Con en s lis s a ailable a ScienceDi ec
Science o Remo e Sensing
jou nal homepage: www.sciencedi ec .com/jou nal/science-o - emo e-sensing
Full leng h a icle
Compa a i e analysis o UAV-based LiDAR and pho og amme ic sys ems o
he de ec ion o e ain anomalies in a his o ical con lic landscape
Ma cel S o cha,b,∗, Benjamin Kisliukc, Thomas Ja me a, Bjö n Waskea, No be de Langeb
aOsnab ück Uni e si y, Ins . o Compu e Science, Remo e Sensing and Digi al Image Analysis, Wachsbleiche 27, 49090 Osnab ück, Ge many
bOsnab ück Uni e si y, Ins . o Compu e Science, En i on. In o ma ics and Municipal Planning, Wachsbleiche 27, 49090 Osnab ück, Ge many
cGe man Resea ch Cen e o A i icial In elligence, Plan-based Robo Con ol G oup, Hambu ge S . 24, 49084 Osnab ück, Ge many
ARTICLE INFO
Keywo ds:
His o ical e ain anomalies
UAV-based emo e sensing
LiDAR s. pho og amme y
Vege a ion impac
ABSTRACT
The documen a ion o his o ical a e ac s and cul u al he i age using high- esolu ion da a ob ained om
unmanned ae ial ehicles (UAVs) is o pa amoun impo ance in he p ese a ion o his o ical knowledge.
This s udy compa es h ee UAV-based sys ems o he de ec ion o his o ically ele an e ain anomalies
in a con lic landscape. Two lase scanne s, a high-end (RIEGL miniVUX-1UAV) and a lowe p iced model
(DJI Zenmuse L1), along wi h a cos -e ec i e op ical came a sys em (pho og amme y using S uc u e om
Mo ion, S M) we e employed in wo s udy si es wi h di e en densi ies o ege a ion. In he s udy a ea wi h
deciduous ees and li le low ege a ion, he DJI Zenmuse L1 sys em pe o ms compa ably o he RIEGL
miniVUX-1UAV, wi h highe comple eness bu lowe co ec ness. The S M me hod demons a ed in e io
pe o mance wi h espec o co ec ness and he F1-sco e, ye achie ed compa able o highe comple eness
alues compa ed o he lase scanne s (maximum 1.0, median 0.84). In he s udy a ea cha ac e ized by dense
nea -g ound ege a ion, he de ec ion esul s a e less op imal. Howe e , he RIEGL miniVUX-1UAV sys em s ill
demons a es supe io esul s in anomaly de ec ion (F1-sco e maximum 0.61, median 0.53) compa ed o he
o he sys ems. The DJI Zenmuse L1 da a showed lowe pe o mance (F1-sco e maximum 0.56, median 0.46).
Bo h lase scanne s exhibi ed enhanced esul s in compa ison o he S M app oach, wi h a maximum F1-sco e
o 0.12. Hence, he S M me hod is iable unde speci ic condi ions, such as de olia ed ees wi hou dense low
ege a ion. The e o e, lowe -cos sys ems can o e cos -e ec i e al e na i es o he high-end LiDAR sys em in
sui able en i onmen s. Howe e , limi a ions pe sis in densely ege a ed a eas.
1. In oduc ion
The iden i ica ion o his o ical aces and emains in landscapes
whe e his o ical con lic s ha e occu ed is o signi ican alue. I is im-
pe a i e o documen his o ical emains in o de o p ese e knowledge
be o e i is los , which is pa icula ly impo an gi en ha many such
landscapes a e subjec o signi ican ans o ma ion p ocesses. These
can be o a physical na u e, o example, h ough p ocesses caused o
accele a ed by clima e change. Simila ly, land de elopmen – encom-
passing cons uc ion, land modi ica ion and land use changes – poses
a signi ican h ea o his o ical emnan s and can, in u n, exace ba e
he impac s o clima e c isis h ough e.g. ege a ion loss, al e ed wa e
balance and inc eased e osion. In addi ion, ans o ma ion p ocesses
can be o a discu si e na u e, which can esul in he eme gence o
ce ain na a i es abou he con lic s o e ime. The documen a ion o
his o ical aces enables he ques ioning and, i necessa y, e u a ion
o alse na a i es abou he con lic s, engage in a discussion on he
∗Co esponding au ho a : Osnab ück Uni e si y, Ins . o Compu e Science, Remo e Sensing and Digi al Image Analysis, Wachsbleiche 27, 49090 Osnab ück,
Ge many.
E-mail add ess: [email p o ec ed] (M. S o ch).
cul u e o ememb ance, and e en ually acili a e he de elopmen o
cul u al he i age. His o ical con lic s in Fi s and Second Wo ld Wa s,
in pa icula , ha e he e o e been he subjec o in es iga ion wi hin
con lic landscape esea ch (Adam e al.,2022;S ichelbau and Cowley,
2016;Van De Sch iek,2022). In his con ex , he e m his o ical e ain
anomalies is used o desc ibe s uc u es in he e ain su ace ha
ypically exhibi an unna u al and egula o m (e.g., ec angula o
ound shaped) and a e mos likely he esul o a his o ical con lic .
A subs an ial body o esea ch demons a es ha emo e sensing
echnology can be employed e ec i ely in he con ex o de ec ing
his o ical aces and emains and in es iga ing con lic ed landscapes. I
is a mo e expedien and he e o e mo e e icien me hod han g ound
inspec ion alone. The majo i y o analyses ha e been conduc ed using
ai bo ne (i.e. om an ai c a ) and spacebo ne sys ems (Ba helme
e al.,2024;Duncan e al.,2023;Luo e al.,2019;Ve hoe en,2017;
h ps://doi.o g/10.1016/j.s s.2024.100191
Recei ed 24 Oc obe 2024; Recei ed in e ised o m 18 Decembe 2024; Accep ed 27 Decembe 2024
Science o Remo e Sensing 11 (2025) 100191
A ailable online 4 Janua y 2025
2666-0172/© 2025 The Au ho s. Published by Else ie B.V. 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/ ).
M. S o ch e al.
Weldegeb iel e al.,2024). Recen ly, he u iliza ion o d one-based
(UAV, unmanned ae ial ehicle) emo e sensing da a o de ec ing
his o ical emains has become inc easingly impo an o e he las ew
yea s (Agudo e al.,2018;Calleja e al.,2018;Campana,2017;Opi z
and He mann,2018). While he a ea co e age achie able wi h a d one
is conside ably less han ha o an ai c a , he signi ican ly highe
esolu ion enables a mo e comp ehensi e eco ding o he g ound
si ua ion, he eby acili a ing a mo e de ailed analysis. Fu he mo e,
d one ligh s can be lexibly adap ed o sui he speci ic a ea unde
in es iga ion and he esea ch ques ion a hand wi h ega d o ligh
planning and da a acquisi ion pa ame e s (Opi z and He mann,2018;
Risbøl and Gus a sen,2018;S o ch e al.,2021).
A a ie y o senso echnologies a e employed in he ield o UAV-
based emo e sensing. Howe e , in his con ex , a comp ehensi e and
sys ema ic compa ison o hese senso s s ill needs o be ca ied ou ,
which is he ocus o his pape . A signi ican dis inc ion can be made
be ween he u iliza ion o passi e op ical- e lec i e came a sys ems on
he one hand and lase scanning (LiDAR, ligh de ec ion and anging)
sys ems on he o he hand. Lase beams a e emi ed and hei e u n
signals a e analyzed, he eby c ea ing a h ee-dimensional poin cloud
o he a ea unde in es iga ion (Luo e al.,2019). LiDAR is capable o
pene a ing ege a ion o a ce ain ex en . This ep esen s a signi ican
ad an age o LiDAR o e o he echnologies in he de ec ion o his o i-
cal emains. The de e io a ion o such emains is pa icula ly p e alen
in open, unp o ec ed e ain. The p ese a ion condi ions a e also poo
in ag icul u al a eas, whe e illage and o he o ms o in e en ion
equen ly esul in signi ican dis u bance. In con as , his o ical aces
and emnan s a e usually well p ese ed unde o es co e , whe e hey
a e be e p o ec ed om ex e nal in luences and e osion p ocesses.
Consequen ly, lase scanning ep esen s an app op ia e me hodology
o he scanning and iden i ica ion o his o ical aces and emains due
o i s capabili y o e ec i ely pene a e ege a ion laye s (C ow e al.,
2007;Ronchi e al.,2020) and UAV-LiDAR has been equen ly used
o his pu pose (Adam e al.,2022;S o ch e al.,2021;Khan e al.,
2017;Masini e al.,2022;S o ch e al.,2023;Zhou e al.,2020).
In many ins ances, including hose p e iously ou lined, d one-based
lase scanning sys ems o he highes quali y a e u ilized o hese
in es iga ions, such as hose belonging o he RIEGL (mini)VUX-SYS
se ies. Thei sys em accu acies and p ecision a e ypically in he ange
o millime e s o 1–1.5cm (Zhou e al.,2020;D eie e al.,2021;
Kükenb ink e al.,2022). Ne e heless, he high cos s associa ed wi h
sys em acquisi ion can o en ac as a ba ie o he implemen a ion
o such sys ems, i espec i e o he speci ic issue o pu pose in ques-
ion (Zhou e al.,2020;Hu e al.,2020). Consequen ly, esea ch is
being conduc ed o de e mine he easibili y o u ilizing al e na i e,
cos -e ec i e sys ems. Fo ins ance, Ronchi e al. (2020) examined he
use o LiDAR and mul ispec al da a o he de ec ion o his o ical land-
scape anomalies. Vilbig e al. (2020) compa ed ai plane-based LiDAR
and UAV pho og amme y o map a his o ic si e. Zhou e al. (2020)
compa ed he RIEGL VUX-1UAV and he Velodyne VLP-16 wi h espec
o hei capaci y o pene a e ege a ion and esul ing g ound poin
densi ies. Hu e al. (2020) de eloped a low-cos UAV-LiDAR sys em
based on he DJI Li ox MID40 o use in o es a eas. I s capabili ies
we e hen compa ed wi h hose o highe -quali y and mo e expensi e
scanne s, including he RIEGL VUX-1UAV, RIEGL miniVUX-1UAV, and
HESAI Panda 40. Salach e al. (2018) and Š one e al. (2023) each
e alua ed a UAV-based LiDAR solu ion (YellowScan Su eyo and DJI
Zenmuse L1, espec i ely) o DTM gene a ion in compa ison o an
al e na i e ha is e en mo e cos -e ec i e: Pho og amme ic poin
clouds we e de i ed om RGB images acqui ed by a digi al came a ha
was moun ed on a d one. Kükenb ink e al. (2022) also compa ed wo
UAV-based LiDAR sys ems (RIEGL) wi h e es ial pho og amme y in
o es en i onmen s.
The s udies e e enced demons a e ha low-cos a ian s can se e
as a easonable al e na i e in ce ain ci cums ances. In some ins ances,
hey we e able o achie e compa able ou comes wi h ega d o accu-
acy o he a ge alue o be de e mined. Ne e heless, some a e no
explici ly conce ned wi h scanning his o ically ele an emains (Kükenb ink
e al.,2022;Hu e al.,2020;Salach e al.,2018;Š one e al.,2023).
O he s who do, howe e , limi hemsel es o a p ima ily manual and
isual in e p e a ion and e alua ion o he da a in ela ion o he
his o ically ele an objec s (Ronchi e al.,2020;Zhou e al.,2020;
Vilbig e al.,2020). The implemen a ion o au oma ed de ec ion and
classi ica ion p ocedu es would ye be essen ial o ca ying ou an
objec i e and compa able analysis. Hence, one aim o his s udy is
o asce ain he ex en o which he classi ica ion ou comes using a
lowe -p iced lase scanning sys em a e compa able o hose o he
highe -p iced sys em, and hus po en ially o e an al e na i e solu ion.
Fu he mo e, an addi ional d one is employed o ga he RGB image
da a om he designa ed su ey a eas, ep esen ing an e en mo e cos -
e ec i e al e na i e o lase scanning. A pho og amme ic me hod is
used o ob ain 3D poin clouds om he image da a, and he classi i-
ca ion esul s o he de ec ion o e ain anomalies a e compa ed wi h
hose o he wo lase scanne s. The second objec i e o his pape is
he e o e o asce ain whe he pho og amme ically gene a ed 3D da a
could ep esen a iable al e na i e o he mo e expensi e lase scanne
sys ems o his esea ch domain.
Ne e heless, a numbe o s udies ha e demons a ed ha he qual-
i y o pho og amme ically gene a ed 3D da a can a y signi ican ly
in compa ison o lase scanning da a, pa icula ly in a eas wi h high
ege a ion densi y (Vilbig e al.,2020;Salach e al.,2018;Š one
e al.,2023). Fo his eason, wo s udy si es we e selec ed o anal-
ysis, cha ac e ized by di e en densi ies o ege a ion. In ligh o he
a o emen ioned conside a ions, he hi d objec i e is o assess e ain
anomaly de ec ion ac oss a ying ege a ion densi ies, using he h ee
UAV-based sys ems, while e alua ing hei sui abili y unde ealis ic,
sub-op imal condi ions.
The e o e, o summa ize, he o e all objec i e o his pape is
o conduc a compa a i e analysis o di e en d one-based sys ems
in he con ex o de ec ing his o ical e ain anomalies. D one-based
lase scanning da a and pho og amme ically gene a ed 3D da a we e
collec ed in a his o ical con lic landscape wi h a ying ege a ion
densi ies, using bo h a high-cos and a low-cos lase scanning sys em,
as well as an RGB d one.
In o de o ensu e he objec i i y o he a o emen ioned com-
pa isons, i is essen ial ha he e alua ion is conduc ed au oma i-
cally h ough he implemen a ion o a classi ica ion p ocedu e, as
opposed o a manual app oach. The use o empla e-ma ching ap-
p oaches o objec -based image analysis o en p esen s he disad an-
age o equi ing he inco po a ion o p io knowledge o se e al
pa ame e s (Lambe s e al.,2019). In con as , he e ec i eness and
possible applica ions o machine lea ning (ML) in de ec ing his o ical
aces and emnan s ha e been demons a ed in a ew s udies (S o ch
e al.,2023;Guyo e al.,2018). Recen ly, he e has u he mo e been
a no able inc ease in he u iliza ion o deep lea ning in he ield o
emo e sensing (Osco e al.,2021), wi h applica ions ex ending o
he de ec ion o his o ical emains (Duncan e al.,2023;T ie e al.,
2021;T o e e al.,2022). Howe e , i s applica ion o d one da a
can p esen challenges in his ega d, as a subs an ial and su icien
amoun o aining da a is ypically equi ed. This can be p oblema ic in
ins ances whe e he numbe o his o ical emains in small su ey a eas
is limi ed. Consequen ly, we ha e chosen o u ilize ML me hodologies
ha a e no associa ed wi h he domain o deep lea ning in he p esen
in es iga ion. Finally, he me ics o comple eness and co ec ness, as
well as hei ha monic mean (F1-sco e), a e u ilized o alida e he
accu acies o he di e en esul s. Comple eness is a measu e o he
numbe o e e ence ins ances ha ha e been co ec ly iden i ied in he
classi ica ion p ocess. I p o ides insigh in o he ex en o which he
g ound u h has been cap u ed. Co ec ness measu es how many o he
objec s iden i ied in he classi ica ion a e ac ually co ec . I he e o e
e lec s he accu acy o he ou pu .
Science o Remo e Sensing 11 (2025) 100191
2
M. S o ch e al.
Fig. 1. Flow cha depic ing he comple e wo k low o de ec ing e ain anomalies.
In o de o a oid he po en ial limi a ions o elying on a single
classi ica ion me hod, his s udy employs wo dis inc ML echniques
o assess i hey pe o m in a simila way: A one-class suppo ec o
machine (Schölkop e al.,2001), which can be applied in an unsupe -
ised manne , and he MaxEn algo i hm (Phillips e al.,2006;Phillips
and Dudík,2008), which is a supe ised me hod. Bo h me hods a e
based on a bina y app oach, in ha he e is only one class o iden i y.
Consequen ly, hey a e e e ed o as one-class classi ie s (OCC) and
ep esen wo widely used and well-es ablished OCC in emo e sensing,
as e idenced by hei applica ion in nume ous s udies (Mack e al.,
2014;Mack and Waske,2017;Rapinel and Hube -Moy,2021;Shi
e al.,2021;Yang e al.,2021;Zu e al.,2024).
2. Ma e ial and me hods
2.1. O e iew o me hodological amewo k
This sec ion p o ides a s uc u ed o e iew o he me hodological
wo k low employed in his s udy. Fig. 1depic s a schema ic ep esen-
a ion o he comple e wo k low in he o m o a low cha .
(1) D one da a is acqui ed in a his o ical con lic landscape (Sec-
ion 2.2) using h ee di e en senso sys ems (Sec ion 2.3). Poin
cloud da a is gene a ed om each o he h ee di e en da a sou ces
(Sec ion 2.4). (2) G ound poin s a e sepa a ed om non-g ound poin s
(Sec ion 2.5) and a Di e en ial Mo phological P o ile (DMP) is used
on he g ound da a o ex ac he e ain ea u es (Sec ion 2.6.1). (3)
Subsequen ly, wo dis inc classi ica ion echniques o anomaly de ec-
ion a e employed in he analysis o each da a se (One-Class Suppo
Vec o Machine, OCSVM, Sec ion 2.6.2, and he MaxEn classi ie , Sec-
ion 2.6.3). (4) In o de o compa e and e alua e he pe o mance o he
di e en acquisi ion sys ems o he de ec ion o e ain anomalies, a
inal alida ion o he classi ica ion esul s is ca ied ou (Sec ion 2.6.4).
2.2. S udy si e
The a ea unde in es iga ion is si ua ed wi hin he Ei el egion in
he wes o Ge many close o he Belgium-Ge man bo de , speci ically
wi hin he alley o Kall (Hü genwald municipali y). In No embe
1944, he US A my and he Weh mach engaged in in ense comba
in his egion. Due o he challenging opog aphy and he equen
e ea s, he Ame ican soldie s cons uc ed a mul i ude o de ensi e
s uc u es, commonly e e ed o as ‘‘ oxholes’’, o sa egua d hemsel es
om enemy assaul s and i le i e (Mille ,2003). The emnan s o hese
ci cula shel e s can s ill be obse ed oday as opog aphical anomalies
wi hin he landscape o he alley o Kall. Mo e ecen p ocesses,
including de o es a ion, he consequences o clima e change, and he
impac o eenac men , a e con ibu ing o he de e io a ion o hese
emains, as e idenced by Adam e al. (2022).
The s udy a ea is si ua ed o he sou heas o Aachen in close
p oximi y o he illage Komme scheid (see Fig. 2). The alley o
Kall is p edominan ly o ien ed in a no h-sou h di ec ion. I is pa ially
o es ed and exhibi s al e na ing zones o high and low ege a ion. In a
manne simila o ha desc ibed by S o ch e al. (2023), wo sub-a eas
we e iden i ied wi hin he alley, each wi h an a ea o app oxima ely
0.5 hec a es. Since hese wo sub-a eas ha e no been excessi ely
al e ed by o es y ac i i ies o he in luence o eenac o s in ecen
yea s, i can be assumed ha ox holes can s ill be de ec ed. In he
ollowing sec ions, he wo s udy a eas will be e e ed o as s udy
a ea Aand s udy a ea B.
S udy a ea A is si ua ed on he eas e n side o he alley o Kall.
The ee popula ion in his egion o he alley emains almos in ac ,
i is hus cha ac e ized by a dense assemblage o deciduous ees, while
he immedia e g ound co e is la gely absen , wi h he excep ion o
a spa se accumula ion o allen lea es. The second s udy a ea (s udy
a ea B) is loca ed on he wes side o he alley. In con as o s udy
a ea A, his egion is dis inguished by dense low ege a ion, including
bushes and sh ubs, which pe sis yea - ound, along wi h o es co e .
Fo nadi images o he wo su ey a eas, cap u ed ia d one and
supplemen ed by g ound pho og aphs o po en ial oxholes, he eade
is di ec ed o Fig. 2.
2.3. Da a acquisi ion
Da a acquisi ion ook place on Ap il 4, 2023. The deciduous ees
in his egion ha e no ye eached hei ull oliage de elopmen a
his ime o yea . The sky was p edominan ly clea , wi h minimal
cloud co e . Th ee di e en d one-based sys ems we e used o collec
da a: Two lase scanne s and one op ical- e lec i e came a sys em. The
la e is he cheapes sys em used and is o be u ilized o gene a e
3D poin clouds om he images using digi al ae ial pho og amme y
me hodology (S uc u e- om-Mo ion, S M).
The high-p iced lase scanning sys em is he RIEGL miniVUX-1UAV
scanne , moun ed unde a DJI Ma ice 600 ca ie d one. The scanne
ope a es ia a o a ing mi o , he eby gene a ing scan lines ha un
ans e se o he di ec ion o ligh . I is able o cap u e up o i e
echoes pe emi ed pulse a a pulse epe i ion equency (PRF) o
100 kHz. The lowe -p iced lase scanning sys em is he DJI Zenmuse
L1 scanne , moun ed unde a DJI Ma ice 300 ca ie d one. I consis s
o a Li ox LiDAR module based on a isley p ism se up. The L1 can
ecei e up o h ee echoes a a scan a e o 240 kHz. Bo h de ices
exhibi an ellip ical lase oo p in ; howe e , his is mo e han wice as
la ge in he case o he DJI Zenmuse L1 as i is o he RIEGL miniVUX-
1UAV. The ine ial measu emen uni s (IMU) in bo h sys ems sample a
a a e o 200 Hz. Howe e , he accu acy o he Applanix APX-20 IMU
implemen ed in he RIEGL miniVUX-1UAV is 1.6 imes be e han ha
o he DJI Zenmuse L1 in e ms o oll and pi ch and 4.3 imes be e in
e ms o heading (see Table 1,DJI,2024b;RIEGL,2024;Mandlbu ge
e al.,2023).
The op ical- e lec i e sys em is a came a moun ed unde a DJI
Phan om 4. We use he RGB-bands o he P4 mul ispec al sys em o
ou s udy, as i includes RTK unc ionali y (see below), which is c ucial
o ou wo k o ensu e p ecise geo e e encing. The came a akes images
wi h a esolu ion o 2.1 megapixels. The wa eleng hs u ilized o isible
ligh imaging a e 450 nanome e s (nm) ±16 nm o he blue band,
560 nm ±16 nm o he g een band and 650 nm ±16 nm o he ed
band DJI (2024a).
All da a mus be geo e e enced as accu a ely as possible so ha he
collec ed da a can be compa ed. Hence, de e mining he ajec o y o
he d one as accu a ely as possible is a pi o al s ep in he gene a ion
and geo e e encing o UAV poin cloud da a. I is he e o e common o
in eg a e a co ec ion signal ob ained by a ixed base s a ion which is
se up a a known and p ecisely measu ed loca ion (Di e en ial Global
Na iga ion Sa elli e Sys em, D-GNSS). In his way, cen ime e -le el
accu acy can be achie ed (D eie e al.,2021).
When he co ec ion signal is ob ained and in eg a ed in eal ime,
he p ocess is called Real Time Kinema ics (RTK). Since bo h he DJI
Science o Remo e Sensing 11 (2025) 100191
3
M. S o ch e al.
Fig. 2. Loca ion and pho os o he esea ch a ea. Pho os aken by M. Adam and M. S o ch.
Table 1
Senso speci ica ions o he wo UAV-based lase scanne s (DJI,2024b;RIEGL,2024;Mandlbu ge e al.,2023).
RIEGL miniVUX-1UAV wi h DJI Zenmuse L1
IMU Applanix APX-20 UAV
Lase Wa eleng h Nea -In a ed Nea -In a ed
Scan Ra e 100 kHz 240 kHz
Max. Numbe o Echoes 5 3
Lase Foo p in @ 50 m 8 cm ×3 cm 25 cm ×3.5 cm
IMU Sampling Ra e 200 Hz 200 Hz
IMU Accu acy: Roll, Pi ch 0.015◦0.025◦
IMU Accu acy: Heading 0.035◦0.15◦
Ma ice 300 and he DJI Phan om 4 a e RTK-capable, an RTK co ec-
ion was ealized by using he DJI base s a ion (DJI D-RTK 2). The
loca ion o he e e ence poin was p e iously measu ed p ecisely using
a GNSS ecei e (S onex S9III+), which ecei es i s D-GNSS co ec-
ion signal om known e e ence g ound s a ions (sa elli e posi ioning
se ice o he Ge man na ional su ey, SAPOS). Howe e , he DJI
Ma ice 600 is no compa ible wi h he D-RTK 2 s a ion. The e o e, an
RTK co ec ion could no be applied du ing he ligh . In his case, he
S onex S9III+ was used o eco d he aw sa elli e signals while he DJI
Ma ice 600 pe o med he ligh . The co ec ion o he aw ajec o y
was hen subsequen ly applied a e he ligh du ing da a p ocessing
(Sec ion 2.4.1).
The ligh s wi h he di e en d one sys ems we e ca ied ou in bo h
a eas in succession on he same day. We op ed o a lying al i ude o
50 me e s abo e g ound le el (AGL) in o de o main ain a sa e dis ance
om he ees wi hin he s udy a ea and o gua an ee con inuous isual
con ac be ween he d one pilo and he UAV h oughou he en i e
ligh du a ion. The RGB da a eco ding achie es an 80% side and on
o e lap.
2.4. Poin cloud gene a ion
2.4.1. RIEGL miniVUX-1UAV scanne
The p ocessing o he LiDAR da a eco ded by he RIEGL miniVUX-
1UAV sys em commences wi h he co ec ion o he ligh ajec o y
ob ained om he d one. As i was no possible o u ilize RTK co ec ion
wi h he Ma ice 600 in ou se up (see Sec ion 2.3), i s ligh pa h
needs o be ec i ied p io o he gene a ion o poin clouds de i ed
om he scanne da a. I is he e o e necessa y o ob ain he GNSS
co ec ion da a om he GNSS base s a ion. The GNSS obse ables
and he aw IMU (ine ial measu emen uni ) da a a e impo ed in o
he Applanix POSPac so wa e. Subsequen ly, he co ec ed ajec o y
(smoo hed bes es ima e o ajec o y, SBET), is combined wi h he aw
scanne da a in RIEGL’s RiPROCESS p og am, and global egis a ion
is pe o med o gene a e geo e e enced poin clouds. Fu he mo e,
RIEGL’s RiPRECISION plugin in RiPROCESS is u ilized o au oma ically
iden i y ie-planes, he eby enhancing c oss- ligh line egis a ion. A
mo e de ailed explana ion o his wo k low can be ound in a a ie y
Science o Remo e Sensing 11 (2025) 100191
4
M. S o ch e al.
o o he s udies (S o ch e al.,2021;D eie e al.,2021;B ede e al.,
2017; en Ha kel e al.,2019).
2.4.2. DJI Zenmuse L1 scanne
The DJI Zenmuse L1 UAV senso has been in oduced in 2021 as a
ela i ely a o dable UAV-based 3D LiDAR senso . I means o ill he
gap be ween pho og amme ic 3D econs uc ion and ypically mo e
expensi e bu e y p ecise LiDAR sys ems. I can be exclusi ely used
as a payload o he DJI Ma ice 300 ca ie d one and combines an
IMU, RGB pho o senso and a LiDAR module based on he p inciple o
he Li ox Mid-40. I o e s wo scan modes which a e called line scan
mode and non- epe i i e (DJI,2024b) enabled by i s se up comp ising a
limi ed numbe o LiDAR anscei e s and a beam de lec ion u ilizing a
Risley p ism. This esul s in he gene a ion o mul iple pa allel igu e-
o -eigh -like scanning ajec o ies, which a e subsequen ly employed
in a na ow, s e ched, nea ly linea scanning band o o a ed in o a
ound pe al shape ha only epea s a e a ull i e a ion o scans. This
hyb id app oach pe mi s he educ ion o cos s while main aining an
adequa e numbe o measu emen poin s (Mandlbu ge e al.,2023;
B azeal e al.,2021).
In his s udy, he non- epe i i e scan mode was selec ed in o de o
enable he acquisi ion o a g ea e numbe o measu emen poin s a
oblique angles along he ligh axis han would be easible in he linea
mode. This can be pa icula ly ad an ageous in ege a ed a eas. The
esul ing da a can only be p ocessed wi h he p op ie a y DJI so wa e
Te a in o 3D su ace models o he en i onmen . In combina ion wi h
he (RTK co ec ed) GNSS da a o he d one, i is possible o c ea e ull
colo 3D en i onmen models.
2.4.3. Op ical- e lec i e RGB images and s uc u e om mo ion
3D in o ma ion is gene a ed om he RGB images using S uc u e
om Mo ion (S M), which is a pho og amme ic echnique ha econ-
s uc s 3D s uc u es om a se ies o 2D images, in his case RGB
images aken om he Phan om 4 d one. The images mus cap u e
he objec s om di e en angles, in o de o ensu e su icien o e lap
and pe spec i e a ia ions. The e o e, a e y high o e lap bo h in and
ac oss he ligh di ec ion was ensu ed (80% o e lap in he di ec ion o
ligh , on o e lap, as well as 80% o e lap ans e se o he di ec ion
o ligh , side o e lap, see Sec ion 2.3). Dis inc i e poin s ( ea u es)
such as co ne s, edges o ex u es a e ecognized in he images. Algo-
i hms such as SIFT (Scale-In a ian Fea u e T ans o m) ex ac hese
ea u es. Subsequen ly, hey a e iden i ied in he a ious images and
co ela ed wi h each o he . The ela i e posi ion and o ien a ion o he
came as (came a poses) is de e mined on he basis o he co espond-
ing ea u es. Finally, a dense 3D poin cloud is gene a ed using he
es ima ed came a poses and co esponding ea u es. This is done by
iangula ion, in which he 3D coo dina es o he poin s a e calcula ed
by de e mining he in e sec ion poin s o he p ojec ions o hese poin s
in he images (see Jiang e al.,2020 o a comp ehensi e o e iew
o S M echnology). The p ocedu e is implemen ed using he Agiso
Me ashape so wa e.
2.5. G ound il e
When gene a ing digi al e ain models om 3D poin cloud da a,
a c ucial s ep is o de e mine which da a poin s ep esen he e ain.
This is known as (g ound) il e ing and in ol es he poin -by-poin
classi ica ion o he da a poin s in o g ound poin s and non-g ound
poin s. Fo his pu pose, he e is a wide ange o il e algo i hms ha
wo k in di e en ways, e.g. su ace-based, TIN-based, and mo phology-
based, o a combina ion o di e en app oaches. The a ious algo i hms
ha e ad an ages and disad an ages depending on he condi ions and
cha ac e is ics o he espec i e s udy a ea. S udies ha e shown ha
mo phologically based me hods can lead o good esul s in complex
e ain S o ch e al. (2021), Moud `
y e al. (2020), Klápš ě e al. (2020).
We he e o e op o he Simple Mo phological Fil e (SMRF) o g ound
il e ing. This il e ing p ocess is based on mo phological ope a ions
(e osion and dila ion). Fi s , he poin cloud is con e ed in o a as e ,
hen a s uc u e mask wi h inc easing size is used o e ode local min-
ima, ollowed by dila ion o es o e he o iginal shape. By combining
hese wo s eps, he poin s close o he g ound a e sepa a ed om
he highe objec s. A e he mo phological ope a ions, a h eshold is
applied o pe o m he inal classi ica ion o he g ound and non-g ound
poin s (Pingel e al.,2013).
The p ocedu e is based on he ollowing pa ame e iza ion: I is
ecommended ha he maximum adius should be se o 10 mo
g ea e , and ha he slope ole ance pa ame e should no exceed a
alue o 10%. The ele a ion h eshold is hen calcula ed by mul iplying
he slope ole ance by he window adius and he cell size (Pingel
e al.,2013). In ou case, we se he cell size o 20 cm. This leads o
a maximum ele a ion h eshold o he ini ial opened su ace o one
me e (0.1 × 0.2 × 10 m∕0.2 m). This co esponds o he app oxima e e -
ical ex en o he e ain anomalies o be expec ed in he s udy a eas,
which consequen ly emain p ese ed as pa o he minimal su ace o
he SMRF algo i hm. Subsequen ly, he inal h eshold o classi y each
LiDAR poin in o g ound o non-g ound is se o 10 cm. This alls wi hin
he usual alue ange o his pa ame e ( ypically 0.05 m–0.2 m) in
o he s udies (Moud `
y e al.,2020;Klápš ě e al.,2020;Š one e al.,
2022). The pa ame e iza ion is applied o he poin clouds o all h ee
d one sys ems in bo h s udy a eas in an iden ical manne , so ha all
da a se s a e ea ed equally o compa abili y.
2.6. Me hodology
2.6.1. Di e en ial mo phological p o ile
Since a e ain model is simply a g id o ele a ion alues, spa ial
in o ma ion seems pa icula ly ele an o da a analysis. Fil e ing ech-
niques, such as mo phological il e s, can simul aneously supp ess he
unimpo an de ails and p ese e he spa ial cha ac e is ics o he o he
egions, such as ce ain shapes in he e ain model. In his s udy, Di -
e en ial Mo phological P o iles (DMP) a e used. The DMP was o iginally
de eloped by Pesa esi and Benedik sson (2001) o segmen ing and
classi ying high- esolu ion sa elli e image y. A mo phological p o ile
(MP) is c ea ed by epea edly applying e osions and dila ions o an
image, mo e speci ically opening (e osion ollowed by dila ion) and
closing (dila ion ollowed by e osion) ope a ions. A s uc u ing mask
(mo ing window) de ines he local neighbo hood a ound each pixel
o hese ope a ions. I s shape depends on he ype o ea u es being
de ec ed. I can be e.g. squa e, ound o ec angula shaped, and also
he size (in pixels) o he s uc u ing elemen mus be speci ied. Then i s
size is inc eased in each i e a ion in o de o c ea e he MP. Finally, he
DMP ex ends he MP by calcula ing he di e ences be ween consecu i e
MP images. In his way, he p o iles cap u e he changes in he image
s uc u e as he size o he s uc u ing elemen a ies. When applied o
an o iginally single-channel g ayscale image, in his case an ele a ion
g id, he DMP p oduces a mul idimensional esul .
DMPs a e widely used in emo e sensing o classi ying sa elli e
images (Huang e al.,2016;Kemmouche e al.,2021), bu also o
objec and anomaly de ec ion in LiDAR da a (S o ch e al.,2023;
Mongus e al.,2014). In his pape , a ci cula shaped s uc u ing mask
is employed because he e ain anomalies o be sea ched o ( oxholes)
a e also ound shaped. The ini ial diame e o he elemen is se a
one me e and inc eased in i e s eps up o a size o wo me e s.
This esul s in a DMP wi h en bands and allows e ain anomalies o
di e en sizes o be highligh ed. As a esul , each pixel has en alues
( i e imes closing, i e imes opening). Subsequen ly, he ou pu o he
DMP is used as inpu ea u es o he one-class suppo ec o machine
(Sec ion 2.6.2) and he MaxEn classi ie (Sec ion 2.6.3).
Science o Remo e Sensing 11 (2025) 100191
5

M. S o ch e al.
2.6.2. One-Class Suppo Vec o Machine
AOne-Class Suppo Vec o Machine (OCSVM) is a special ype o
suppo ec o machine ha is used o de ec ing anomalies, i.e. iden-
i ying anomalous da a poin s ha de ia e om a single class. I was
o iginally p oposed by Schölkop e al. (2001).
As wi h o he SVMs, a ke nel unc ion is used wi h he OCSVM o
ans o m he da a in o a highe -dimensional space. Commonly used
ke nel unc ions a e he linea ke nel, he polynomial ke nel and he
adial basis unc ion (RBF) ke nel. The OCSVM is hen only ained wi h
he da a poin s o one class. These pixels ep esen he no mal, ypical
s a e – in his case, e ain wi hou e ain anomalies. The ma hema ical
basis o he OCSVM is o ind a decision equa ion ha sepa a es
he aining da a poin s in high-dimensional space. This equa ion is
based on a pa ame e 𝜈, which con ols wha p opo ion o he da a
is conside ed an ou lie and how s ic ly he bounda y be ween no mal
and anomalous da a is de ined. A e aining, OCSVM is used on es
da a o classi y he da a poin s. A new da a poin (i.e., a pixel no seen
du ing aining) is classi ied as an anomaly i i lies on he nega i e side
o he sepa a ing hype plane in ea u e space.
A numbe o s udies ha e employed OCSVM o he pu pose o
anomaly de ec ion in emo e sensing da a (Ananias and Neg i,2021;
Chen e al.,2024;Coca and Da cu,2021). A ecen s udy has also
demons a ed he e icacy o OCSVM in he de ec ion o e ain anoma-
lies in LiDAR da a (S o ch e al.,2023). Simila ly, we also use he
OCSVM in an unsupe ised manne . The da a se s a e di ided in o
pa ches, 2∕3 o hem a e used o aining, 1∕3 is used o independen
classi ica ion. Then his p ocess is epea ed un il all po en ial combi-
na ions o pa ches ha e been u ilized once o ain he OCSVM. The
mean signed dis ance o he sepa a ing hype plane, calcula ed o e all
i e a ions, is employed as he basis o he inal classi ica ion decision.
In he e en ha he a o emen ioned dis ance is nega i e, he pixel in
ques ion is classi ied as anomalous.
In o de o asce ain he op imal pa ame e alues o he OCSVM
pa ame e 𝜈, a se ies o i e a ions is conduc ed using alues be ween 0.01
and 0.05 (in 0.5hinc emen ), which is a pa ame e ange also used
in ela ed esea ch (Mack e al.,2014;Li and Xu,2010). The RBF
ke nel 𝐾(𝑥, 𝑥𝑖) = exp(−𝛾||𝑥−𝑥𝑖||2)is employed as he unde lying ke nel
in his pape , as i is he mos commonly applied ke nel in nume ous
o he s udies (Mack e al.,2014;Mack and Waske,2017;Ananias and
Neg i,2021). Howe e , as he inal classi ica ion decision is based on a
la ge numbe o indi idual OCSVM uns, he ke nel pa ame e 𝛾is no
subjec ed o addi ional a ia ion. Ins ead, i is main ained a i s de aul
alue, which is equi alen o he ecip ocal o he numbe o ea u es.
2.6.3. MaxEn classi ie
The MaxEn algo i hm (Maximum En opy) was ini ially de eloped
and employed in he ield o ecology o he pu pose o modeling
he po en ial geog aphical dis ibu ion o species (Phillips e al.,2006;
Phillips and Dudík,2008). I is based on he es ima ion o p obabili y
densi ies. The objec i e is o asce ain he p obabili y o occu ence
o he a ge class, gi en he en i onmen al condi ions, i.e. 𝑃(𝑦|𝑥𝑖)
which deno es he condi ional p obabili y o he p esence o he a ge
class 𝑦gi en he independen inpu a iables 𝑥𝑖. Using Bayes’ ule,
his can be ans o med in o [𝑃(𝑥𝑖|𝑦)𝑃(𝑦)]∕𝑃(𝑥𝑖). Since he a p io i
p obabili y o he a ge class, i.e. 𝑃(𝑦)is no known, MaxEn es ima es
he densi y a io 𝑃(𝑥𝑖|𝑦)∕𝑃(𝑥𝑖). In o de o es ima e he condi ional
p obabili y densi y 𝑃(𝑥𝑖|𝑦), in o ma ion on he inpu a iables 𝑥𝑖a
sample loca ions whe e he a ge class 𝑦was obse ed a e equi ed.
In con as , he uncondi ional p obabili y 𝑃(𝑥𝑖)is es ima ed om back-
g ound samples, wi hou in o ma ion on whe he he a ge class is
p esen o no . The inco po a ion o unlabeled da a in o he modeling
p ocess, in conjunc ion wi h he posi i e samples, ca ego izes MaxEn
as a PU-classi ie (Posi i e and Unlabeled) (Mack e al.,2014;Eli h
e al.,2011).
Simila o he OCSVM, he en bands o he DMP a e employed as
inpu da a. He e, hey ep esen he independen inpu a iables 𝑥𝑖.
Fu he mo e, selec ed loca ions (pixels) a e inco po a ed in o he model
a which he a ge class, namely e ain anomalies indica ing oxholes,
whe e obse ed (see Sec ion 2.6.4 on he acquisi ion o e e ence da a).
To a oid eliance on he selec ion o dis ibu ion o he aining da a,
his classi ie is also i e a ed mul iple imes. One- hi d o he e e ence
da a is u ilized as aining samples, while he emaining wo- hi ds a e
employed o alida ion pu poses. This p ocess is epea ed 100 imes,
wi h a andomized new dis ibu ion be ween he aining and alida-
ion samples each ime. Finally, a h eshold is equi ed in o de o
con e he con inuous ou pu o MaxEn in o a bina y classi ica ion.
A alue o 0.5 is ypically selec ed o his pu pose o decide o
each loca ion (i.e., o each pixel) i i is classi ied as an anomaly o
no (Mack e al.,2014;O iz e al.,2013).
2.6.4. Valida ion
The e e ence da a is gene a ed h ough a p ocess o manual in-
e p e a ion and digi iza ion o he da a. The posi ions in he da a
a e independen ly anno a ed by h ee expe s in o de o iden i y any
e ain anomalies ha may indica e he p esence o oxholes. Subse-
quen ly, he deg ee o o e lap be ween he e e ence and classi ica ion
da ase s is asce ained. A ma ch be ween he e e ence and he clas-
si ica ion esul is indica i e o a co ec ly ecognized ins ance (i.e., a
e ain anomaly) which is e e ed o as a ue posi i e (TP). In con as ,
an objec om he classi ica ion da ase ha does no ha e a spa ial
ma ch in he e e ence da ase is classi ied as a alse posi i e (FP). An
un ecognized objec om he e e ence da a se is classi ied as a alse
nega i e (FN). F om his, comple eness (o ecall, Eq. (1)), co ec ness
(o p ecision, Eq. (2)) and he F1-sco e (Eq. (3)) a e de e mined, each
o which has a alue ange om 0 o 1 (see e.g. Zou e al.,2016).
Comple eness (Recall) =TP
TP +FN (1)
Co ec ness (P ecision) =TP
TP +FP (2)
F1-Sco e = 2⋅
Comple eness ⋅Co ec ness
Comple eness +Co ec ness (3)
In o de o asce ain whe he he esul s, e.g. wi h ega d o he use
o he di e en senso s, di e signi ican ly om each o he o whe he
hese di e ences a e oo small and he e o e andom, a s a is ical es
o he signi icance o he di e ences on independen samples is ca ied
ou . In o de o apply a - es , o ins ance, i is necessa y o ensu e
ha he da a in ques ion ollows a no mal dis ibu ion. In he e en
ha his is no he case, a Mann–Whi ney U es may be conduc ed.
As p e iously s a ed, he selec ion o aining and es da a o each
un is andomized. Consequen ly, he classi ica ion ou comes can be
conside ed o be he esul s o independen andom samples, and he
a o emen ioned es p ocedu es a e he e o e app op ia e o use.
3. Resul s
3.1. Poin cloud compa ison
Table 2summa izes he poin densi y esul ing om he di e en
da a acquisi ion me hods in he wo s udy a eas. The lowes poin
densi y is achie ed by he RIEGL miniVUX-1UAV scanne . I is h ee
o ou imes highe o he da a om he DJI Zenmuse L1 sys em. The
poin densi y ob ained h ough he S M me hod is si ua ed be ween ha
o he da a om he wo lase scanne s. This also applies o he il e ed
poin clouds, i.e. poin s ha ep esen he e ain (see he wo bo om
ows in Table 2).
Fig. 3depic s app oxima ely 60 m × 5 mexempla y sec ions o he
il e ed poin clouds om s udy a eas A and B, de i ed om da a
ob ained by he di e en sys ems. S udy a ea A (le ) is ypi ied by
a deciduous o es wi hou nea -g ound ege a ion. E en hough he
S M me hod (a3) is unable o de i e any meaning ul in o ma ion abou
he ee ege a ion om he RGB images, all h ee sys ems appa en ly
Science o Remo e Sensing 11 (2025) 100191
6
M. S o ch e al.
Table 2
Poin densi ies o he poin clouds using di e en acquisi ion echniques.
RIEGL miniVUX-1UAV DJI Zenmuse L1 RGB images and S M
A ea A A ea B A ea A A ea B A ea A A ea B
Poin Densi y 245 p s/m2192 p s/m2791 p s/m2779 p s/m2309 p s/m2453 p s/m2
A g. Ho iz. Poin Spacing 6cm 7cm 4cm 4cm 6cm 5cm
G ound Poin Densi y 105 p s/m278 p s/m2375 p s/m2313 p s/m2201 p s/m2150 p s/m2
A g. G ound Poin Spacing 10 cm 11 cm 5cm 6cm 7cm 8cm
Fig. 3. 60 m × 5 mexempla y sec ions o he il e ed poin clouds om s udy a eas A and B, de i ed om he RIEGL miniVUX-1UAV scanne (a1, b1), DJI Zenmuse L1 scanne
(a2, b2), RGB images and S uc u e om Mo ion (a3, b3).
demons a ed a compa able capabili y o eco d he ea h’s su ace
opog aphy (g ound poin s a e indica ed by a b own colo ). In s udy
a ea B ( igh ), which is dis inguished by ma kedly mo e p onounced
g ound-le el ege a ion in compa ison o s udy a ea A, i is e iden ha
he e a e di e ences in he pe o mance o he sys ems. On a e age, he
poin densi y o he g ound poin s o he DJI sys em is mo e han h ee
imes ha o he RIEGL sys em (see Table 2). Ne e heless, in a eas
wi h dense ege a ion, pa icula ly bushes and sh ubbe y (le hal o
b1–b3 in Fig. 3), an examina ion o he g ound poin da a om he
DJI Zenmuse L1 (b2) e eals he p esence o gaps. The co e age o he
e ain om he S M da a se (b3) appea s o be e en mo e incomple e
in hese a eas. The RIEGL miniVUX-1UAV (b1), howe e , cap u es he
e ain su ace in hese egions wi hou p oducing da a oids.
3.2. Te ain anomaly de ec ion
Fig. 4depic s an example o he esul s o he anomaly de ec ion
implemen a ion wi hin he wo s udy a eas, using he da a om he
RIEGL miniVUX-1UAV and he OCSVM. The e e ence da a (see Sec-
ion 2.6.4) is delinea ed in ed, he classi ica ion ou pu shows in black
colo whe e an anomaly has been iden i ied.
Figs. 5and 6depic he quan i a i e alida ion esul s g aphically
in he o m o box plo s. Comple eness, co ec ness and he F1-sco e a e
compa ed in e ms o he da a sou ce (RIEGL miniVUX-1UAV shown in
o ange, DJI Zenmuse L1 shown in blue, op ical images and S M shown
in g een) and he anomaly de ec ion algo i hm used (OCSVM s. Max-
En ). Tables 3and 4demons a e whe he he obse ed di e ences a e
s a is ically signi ican . As he Shapi o–Wilk es o no mal dis ibu ion
indica ed ha his was no consis en ly p esen ac oss all da a se s, he
Mann–Whi ney U es was ca ied ou uni o mly o all da a. Finally,
Table 5summa izes he median F1-sco es.
In s udy a ea A (Fig. 5), i is no iceable wi h ega d o he com-
ple eness ha he medians o he comple eness o he classi ica ions
all lie be ween 0.68 and 0.84. In gene al, he esul s o he MaxEn
me hod a e sligh ly be e han hose o he OCSVM. Wi h ega d o he
da a sou ces, howe e , he e a e di e ences in he esul s in ela ion
o he wo classi ica ion me hods used. While he di e ence be ween
he wo me hods is a he small o he da a om he RIEGL sys em
(di e ence in he median o 0.04), i is 0.1 o he DJI sys em and
ises o 0.16 o he S M da a. Fo he la e wo, he di e ences a e
s a is ically signi ican . This also applies when compa ing MaxEn om
he S M and DJI da a o he MaxEn classi ica ion using he RIEGL da a
(see Table 3). In con as , when using OCSVM, he comple eness o he
classi ica ion esul s does no di e signi ican ly.
Simila o he comple eness, i can be obse ed ha he MaxEn
me hod achie es sligh ly be e co ec ness alues han he OCSVM,
a leas o he da a om he wo lase scanne s used, e en i he
di e ence in he da a o he DJI sys em is no signi ican . The e e se
is ue o da a gene a ed wi h S M. He e, OCSVM achie es highe
co ec ness alues han he MaxEn classi ie . O e all, he co ec ness
achie ed by he MaxEn me hod on he RIEGL da a is he highes
Science o Remo e Sensing 11 (2025) 100191
7
M. S o ch e al.
Table 3
Signi icance o he di e ences in he alida ion me ics (s udy a ea A). RIEGL = RIEGL miniVUX-1UAV, DJI = DJI Zenmuse L1, S M = S uc u e om Mo ion based on RGB
images. Fo he en ies ma ked wi h an X, he wo esul s a e s a is ically signi ican ly di e en (𝛼= 0.05).
S udy a ea A Comple eness Co ec ness F1-Sco e
RIEGL - OCSVM
RIEGL - MaxEn
DJI - OCSVM
DJI - MaxEn
S M - OCSVM
S M - MaxEn
RIEGL - OCSVM
RIEGL - MaxEn
DJI - OCSVM
DJI - MaxEn
S M - OCSVM
S M - MaxEn
RIEGL - OCSVM
RIEGL - MaxEn
DJI - OCSVM
DJI - MaxEn
S M - OCSVM
S M - MaxEn
RIEGL - OCSVM X X X X X X X X X
RIEGL - MaxEn X X X X X X X X X
DJI - OCSVM X X X X X
DJI - MaxEn X X X X X
S M - OCSVM X X X
Fig. 4. An exempla y p esen a ion o he classi ica ion esul s o e ain anomaly
de ec ion in he wo s udy a eas. He e: Applica ion o he one-class suppo ec o
machine (OCSVM) based on he RIEGL miniVUX-1UAV da a se s. Top: Hillshade
isualiza ion o he e ain models. Bo om: Classi ica ion esul s.
(median is 0.72). The di e ences o all o he a ian s a e s a is ically
signi ican .
O e all, his pa e n is e lec ed in he F1-sco e. The highes F1-
sco es a e achie ed by he MaxEn me hod on he RIEGL and DJI lase
scanne da a (median 0.71 and 0.69, di e ence no signi ican ). When
using he OCSVM classi ie , his is sligh ly lowe : 0.67 (RIEGL) and 0.62
(DJI). Howe e , his di e ence is s a is ically signi ican . The F1 alues
o he S M da a a e signi ican ly lowe o bo h classi ica ion me hods
in compa ison. Thei maximum does no exceed 0.61, he medians
a e 0.56 and 0.51.
In s udy a ea B (Fig. 6,Table 4), whe e dense unde s o ey ege-
a ion like bushes and sh ubs a e p esen , i can gene ally be s a ed
ha he alida ion me ics o he me hods wi h he lase scanne da a
(RIEGL and DJI) a e highe han he esul s wi h he da a gene a ed
om S M. In e ms o comple eness, i can be seen ha he esul s a e
highe when using he RIEGL da a han when using he DJI da a. Taking
in o accoun he wo di e en classi ica ion me hods, he median o
he comple eness is 0.70 (RIEGL) and 0.60 (DJI) when applying he
OCSVM (di e ence is signi ican ), i is 0.86 (RIEGL) and 0.71 (DJI)
when using MaxEn (no signi ican ). In a di ec compa ison o he
classi ica ion me hods, MaxEn he e o e achie es highe comple eness
alues han OCSVM ( he di e ence is only signi ican o he DJI da a).
A he same ime, he di e en in e qua ile anges (IQR) o he wo
classi ica ion me hods a e appa en . Wi h a alue o 0.43 and 0.29,
he IQR is 4 and 3 imes highe when using MaxEn han he esul s
ob ained wi h OCSVM. When using he da a gene a ed using S M, he
median o comple eness is a ound 0.30 in each case, ega dless o
he classi ica ion me hod. In gene al, i should be no ed ha , due o
he objec -based alida ion and he limi ed numbe o anomalies in
his s udy a ea ( he e e ence da a se o s udy a ea B comp ises 10
anomalies), he comple eness can only a ain alues in 0.1 inc emen s
he e. Consequen ly, i is possible ha he median is equal o ei he
he i s o hi d qua ile. This phenomenon is obse ed in he esul s
de i ed om he RIEGL miniVUX1-UAV da a se (bo h classi ie s) and
he DJI Zenmuse L1, u ilizing he OCSVM.
As wi h comple eness, he da a om he wo lase scanne s also
show signi ican ly highe esul s han he S M in e ms o co ec ness.
A di ec compa ison o he da a om he wo lase scanne s shows ha
he co ec ness alues achie ed wi h he RIEGL da a a e sligh ly highe
han alues achie ed wi h he DJI da a. Howe e , he di e ence is no
signi ican when using OCSVM as classi ie (medians 0.44 s. 0.40), i
is signi ican when using MaxEn (medians 0.18 s. 0.13). The e o e, in
con as o comple eness, he OCSVM co ec ness alues a e gene ally
highe han hose o MaxEn .
The F1-sco e he e o e beha es in a simila way. The highes F1-
sco e is achie ed by he OCSVM classi ie on he RIEGL lase scanne
da a (maximum alue 0.61, median 0.53), he second bes esul is
ob ained when using he OCSVM algo i hm on he DJI da a (maximum
alue 0.56, median 0.46). The di e ence is signi ican , as wi h all o he
a ian s. Once again, bo h classi ica ion me hods achie e by a he
lowes esul s on he da a gene a ed wi h S M (do no exceed a alue
o 0.12).
Table 5p esen s a summa y o he median F1-sco es achie ed o
he wo s udy a eas, h ee acquisi ion sys ems, and wo classi ica ion
me hods.
4. Discussion
The o e all objec i e o his s udy was o compa e di e en UAV
senso sys ems o de ec ing e ain anomalies. Da a se s om wo s udy
si es we e analyzed, using wo di e en me hods (OCSVM, MaxEn ).
One aim was o e alua e he pe o mance o a lowe -cos lase
scanne (in his case DJI Zenmuse L1) in compa ison o a high-p iced
sys em (in his case RIEGL miniVUX-1UAV). S udy a ea A was cha -
ac e ized by deciduous o es s ands wi hou nea -g ound ege a ion.
He e, he DJI sys em achie ed a g ound poin densi y ha is app ox. 3-
4 imes highe han ha o he RIEGL sys em. Despi e he conside able
dispa i y in e ain poin densi y, he classi ica ion maps emain la gely
compa able. The comple eness o he DJI da a se is sligh ly highe
han ha o he RIEGL da a se , whe eas he opposi e is ue wi h
ega d o co ec ness. This gene al end appea s o be independen
o he speci ic classi ie , as his beha io can be obse ed wi h bo h
he OCSVM and he MaxEn classi ie . In s udy a ea B, he ege a ion
consis ed o addi ional dense low ege a ion. He e, he da a om he
Science o Remo e Sensing 11 (2025) 100191
8
M. S o ch e al.
Fig. 5. Valida ion quan i ies in s udy a ea A.
Fig. 6. Valida ion quan i ies in s udy a ea B.
Science o Remo e Sensing 11 (2025) 100191
9