AI-D i en Habi a Change De ec ion in Alpine
P o ec ed A eas: Founda ion Models s. Di ec
Change De ec ion
Ha ald K is en1[000−0003−0395−8428], Daniel Kulme 2[0009−0007−1993−1728], and
Manuela Hi schmugl1,2[0000−0002−5224−0992]
1Uni e si y o G az, 8010 G az, Aus ia
2Joanneum Resea ch, 8010 G az, Aus ia
Abs ac . Rapid clima e change in alpine ecosys ems demands equen
habi a moni o ing, ye adi ional manual mapping app oaches a e p o-
hibi i ely expensi e o he empo al esolu ion needed. This s udy p esen s
a comp ehensi e compa ison o wo undamen al change de ec ion pa adigms
using long- e m alpine habi a da a: (1) pos classi ica ion - indepen-
den habi a maps compa ed empo ally e sus (2) di ec change de-
ec ion - bi- empo al image y p ocessed end- o-end. We sys ema ically
e alua e mode n AI a chi ec u es ac oss bo h pa adigms on a unique 20-
yea alpine habi a da ase om Gesäuse Na ional Pa k, Aus ia. Pos -
classi ica ion app oaches compa e geospa ial ounda ion models P i h i-
EO-2.0 and Clay 1.0 agains adi ional con olu ional neu al ne wo ks
(CNNs) based on U-Ne a chi ec u e. Di ec change de ec ion e alua es
specialized ans o me s ChangeViT agains U-Ne baselines. Using e y
high esolu ion mul imodal da a including RGB, nea -in a ed (NIR), Li-
DAR ele a ion, and e ain a ibu es ac oss 4,480 documen ed habi a
changes o e 15.3 km², ou analysis e eals key pe o mance di e ences.
Wi hin pos -classi ica ion, ounda ion models signi ican ly ou pe o m
CNNs wi h Clay achie ing 0.60 O e all Accu acy (OA) e sus U-Ne ’s
0.49 OA. Be ween pa adigms, di ec me hods achie e supe io change
de ec ion wi h In e sec ion O e Union (IoU) sco es o 0.48-0.49 com-
pa ed o pos -classi ica ion app oaches a IoU 0.28. The esul s p o ide
ac ionable guidance o p o ec ed a ea manage s choosing be ween AI
pa adigms o au oma ed habi a moni o ing.
Keywo ds: Change De ec ion ·Remo e Sensing ·P o ec ed A eas ·
Founda ion Models ·Alpine Habi a s
1 In oduc ion
Alpine p o ec ed a eas ace unp eceden ed landscape ans o ma ion due o cli-
ma e change, wi h ecosys ems wa ming a wice he global a e age a e [3].
T adi ional habi a mapping elies on manual in e p e a ion o ae ial image y,
exempli ied by he Habi Alp p ojec [5], bu his app oach is oo labo -in ensi e
o he equen moni o ing ha apidly changing alpine ecosys ems equi e.
2 H. K is en e al.
Recen ad ances in geospa ial AI, pa icula ly Geospa ial Founda ion Models
(FMs) demons a e ema kable capabili ies when ine- uned o Ea h Obse a-
ion asks [7], [4]. Mo eo e , specialized Vision T ans o me a chi ec u es ha e
eme ged o di ec change de ec ion [8]. Howe e , a undamen al ques ion e-
mains: Which au oma ed change de ec ion pa adigm, pos -classi ica ion e sus
di ec change de ec ion, is op imal o ecological moni o ing?
This gap is signi ican gi en ecen analyses showing ha complex models do
no always ou pe o m es ablished baselines, emphasizing he need o igo ous
benchma king [1]. The unique cha ac e is ics o ecological da a se s may a o
di e en app oaches han u ban o ag icul u al applica ions.
Resea ch ques ions
1. How well can AI p ocedu es de ec changes in p o ec ed a ea habi a s?
2. Which pa adigm is mo e sui able: di ec change de ec ion o pos -classi ica ion
me hods?
3. Do ounda ion models imp o e pe o mance compa ed o CNN-based ap-
p oaches?
4. Wha is he added alue o including addi ional inpu da a se s such as
LiDAR-de i ed me ics?
2 Me hodology
2.1 S udy a ea and Da a
Gesäuse Na ional Pa k p o ides he Habi Alp da ase (2003, 2013) wi h 23 habi-
a classes ocusing on o es s uc u e. The da ase encompasses 15.3 km²wi h
4,480 documen ed changes. Da a includes RGB image y (2003, 2013, 2020, 2024),
NIR, and LiDAR-de i ed p oduc s (nDSM, DTM, oughness).
2.2 App oach 1 - Pos Classi ica ion
We ine une U-Ne s [6] using p e ained ImageNe weigh s [2] o seman ic seg-
men a ion. We also ine- une wo geospa ial ounda ion models: P i h i-EO-2.0-
300M (Vision T ans o me p e ained on 4.2M NASA HLS V2 samples) [7] and
Clay 1.0 (ViT p e ained on 47M sa elli e images and 24M op ical samples) [4].
Seman ic segmen a ion is pe o med using da a om 2013 wi h RGB + NIR +
LiDAR inpu laye s o aining/ alida ion, hen applied o 2020/2024 da a o
iden i y habi a changes h ough seman ic classi ica ion compa ison.
2.3 App oach 2 - Di ec Change
We use p e ained U-Ne s simila o app oach 1, wi h s acked empo al bands
o bina y and mul iclass change p edic ion. We also e alua e ChangeViT, which
employs ision ans o me s o bina y change de ec ion, excelling a la ge-scale
pa e n ecogni ion and de ec ing g adual and sudden changes [8]. End- o-end
aining is pe o med on bi- empo al RGB image y (2003-2013) wi h labeled
changes, hen applied o 2013-2020/2024 da a o p ospec i e change de ec ion.
AI-D i en Habi a Change De ec ion in Alpine P o ec ed A eas 3
2.4 E alua ion and Enhancemen
E alua ion me ics include Jacca d Index / In e sec ion O e Union (IoU),
O e all Accu acy (OA), and F1-sco es sui able o imbalanced da ase s. Pos -
p ocessing il e s elimina e change ansi ions ha a e ecologically impossible in
he gi en ime ame (e.g., ock → o es , clea cu →ma u e s age), o imp o e
seman ic consis ency. Mul i-modal analysis e alua es RGB-only e sus RGB +
NIR e sus RGB + NIR + ele a ion combina ions ac oss bo h pa adigms.
3 P elimina y Resul s
3.1 Pe o mance Compa ison
Table 1 summa izes pe o mance ac oss bo h pa adigms. Me ics e lec di e en
e alua ion s a egies due o ongoing da a p epa a ion: pos -classi ica ion shows
seman ic segmen a ion accu acy on 2013 habi a da a, while di ec change de-
ec ion shows change de ec ion pe o mance using 2003-2013 ansi ions.
Table 1. Pe o mance compa ison ac oss pa adigms (di e en e alua ion asks)
Pa adigm Me hod IoU OA
Seman ic Segmen a ion Task (2013 habi a s)
Pos -Classi ica ion U-Ne RGB+NIR+nDSM 0.24 0.43
P i h i-EO-2.0 0.26 0.54
Clay 1.0 0.28 0.60
Change De ec ion Task (2003-2013 ansi ions)
Di ec Change De ec ion U-Ne Bina y 0.48 0.63
ChangeViT Bina y 0.48 0.64
U-Ne Mul iclass 0.35 0.51
Wi h Clay achie ing he bes Pos -Classi ica ion esul s, ounda ion models
pe o m be e han adi ional CNNs. Mul i-modal da a in eg a ion p o es c u-
cial: adding ele a ion da a imp o es U-Ne pe o mance om 0.20 o 0.43 OA.
Bina y change de ec ion achie es high pe o mance ac oss a chi ec u es (IoU:
0.48-0.49), mul iclass app oaches emain mo e challenging. ChangeViT ma ches
U-Ne pe o mance o bina y asks bu o e s be e scalabili y.
3.2 Key insigh s
Founda ion models signi ican ly ou pe o m adi ional CNNs o habi a clas-
si ica ion, wi h Clay achie ing 40% be e OA han U-Ne , while di ec change
de ec ion me hods excel a bina y change iden i ica ion. Ele a ion da a consis-
en ly imp o es pe o mance ac oss all app oaches, highligh ing he impo ance
o h ee-dimensional s uc u al in o ma ion o alpine habi a disc imina ion.
4 H. K is en e al.
4 Expec ed Impac and Fu u e Wo k
This esea ch p o ides a comp ehensi e benchma king amewo k o AI-based
habi a moni o ing in p o ec ed a eas, di ec ly add essing conse a ion man-
age s’ ope a ional needs. Resul s demons a e ha pa adigm choice depends on
speci ic moni o ing objec i es: di ec me hods excel o change de ec ion, while
ounda ion models o e supe io seman ic unde s anding.
Fu u e wo k includes gene a ion o 2020 aining da a and use o 2024 LiDAR
da a wi h o es s uc u e a ibu es. Findings enable p o ec ed a ea manage s
o make in o med decisions abou AI adop ion, po en ially educing moni o ing
cos s while inc easing empo al esolu ion o clima e change adap a ion.
Re e ences
1. Co ley, I., Robinson, C., O iz, A.: A Change De ec ion Reali y Check (Ap 2024).
h ps://doi.o g/10.48550/a Xi .2402.06994, h p://a xi .o g/abs/2402.06994,
a Xi :2402.06994 [cs]
2. Deng, J., Dong, W., Soche , R., Li, L.J., Li, K., Fei-Fei, L.: Im-
ageNe : A la ge-scale hie a chical image da abase. In: 2009 IEEE
Con e ence on Compu e Vision and Pa e n Recogni ion. pp.
248–255 (Jun 2009). h ps://doi.o g/10.1109/CVPR.2009.5206848,
h ps://ieeexplo e.ieee.o g/documen /5206848, iSSN: 1063-6919
3. Eu opean En i onmen Agency: Regional clima e change and adap a ion : he Alps
acing he challenge o changing wa e esou ces. Publica ions O ice, LU (2009),
h ps://da a.eu opa.eu/doi/10.2800/12552
4. Founda ion, C.: Clay 1.0 (2023), h ps://hugging ace.co/made-wi h-clay/Clay, doi
= 10.57967/h /1537
5. Hauens ein, P., Ind a-Cama hias, L.: CC-HABITALP Ka ie ung 1954 - 2003 -
2015. Vo läu ige echnische Schlussbe ich , Hauens ein Geoin o ma ik, Tamins
(Oc 2018)
6. Ronnebe ge , O., Fische , P., B ox, T.: U-Ne : Con olu ional Ne wo ks o Biomed-
ical Image Segmen a ion (May 2015). h ps://doi.o g/10.48550/a Xi .1505.04597,
h p://a xi .o g/abs/1505.04597, a Xi :1505.04597 [cs]
7. Szwa cman, D., Roy, S., F acca o, P., Gíslason, P.E., Blumens iel, B., Ghosal,
R., Oli ei a, P.H.d., Almeida, J.L.d.S., Sedona, R., Kang, Y., Chak abo y, S.,
Wang, S., Gomes, C., Kuma , A., T uong, M., Godwin, D., Lee, H., Hsu,
C.Y., Asanjan, A.A., Mujeci, B., Shidham, D., Keenan, T., A e alo, P., Li,
W., Alemohammad, H., Olo sson, P., Hain, C., Kennedy, R., Zad ozny, B.,
Bell, D., Ca alla o, G., Wa son, C., Maskey, M., Ramachand an, R., Mo eno,
J.B.: P i h i-EO-2.0: A Ve sa ile Mul i-Tempo al Founda ion Model o Ea h
Obse a ion Applica ions (Feb 2025). h ps://doi.o g/10.48550/a Xi .2412.02732,
h p://a xi .o g/abs/2412.02732, a Xi :2412.02732 [cs]
8. Zhu, D., Huang, X., Huang, H., Shao, Z., Cheng, Q.: ChangeViT: Un-
leashing Plain Vision T ans o me s o Change De ec ion (Jun 2024).
h ps://doi.o g/10.48550/a Xi .2406.12847, h p://a xi .o g/abs/2406.12847,
a Xi :2406.12847 [cs]
