea a e concen a ed in a s ip unning om he sou hwes o
he sou heas , along wi h speci ic no he n egions.
Sha i i e al.27 in he ield o emo e sensing ad oca ed o he use o Pola ime ic Syn he ic Ape u e Rada
echnology (PolSAR) when SAR images ace challenges due o speckle noise. They highligh ed he e ec i eness
o PolSAR in cap u ing images ac oss di e en pola iza ions as a p ac ical and al e na i e solu ion. The Fas
ICA me hod is s ongly endo sed o i s p o iciency in educing speckle, p ese ing de ails, and demons a ing
ema kable speed.
Kossa i e al.28 in oduced a apid me hod o dimensioning he A i ude De e mina ion and Con ol Sys em
(ADCS) o Ea h obse a ion sa elli es. They applied a ma ching diag am echnique, well-es ablished in ai c a
indus ies o ai c a design. The s udy emphasized spa ial and empo al esolu ions as he key pe o mance
equi emen s (PRs).
Yuh e al.29 conduc ed a compa a i e analysis o ou dis inc machine lea ning algo i hms o moni o changes
in Land Use and Land Co e (LULC) in no he n Came oon. Thei s udy u ilized Landsa 7 ETM and Landsa 8
OLI image y om No embe 2000 and No embe 2020. KNN, SVM, RF, and ANN we e among he algo i hms
ha we e assessed, all o hem showed a commendable le el o accu acy. The KNN algo i hm p oduced a Kappa
Coe icien o 89% and an O e all Accu acy o 91.1% o he yea 2020. Likewise, he ANN algo i hm p oduced
a high 94% Kappa Coe icien along wi h a high 95.8% O e all Accu acy. The RF algo i hm demons a ed a 94%
Kappa Coe icien and an O e all Accu acy o 90.3%. Wi h a Kappa Coe icien o 87%, he O e all accu acy
o SVM was 88.6%. The s udy’s conclusions showed ha he e was a no able educ ion in he amoun o o es
co e be ween 2000 and 2020 as a esul o he con e sion o hese o es ed egions in o ag icul u al land, mos ly
o he p oduc ion o c ops.
Mo adi e al.30 explo ed changes in o es co e in he Zag os Moun ains, Wes e n I an, u ilizing Landsa
image y. They applied a CNN deep lea ning algo i hm o disce n al e a ions in he landscape. The esul s o hei
s udy e ealed a subs an ial decline in o es co e o e he pas hi y yea s. The CNN algo i hm p o ed e ec i e
in dis inguishing oak o es om wa e and ag icul u al classes. Thei esea ch achie ed a high accu acy o 97%
and a Kappa coe icien o 94.7% when u ilizing Landsa TM image y. Simila ly, wi h Landsa ETM image y,
hey a ained a 95% O e all accu acy and a Kappa coe icien o 94.1%.
This wo k is o ganized as ollows. The me hods and ma e ial a e discussed in “Me hods and ma e ial” while
he esul s o ou expe imen a ion a e discussed in “Expe imen s and esul s”. Discussion on ou p oposed
algo i hms and ob ained esul s a e en ailed in “Discussion”. Las ly, we succinc ly conclude in “Conclusion”
along wi h some u u e p oposi ions.
Me hods and ma e ial
To ini ia e he o es co e de ec ion p ojec , he Abbo abad egion has been designa ed as he pilo a ea. This
geog aphically di e se a ea is cha ac e ized by olling hills and en eloped by lush g een moun ains, making
i a enowned summe e ea admi ed o i s o es ed cha m. The p ocess o ga he ing accu a e da a and
geog aphical poin s un olded in mul iple s ages. Ini ially, on-si e inspec ions a e conduc ed o ca ego ize he
classes, and he ollowing ou classes a e chosen.
i.Fields
ii.Fo es
iii.U ban a ea
i .Sh ubs.
Secondly, o enhance he eliabili y o he da a, he shape ile o Abbo abad is acqui ed om he Pakis an
Fo es Ins i u e in Peshawa , a well- espec ed o ganiza ion in he coun y. To ensu e da a accu acy and u ilize
cu ing-edge echnology, he "Geosu ey App," an indigenous applica ion de eloped by he Na ional Cen e o
Big Da a and Cloud Compu ing (NCBC) in Peshawa (h ps:// www. ncbcp eshaw a . com), is employed. U ilizing
he “GeoSu ey App” o he Fields class, we me iculously choose and ou line a o al o 900 polygons. The Fo es
class comp ises 901 ca e ully selec ed polygons. Likewise, o he U ban class, we iden i y and pick 900 poly-
gons. The da a pe aining o Sh ubs polygons is e ie ed om he Fo es y Planning and Moni o ing Sys em in
Peshawa . Figu e1 displays he shape ile o he Abbo abad dis ic .
In he hi d phase, Sen inel-2 sa elli e images a e acqui ed. The expe imen a ion in ol es wo king wi h bo h a
single downloaded image and a empo ally sequenced downloaded image. Speci ically, a Sen inel-2 single image
om Oc obe 27 h, 2021, o Dis ic Abbo abad is ob ained. Fo he empo al image se , ou images om
Sep embe 2nd, 2021, Oc obe 27 h, 2021, No embe 11 h, 2021, and Decembe 11 h, 2021, a e downloaded.
5
Vol.:(0123456789)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
The ollowing p ocedu es a e execu ed on hese downloaded images:
1. P ep ocessing is ca ied ou using SNAP Desk op, wi h esampling pa ame e s being con igu ed.
2. The esampled da a is subsequen ly employed o u he p ocessing.
3. All he images a e laye -s acked.
4. A mask is cons uc ed, and his mask is applied o ex ac he Abbo abad image om he shape ile.
5. A CSV ile is gene a ed o he egion o in e es (ROI).
In he ou h phase, ollowing he es ablishmen o he Remo e Sensing Da ase , expe imen s a e conduc ed
using A i icial In elligence Neu al Ne wo k algo i hms wi h di e se pa ame e s and Random Fo es algo i hm.
A i icial Neu al Ne wo ks (ANNs) p o icien ly manage di e se emo e sensing da a, inco po a ing bo h mul i-
spec al and hype spec al image y. Thei e sa ili y allows o seamless adap a ion o he di e se spec al bands
and esolu ions commonly encoun e ed in a ious emo e sensing applica ions. A i icial Neu al Ne wo ks
(ANNs) ha e p o en e ec i e ac oss di e se applica ions in emo e sensing, such as land co e classi ica ion8,
objec de ec ion11, ege a ion and c ops moni o ing4, and e ain analysis8. A i icial Neu al Ne wo ks ind
applica ions in a ious domains such as image p ocessing and cha ac e ecogni ion31, classi ica ion32, o ecas -
ing, enhancemen 33, analysis34, es ima ion, and p edic ion35. Thei adap abili y ende s hem sui able o a b oad
spec um o asks wi hin he ield. Speci ically, ne wo ks wi h a signi ican numbe o pa ame e s may be p one
o o e i ing, cap u ing noise o speci ic pa e ns in he T aining da a ha may no gene alize e ec i ely o new,
unseen da a. The whole p ocedu e is depic ed in Fig.2.
Wi hin he machine lea ning domain, Random Fo es (RF) is widely acknowledged as a equen ly employed
ensemble lea ning echnique sui able o bo h classi ica ion and eg ession asks. In his wo k, we op ed o RF
due o i s basic ensembled s uc u e and compu a ional easibili y as compa ed o o he bagging and boos ing
based ensemble lea ning echniques (i.e., XGBoos , CATBoos e c.). As i can be seen in Fig.3, i ope a es as an
ensemble model, gene a ing mul iple decision ees using andomly selec ed subse s o T aining samples and
a iables. The Random Fo es (RF) classi ie demons a es educed sensi i i y36 in compa ison o o he s eam-
lined machine lea ning classi ie s conce ning he quali y o T aining samples and o e i ing conce ns. Random
Fo es s may equi e subs an ial compu a ional esou ces, especially when dealing wi h a subs an ial numbe o
ees and ea u es. The T aining and e alua ion o a la ge ensemble can be compu a ionally demanding.
Figu e1. Dis ic Abbo abad shape ile.

6
Vol:.(1234567890)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
A i icial neu al ne wo ks
Classi ica ion
Khan and Minallah4 employed he A i icial Neu al Ne wo k algo i hm in hei s udy. They emphasized ha
a i icial neu ons se e as he undamen al componen s o A i icial Neu al Ne wo ks37. Fo he implemen a-
ion o a neu al ne wo k, a minimum o h ee laye s is equi ed, namely he Inpu Laye , he Hidden Laye , and
he Ou pu Laye , as depic ed in Fig.4. The Inpu Laye ansmi s inpu o he Hidden Laye , also known as he
middle laye , which add esses p oblems by u ilizing mul iple P ocessing Elemen s (PE). The Ou pu Laye , he
inal laye , gene a es ou pu based on gi en inpu pa ame e s.
Ini ially, e e y A i icial Neu al Ne wo k goes h ough T aining o unde s and and compa e i s eac ions when
gi en new pixels, igu ing ou which side o a linea sepa a ing line hey all on35. A e ha , he p ocessing pa
depends on he inpu s and weigh s om he laye be o e. The P ocessing Elemen (PE) handles a se o inpu s,
like X = × 1, × 2, × 3……xN, whe e w is he connec ion weigh , θ is a bias, and Z0 is he Ou pu Laye .
Feed o wa d neu al ne wo k
A Feed-Fo wa d Neu al Ne wo k (FFNN) u ilizes a laye o in e connec ed neu ons o he p ocessing and
ansmission o in o ma ion. I alls unde he ca ego y o A i icial Neu al Ne wo ks ha use a supe ised
classi ica ion me hod o app oxima e a classi ie . Du ing FFNN T aining, adjus men s a e made o he weigh s
a he nodes wi h he goal o educing he dispa i y be ween he ac i a ion o he ou pu nodes and he inpu .
The ne wo k mus lea n he app op ia e weigh s and biases o p ecisely classi y he inpu da a. Ce ain ea u es
o Feed-Fo wa d Ne wo ks encompass:
• P ocessing Elemen s (PEs) a e s uc u ed in laye s, whe ein he inpu laye accep s inpu da a, he ou pu laye
p oduces ou pu s, and he in e media y laye s, known as hidden laye s, do no ha e ex e nal connec ions
bu exclusi ely in e ac wi h o he laye s wi hin he model.
• In o ma ion a els in a single di ec ion, mo ing om he inpu laye h ough he hidden laye and eaching
he ou pu laye .
• FFNNs a e non-cyclic, signi ying he absence o eedback connec ions in he ne wo k, which inhibi s neu ons
om exchanging in o ma ion wi h each o he in a e e se manne .
Figu e2. A i icial neu al ne wo k algo i hm me hodology.
7
Vol.:(0123456789)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
Connec ions a e es ablished, wi h a P ocessing Elemen (PE) such as H1 connec ed o inpu s x1, x2, and x3,
and H2 linked o inpu s x1, x2, and x3, as depic ed in Figu e4. Equa ion (2) accoun s o all he weigh s in play.
PE compu es he ma ix p oduc o he hidden laye wi h hese weigh s, includes i s own bias, and subsequen ly
applies he ac i a ion unc ion. The ma ix compu a ion is p esen ed as:
=




W11 W12 W13
W21 W22 W23




∗





x
1
x2
x3





(2)
=




W11 ∗x1+W12 ∗x2+W13 ∗x3
W21 ∗x1+W22 ∗x2+W23 ∗x3




(3)
H
1=

WIJ ∗li+Bi

Figu e3. A i icial neu al ne wo k algo i hm me hodology.
Figu e4. A i icial neu al ne wo k algo i hm.
8
Vol:.(1234567890)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
The Hidden laye s’ alue is compu ed by adding up he p oduc s o he inpu alues and hei co esponding
weigh s, as desc ibed in Eq.(3). The key pu pose o his compu a ion is o asce ain how he sys em should be
adjus ed o ma ch he ou pu wi h he desi ed a ge . E en sligh modi ica ions in weigh s can esul in subs an ial
changes in ou pu 30. This a ibu e acili a es he lea ning p ocess.
Pa ame e s o neu al ne wo k
Be o e es ablishing he pa ame e s, ce ain decisions mus be made, including de e mining he numbe o laye s
o be employed. Gene ally, h ee laye s a e deemed sa is ac o y, wi h he i s designa ed as he Inpu Laye , he
subsequen one as he Hidden laye , and he las one as he Ou pu laye . The inpu laye ypically ecei es nodes
co esponding o he numbe o componen s ( ea u es) in he pixel ec o s. The ollowing pa ame e s in Table1
ha e been se o he ANN algo i hm.
Random o es algo i hm
The Random Fo es is a supe ised classi ica ion machine lea ning algo i hm ha cons uc s and g ows mul iple
decision ees o o m a " o es ." I is employed o bo h classi ica ion and eg ession p oblems shown in Fig.5.
In classi ica ion, i builds decision ees on a ious samples and akes a majo i y o e, while in eg ession, i
calcula es he a e age o di e en samples. A no able ea u e o he Random Fo es Algo i hm is i s abili y o
handle da ase s wi h ca ego ical a iables o classi ica ion, leading o imp o ed esul s.
How andom o es algo i hm wo k
The Random Fo es algo i hm employs Bagging o Boo s ap Agg ega ion Techniques. Bagging en ails gene a ing
mul iple T aining subse s om he sample T aining da a wi h eplacemen , and he ul ima e ou pu is decided by
he majo i y o o es. Boo s ap andomly selec s ows and ea u es om he da ase o c ea e sample da ase s o
each model. Agg ega ion consolida es hese sample da ase s h ough majo i y o ing o gene a e he inal ou pu .
Boo s ap Agg ega ion is e ec i e in mi iga ing he a iance o high- a iance algo i hms, like decision ees.
S eps in ol ed in andom o es algo i hm
(4)
O
=
(WIJ
∗
H
+
B)
Table 1. Pa ame e s o ANN algo i hm.
S. no. Pa ame e s Value/ ype
1 Lea ning a e 0.01
2 T aining momen um 0.900
3 T aining RMS exi c i e ia 0.100
4 Numbe o hidden laye 2
5 Numbe o aining i e a ions/epochs 50, 100, 200, 300
6 Ac i a ion unc ion Relu
7 No o neu ons in hidden laye 64
8 Ba ch size 32
Figu e5. Random o es algo i hm.
9
Vol.:(0123456789)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
1. Random Fo es s ope a e on a gi en da ase wi h N eco ds and K ou pu s, whe e N ep esen s he numbe
o samples, and K deno es he numbe o classes.
2. A decision ee is c ea ed o each se o samples o p oduce he ou pu .
3. In classi ica ion, he inal ou pu is de e mined by assigning g ea e impo ance o he majo i y o o es.
Sa i a Desdhan y and Rus am38 implemen ed he Random Fo es algo i hm in hei esea ch. In hei me h-
odology, hey de ine S = {(xi, yi)}, whe e xi ep esen s he nume ical ea u e, and yi co esponds o he espec i e
labels. Assuming he Random Fo es has T ea u es, and P deno es he numbe o ees in he o es , N ees
a e andomly selec ed in he Random Fo es , and each is employed o cons uc a decision ee. This p ocess is
epea ed P imes, and a each node, a small subse o ea u es is c ea ed. The bes ea u e o each subse is hen
de e mined. The ou come o his p ocedu e is a selec ed ea u e A ha achie es he highes sco e38: he algo i hm
is p esen ed in he able below38.
Ini ializa ion: A aining se S: = {(xi,yi)}, T ea u es, and numbe o ees in o es P
1. Selec M ees om he da ase , in o de o o cons uc a decision ee
2. Redo he p e ious s ep P imes
3. A each node:
4. Cons uc a small subse o F, call i
5. Sepa a e he mos app op ia e ea u es in
6. The ca ego y ha gains he majo i y o es will be gi en a new eco d
The Ou pu will be he selec ed ea u es ha ha e he highes accu acy sco e
Algo i hm Random o es .
The ollowing Pa ame e s ha e been se in Table2 o he Random Fo es algo i hm.
These ou comes play a c ucial ole in de e mining he O e all Accu acy and Kappa Coe icien o bo h he
T aining and Tes ing Da a se s.
O e all accu acy s ands ou as a equen ly used e alua ion me ic. I signi ies he a io o accu a ely classi ied
ins ances, o da a poin s, o he o al numbe o ins ances in a da ase . This me ic se es as a undamen al
benchma k o assessing he model’s pe o mance in e ms o co ec classi ica ions ac oss he en i e da ase .
The o mula o he Kappa coe icien is as ollows in Eq. (5):
The Kappa coe icien , also known as Cohen’s Kappa, is a s a is ic ha measu es he ag eemen be ween
obse ed and expec ed classi ica ion esul s while conside ing he possibili y o ag eemen occu ing by chance.
The o mula o he Kappa coe icien is as ollows in Eq.(6):
O e all Ag eemen is he p opo ion o obse ed ag eemen be ween he classi ied esul s and he e e ence
(g ound u h) da a.
Chance Ag eemen is he expec ed ag eemen due o chance. I is calcula ed based on he ma ginal
p obabili ies o ag eemen o each class.
Conce ning he T aining Da a, as delinea ed in Table3, we ha e selec ed 17,101 pixels o he Fields class,
33,045 pixels o he Fo es class, 3678 pixels o he Sh ubs class, and 9542 pixels o he U ban class. Rega ding
he Tes ing Da a, as indica ed in Table1, 7377 pixels a e chosen o he Fields class, 14,166 pixels o he Fo es
class, 2058 pixels o he Sh ubs class, and 4191 pixels o he U ban class.
Th ough he u iliza ion o he Random Fo es Supe ised machine lea ning Classi ica ion Algo i hm, we
a ained a T aining O e all accu acy o 99.79% and Tes ing O e all accu acy o 97%. Fu he mo e, he applica ion
(5)
O e all Accu acy
=
Sum o Co ec ly Classi ied Pixels
To al Numbe o Pixels
×
100%
(6)
Kappa Coe icien
=
O e all Ag emen
−
Chance Ag eemen
1
−
Chance Ag eemen
Table 2. Pa ame e o andom o es algo i hm.
S. no. Pa ame e Value
1 Maximum dep h 5, 10, 20
Table 3. To al numbe o aining and es ing pixels.
Class T aining pixel Tes ing pixels
Fields 17,101 7377
Fo es 33,045 14,166
Sh ubs 3678 2058
U ban 9542 4191
16
Vol:.(1234567890)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
u ilizing an A i icial Neu al Ne wo k (ANN) algo i hm and Sen inel-2 da a su passed his pe o mance wi h
an O e all accu acy o 97.75% and 0.965 Kappa Coe icien .
Conclusion
Pakis an aces challenges as a o es -poo coun y, wi h less han 6% o i s o al a ea co e ed by o es s. To add ess
his issue, ad anced echniques employing Machine Lea ning and Deep Lea ning algo i hms we e applied o
o es co e classi ica ion in Dis ic Abbo abad. No ably, he A i icial Neu al Ne wo k (ANN) and Random
Fo es algo i hms we e employed, yielding s a e-o - he-a esul s in e ms o O e all Accu acy and Kappa
Coe icien . The ANN algo i hm demons a ed ema kable pe o mance, achie ing a bes O e all Accu acy
o 97.75% and a Kappa Coe icien o 0.965. Howe e , o ackle he o e i ing p oblem inhe en in ANN, he
Random Fo es algo i hm was in oduced. This app oach esul ed in a commendable O e all Accu acy o 96.98%
and a Kappa Coe icien o 0.954, pa icula ly when using a maximum dep h o 20 o he Tempo al Laye
s acked image. Applying he ANN algo i hm o he en i e 166,103 hec a es a ea o Abbo abad e ealed a o es
co e o 51,613 hec a es, cons i u ing 31.07% o he o al dis ic a ea. Meanwhile, u ilizing he Random Fo es
algo i hm iden i ied a o al o es co e o 51,774 hec a es, equi alen o 31.17% o he Abbo abad dis ic
egion. To ele a e he p ecision o o es co e classi ica ion, inco po a ing hype spec al sa elli e image y is
ecommended. Addi ionally, o u u e enhancemen s, deep lea ning algo i hms such as Con olu ional Neu al
Ne wo ks (CNN), Long Sho -Te m Memo y ne wo ks (LSTM), and Ga ed Recu en Uni s (GRU) will be
explo ed o hei po en ial in ad ancing classi ica ion accu acy.
Da a a ailabili y
The da ase s used and/o analysed du ing he cu en s udy a e a ailable om he co esponding au ho on
easonable eques .
Recei ed: 25 Sep embe 2023; Accep ed: 27 Decembe 2023
Re e ences
1. 2019 Fo es y Sec o Re iew: Pakis an. h ps:// doi. o g/ 10. 4060/ ca486 9en (FAO, 2020).
2. In oduc ion o Landscapes o Pakis an. cap.unep.o g. Accessed 26 July 2010.
3. Rasheed, F., Yaqoob, S. & Ahmad, A. H. Fo es ypes o Pakis an. In Tex book o Applied Fo es y. (Uni e si y o Ag icul u e
Faisalabad).
4. Khan, W. e al. On he pe o mance o empo al s acking and ege a ion indices o de ec ion and es ima ion o obacco c op. IEEE
Access 8, 103020–103033. h ps:// doi. o g/ 10. 1109/ ACCESS. 2020. 29980 79 (2020).
5. Li, J., Wang, L., Fang, P., Xu, W. & Dai, Q. Fo es ype mapping a a egional scale based using mul i empo al Sen inel-2 image y.
In 2021 IEEE In e na ional Geoscience and Remo e Sensing Symposium IGARSS, B ussels, Belgium. 4228–4231 h ps:// doi. o g/ 10.
1109/ IGARS S47720. 2021. 95540 83 (IEEE, 2021).
6. Liu, C. e al. Land use/land co e changes and hei d i ing ac o s in he no heas e n Tibe an Pla eau based on geog aphical
de ec o s and Google Ea h Engine: A case s udy in Gannan P e ec u e. Remo e Sens. 12(19), 3139. h ps:// doi. o g/ 10. 3390/ s121
93139 (2020).
7. Jiang, J., Xiang, J., Yan, E., Song, Y. & Mo, D. Fo es -CD: Fo es change de ec ion ne wo k based on VHR images. IEEE Geosci.
Remo e Sens. Le . 19, 1–5. h ps:// doi. o g/ 10. 1109/ LGRS. 2022. 31886 36 (2022).
8. Sun, L. & Schulz, K. The imp o emen o land co e classi ica ion by he mal emo e sensing. Remo e Sens. 7(7), 8368–8390. h ps://
doi. o g/ 10. 3390/ s707 08368 (2015).
9. Pal, M. Random o es classi ie o emo e sensing classi ica ion. In . J. Remo e Sens. 26(1), 217–222. h ps:// doi. o g/ 10. 1080/
01431 16041 23312 69698 (2005).
10. T ochim, E. D., P akash, A., Kane, D. L. & Romano sky, V. E. Remo e sensing o wa e acks. Ea h Sp. Sci. 3(3), 106–122. h ps://
doi. o g/ 10. 1002/ 2015E A0001 12 (2016).
11. Gall, J., Yao, A., Raza i, N., Van Gool, L. & Lempi sky, V. Hough o es s o objec de ec ion, acking, and ac ion ecogni ion. IEEE
T ans. Pa e n Anal. Mach. In ell. 33(11), 2188–2202. h ps:// doi. o g/ 10. 1109/ TPAMI. 2011. 70 (2011).
12. Smi h, A. Image segmen a ion scale pa ame e op imiza ion and land co e classi ica ion using he Random Fo es algo i hm. J.
Spa . Sci. 55(1), 69–79. h ps:// doi. o g/ 10. 1080/ 14498 596. 2010. 487851 (2010).
13. Feng, T., Ma, H. & Cheng, X. G eenhouse ex ac ion om high- esolu ion emo e sensing image y wi h imp o ed andom o es .
In IGARSS 2020—2020 IEEE In e na ional Geoscience and Remo e Sensing Symposium, Waikoloa, HI, USA. 553–556 h ps:// doi.
o g/ 10. 1109/ IGARS S39084. 2020. 93241 47 (IEEE, 2020).
14. Benson, M. L., Pie ce, L., Be gen, K. & Sa abandi, K. Model-based es ima ion o o es canopy heigh and biomass in he Canadian
Bo eal Fo es using ada , LiDAR, and op ical emo e sensing. IEEE T ans. Geosci. Remo e Sens. 59(6), 4635–4653. h ps:// doi. o g/
10. 1109/ TGRS. 2020. 30186 38 (2021).
15. E ani a d, Y., Lo iNasi abad, M. & S e eńczak, K. Assessmen o I an’s mang o e o es dynamics (1990–2020) using Landsa
ime se ies. Remo e Sens. 14(19), 4912. h ps:// doi. o g/ 10. 3390/ s141 94912 (2022).
16. Wallne , A. e al. A emo e sensing-guided o es in en o y concep using mul ispec al 3D and heigh in o ma ion om ZiYuan-3
sa elli e da a. Fo . In . J. Fo . Res. 95(3), 331–346. h ps:// doi. o g/ 10. 1093/ o es y/ cpab0 55 (2022).
17. Islam, Md. R., Khan, Md. N. I., Khan, Md. Z. & Roy, B. A h ee decade assessmen o o es co e changes in Nijhum dwip na ional
pa k using emo e sensing and GIS. En i on. Chall. 4, 100162. h ps:// doi. o g/ 10. 1016/j. en c. 2021. 100162 (2021).
18. Gao, Y., Sku sch, M., Paneque-Gál ez, J. & Ghila di, A. Remo e sensing o o es deg ada ion: A e iew. En i on. Res. Le . 15(10),
103001. h ps:// doi. o g/ 10. 1088/ 1748- 9326/ abaad7 (2020).
19. Ba mpou is, P., Papaioannou, P., Dimi opoulos, K. & G ammalidis, N. A e iew on ea ly o es i e de ec ion sys ems using op ical
emo e sensing. Senso s 20(22), 6442. h ps:// doi. o g/ 10. 3390/ s2022 6442 (2020).
20. Negassa, M. D., Mallie, D. T. & Gemeda, D. O. Fo es co e change de ec ion using geog aphic in o ma ion sys ems and emo e
sensing echniques: A spa io- empo al s udy on Kom o p o ec ed o es p io i y a ea, Eas Wollega Zone, E hiopia. En i on. Sys .
Res. 9(1), 1. h ps:// doi. o g/ 10. 1186/ s40068- 020- 0163-z (2020).
21. Sha i i, A., Amini, J. & Ta eishi, R. Es ima ion o o es biomass using mul i a ia e ele ance ec o eg ession. Pho og amm. Eng.
Remo e Sens. 82(1), 41–49. h ps:// doi. o g/ 10. 14358/ PERS. 83.1. 41 (2016).

17
Vol.:(0123456789)
Scien i ic Repo s | (2024) 14:69 | h ps://doi.o g/10.1038/s41598-023-50863-1
www.na u e.com/scien i ic epo s/
22. Ghade izadeh, S., Abbasi-Moghadam, D., Sha i i, A., Ta iq, A. & Qin, S. Mul iscale dual-b anch esidual spec al–spa ial ne wo k
wi h a en ion o hype spec al image classi ica ion. IEEE J. Sel. Top. Appl. Ea h Obs. Remo e Sens. 15, 5455–5467 h ps:// doi. o g/
10. 1109/ JSTARS. 2022. 31887 32 (2022).
23. Sha i i, A. De elopmen o a me hod o lood de ec ion based on Sen inel-1 images and classi ie algo i hms. Wa e En i on. J.
35(3), 924–929. h ps:// doi. o g/ 10. 1111/ wej. 12681 (2021).
24. Ta iq, A. e al. Flash lood suscep ibili y assessmen and zona ion by in eg a ing analy ic hie a chy p ocess and equency a io
model wi h di e se spa ial da a. Wa e 14(19), 3069. h ps:// doi. o g/ 10. 3390/ w1419 3069 (2022).
25. Mohammadi, M., Sha i i, A., Hosseingholizadeh, M. & Ta iq, A. De ec ion o oil pollu ion using SAR and op ical emo e sensing
image y: A case s udy o he Pe sian Gul . J. Indian Soc. Remo e Sens. 49(10), 2377–2385. h ps:// doi. o g/ 10. 1007/ s12524- 021-
01399-2 (2021).
26. Zamani, A., Sha i i, A., Felega i, S., Ta iq, A. & Zhao, N. Ag o clima ic zoning o sa on cul u e in Miyaneh Ci y by using WLC
me hod and emo e sensing da a. Ag icul u e 12(1), 118. h ps:// doi. o g/ 10. 3390/ ag ic ul u e1201 0118 (2022).
27. Sha i i, A., Amini, J., S i Suman yo, J. T. & Ta eishi, R. Speckle educ ion o PolSAR images in o es egions using as ICA algo-
i hm. J. Indian Soc. Remo e Sens. 43(2), 339–346. h ps:// doi. o g/ 10. 1007/ s12524- 014- 0423-3 (2015).
28. Kosa i, A., Sha i i, A., Ahmadi, A. & Khoshsima, M. Remo e sensing sa elli e’s a i ude con ol sys em: Rapid pe o mance sizing
o passi e scan imaging mode. Ai c . Eng. Ae osp. Technol. 92(7), 1073–1083. h ps:// doi. o g/ 10. 1108/ AEAT- 02- 2020- 0030 (2020).
29. Yuh, Y. G., T acz, W., Ma hews, H. D. & Tu ne , S. E. Applica ion o machine lea ning app oaches o land co e moni o ing in
no he n Came oon. Ecol. In o m. 74, 101955. h ps:// doi. o g/ 10. 1016/j. ecoin . 2022. 101955 (2023).
30. Mo adi, E. & Sha i i, A. Assessmen o o es co e changes using mul i- empo al Landsa obse a ion. Assess. Fo . Co e Change
Using Mul i-Tempo al Landsa Obs. 25(2), 1351–1360. h ps:// doi. o g/ 10. 1007/ s10668- 021- 02097-2 (2023).
31. Topouzelis, K., Ka a hanassi, V., Ka a hanassi, V., Pa lakis, P. & Rokos, D. Po en iali y o eed- o wa d neu al ne wo ks o clas-
si ying da k o ma ions o oil spills and look-alikes. Geoca o In . 24(3), 179–191 (2009).
32. Lekha, L. S. & Kuma , S. S. “Classi ica ion and mapping o land use land co e change in Kanyakuma i dis ic wi h emo e sensing
and GIS echniques. En i on. Sci. Geog . 13(1), 9 (2018).
33. Mas, J. F. & Flo es, J. J. The applica ion o a i icial neu al ne wo ks o he analysis o emo ely sensed da a. In . J. Remo e Sens.
29(3), 617–663. h ps:// doi. o g/ 10. 1080/ 01431 16070 13521 54 (2008).
34. Jako le , S., Eglynas, T. & Voznak, M. Applica ion o neu al ne wo k p edic i e con ol me hods o sol e he shipping con aine
sway con ol p oblem in Quay c anes. IEEE Access 9, 78253–78265. h ps:// doi. o g/ 10. 1109/ ACCESS. 2021. 30839 28 (2021).
35. Tayyebi, A., Pijanowski, B. C. & Tayyebi, A. H. An u ban g ow h bounda y model using neu al ne wo ks, GIS and adial pa am-
e e iza ion: An applica ion o Teh an, I an. Landsc. U ban Plan. 100(1–2), 35–44. h ps:// doi. o g/ 10. 1016/j. landu bplan. 2010. 10.
007 (2011).
36. Mas o, P., Masiello, G., Se io, C. & Pepe, A. Change de ec ion echniques wi h syn he ic ape u e ada images: Expe imen s wi h
andom o es s and Sen inel-1 obse a ions. Remo e Sens. 14(14), 3323. h ps:// doi. o g/ 10. 3390/ s141 43323 (2022).
37. Khai unniza-Bejo, S., Mus a ha, S. & Wan Ismail, W. I. Applica ion o a i icial neu al ne wo k in p edic ing c op yield: A e iew.
J. Food Sci. Eng. 4, 1–9 (2014).
38. Desdhan y, V. S. & Rus am, Z. Li e cance classi ica ion using andom o es and ex eme g adien boos ing (XGBoos ) wi h
gene ic algo i hm as ea u e selec ion. In 2021 In e na ional Con e ence on Decision Aid Sciences and Applica ion (DASA), Sakhee ,
Bah ain. 716–719 h ps:// doi. o g/ 10. 1109/ DASA5 3625. 2021. 96823 11 (IEEE, 2021).
Acknowledgemen s
The esea ch was co unded by he Eu opean Union wi hin he REFRESH p ojec -Resea ch Excellence o Region
Sus ainabili y and High- ech Indus ies ID No. CZ.10.03.01/00/22_003/0000048 o he Eu opean Jus T ansi ion
Fund.
Au ho con ibu ions
Concep ualiza ion, G.A. and N.M.; me hodology, J.F., A.S., and W.K.; alida ion, J.F., N.M., and W.K.; o mal
analysis, G.A., N.M. and A.S.; esou ces, G.A. and N.M.; da a cu a ion, A.S. and J.F.; w i ing o iginal d a
p epa a ion, G.A.; w i ing e iew and edi ing, G.A., N.M. and J.F.; isualiza ion, A.S. and W.K.; supe ision,
N.M.; p ojec adminis a ion, J.F., unding acquisi ion, J.F. All au ho s ha e ead and ag eed o he published
e sion o he manusc ip .
Compe ing in e es s
The au ho s decla e no compe ing in e es s.
Addi ional in o ma ion
Co espondence and eques s o ma e ials should be add essed o J.F.
Rep in s and pe missions in o ma ion is a ailable a www.na u e.com/ ep in s.
Publishe ’s no e Sp inge Na u e emains neu al wi h ega d o ju isdic ional claims in published maps and
ins i u ional a ilia ions.
Open Access This a icle is licensed unde a C ea i e Commons A ibu ion 4.0 In e na ional
License, which pe mi s use, sha ing, adap a ion, dis ibu ion and ep oduc ion in any medium o
o ma , as long as you gi e app op ia e c edi o he o iginal au ho (s) and he sou ce, p o ide a link o he
C ea i e Commons licence, and indica e i changes we e made. The images o o he hi d pa y ma e ial in his
a icle a e included in he a icle’s C ea i e Commons licence, unless indica ed o he wise in a c edi line o he
ma e ial. I ma e ial is no included in he a icle’s C ea i e Commons licence and you in ended use is no
pe mi ed by s a u o y egula ion o exceeds he pe mi ed use, you will need o ob ain pe mission di ec ly om
he copy igh holde . To iew a copy o his licence, isi h p:// c ea i eco mmons. o g/ licen ses/ by/4. 0/.
© The Au ho (s) 2024