T aining and p edic ion o Fco e Abundances o Sen inel 2 image y
Paul Ka lshö e , Ke in Kühl, Pe e Schwind, U a Heiden
DLR Obe p a enho en (DLR), Ea h Obse a ion Cen e (EOC), Ge many
Non-pho osyn he ic
ac i e ege a ion
A: Xylan & Cellulose
B: Lignin & Cellulose
C: Cellulose
Ba e soil
D: Clay
E: Ca bona es
Sen inel-2
(mul ispec al)
EnMAP
(hype spec al)
S2 bands
modi ied Hyb idSNmodel
EO-based ba e soil e lec ance composi es (SRC) a e impo an :
• o la ge scale soil pa ame e analysis
• o e alua e he impac o soil e osion p ocesses.
P os and cons o hype spec al image da a
•Enables pixel le el sepa a ion o ba e soil (BS) and d y ege a ion (NPV)
using spec oscopy
•Limi ed co e age and in equen obse a ions educe usabili y o la ge-
scale, ime-sensi i e applica ions
Downs eam Task: Soil Re lec ance Composi e
SCMaP ained
SRC PV+IR2
PV+IR2index-basedSRC
EnMAP ained
SRC s2 60
Based on min 60% BS abundance / pixel
2,5 km
Ba e soil e lec ance
Conclusion
•The model makes good p edic ion, bu
aining bias is p oblema ic, especially
in unclea si ua ions
•Usage o spec al in o ma ion
meaning ul
•Spa ial ea u es appea e y impo an ,
e en i spec al in o ma ion is cons an
•SRC quali a i ely con incing and
compa able o index-based SRC (mean
RSME 0.013 on Tile in Belgium )
•Focus on cen al Eu ope o aining F om 656 S2-
EnMAP ma chups, he mos p omising 115 a e
selec ed (based on oo p in o e lap, cloudco e ,
e c) (103 aining, 12 alida ion)
•41374077 ainingpa cheswi h size 25x25x10;
andomly lipping aining pa ches in X/Y o educe
isk o o e i ing
•T aining on 4 A100-SXM4-80GB GPUs and 1TB RAM
•Tenso low2.6, Cuda11.0, numpy1.21
•T aining o 4 epochs (8h pe epoch) wi h min
alida ion loss 0.011
P o ided by and modi ied
om M. Bachmann
³Schwind, P. e al. (2024). Using Deep Lea ning To Gene a e F ac ional Vege a ion Co e F om Mul ispec al Da a. 13 h EARSeL Wo kshop on Imaging
Spec oscopy, 2024-04-16 -2024-04-19, Valencia, Spain.
Concep : Le e age he g owing hype spec al a chi e o EnMAP o p edic
ac ional ege a ion co e in Sen inel-2 empo al s acks³
T aining o a modi ied Hyb idSNmodel (see below) wi h S2-EnMAP empo al
pai s ( lowcha le ) ha p edic s co e (pe cen age o PV/NPV/BS pe pixel.
Foo p in s (le )
( ed) S2 scenes in ol ed in aining p ocess
(black) EnMAPlabels
Dis ibu ion o labels ( igh )
Pe class (PV, NPV, BS), equency (his og am) o a ious
abundance le els (10% size o bins) -> aining da a is
no uni o mly dis ibu ed
equency
Binned Fco e pe class
Dis ibu ion o Inpu
Labels pe co e class
Resul (Valida ion –no seen du ing T aining)
In es iga ion o model’s decision making
In luence o spec al ea u es ( ia Occlusion)
G een Vege a ion (PV)
Red-edge (Chlo ophyl-a)
S2-Bands (5 and 6) mos
impo an , ollowed by
B11 –howe e all bands
signi ican
D y Vege a ion (NPV)
OnlyRed-edge S2-Bands
(5 and 6) and Band 11 is
impo an
Ba e su ace / soil (BS)
Impo ance is well
dis ibu ed ac oss all
ea u es
Shape o spec a is mos
impo an
In luence o spa ial ea u es ( ia Pa ch-Al e a ion)
Examples o Pa ch Al e a ion o es model esponse →done o 1000s o pa ches
O iginal 25x25x10 pa ch
Baseline / labeled da a
S eaking
Spe ain ac , bu shape wo se
Shu le se o pixels
his case, 3 deep om bo de Lin. Comb. Wi h cloud spec a
20%
Me hod
mean(
p ed –baseline_p ed)
Commen
Flip en i e Pa ch
[ 0.002 0.003
-0.005]
No impac , since pa o aining ou ine
S eaking Pa ch
b)
[ 0.040
-0.010 -0.029]
Ve y esilien
Shu le Pe iphe y (1)
[ 0.003
-0.004 0.000]
Ve y esilien o 1px deep changes a bo de
Shu le Pe iphe y (3)
c)
[ 0.665
-0.495 -0.118 ]
Ve y sensi i e o 3px deep changes, huge d i o PV
Rand. Cen e (3x3)
[ 0.379
-0.314 -0.061 ]
Sensi i e, bu mo e esilien han c) ( ewe pixels a ec ed)
A i icial Cloud (d)
[ 0.386
-0.276 -0.106 ]
Sensi i e, lin. combina ion o (.8 * spec a, .2 * cloud)
ba cd
Model en o ces ha sum o co e adds up o 1,
Wi h inc easing unce ain y, model ends owa ds PV (bias o aining da a)
Heigh o g een ba
indica es ea u e
impo ance
NPV
BS
PV PV
PV
In es iga ing F ac ional Vege a ion Co e ( co e ) om Sen inel-2 as
Inpu o Ba e Su ace Composi es
Exis ing p oduc : SoilSui e5y
•Mosaic and empo al composi e o all
“ba e su ace pixels” ound in S2 da a
•Classi ica ion is bina y (soil, no soil), --
→ ac ional co e is mo e de ailed
Heiden e al, SoilSui e–Sen inel-2 –Eu ope, 5 yea
composi e (2018-2022), DOI: 10.15489/qkud8cudg596
PV (g een) NPV (blue) and BS ( ed). Clouds and u ban masked. Le : Labels ( co e de i ed
om EnMAP), igh : P edic ion ( co e p ed ic ed om Sen inel-2). NW o Basel Ai po
T ue colo example o SRC in 31UFS (Belgium)
Mean RSME be ween bo h SRCs is 133 o e lec ance da a scaled o [0,10000]
