AerialStamp: dating aerial orthophotos from visual cues and Sentinel-2 time series

AerialStamp: dating aerial orthophotos from visual
cues and Sentinel-2 time series
Clemens Mosig1,*, David Montero1,2, Miguel D. Mahecha 1,2,3,4, Janusch Vajna-Jehle5, Teja
Kattenborn5
1 Institute for Earth System Science and Remote Sensing, Leipzig University, Leipzig, Germany
2 German Centre for Integrative Biodiversity Research (iDiv), Leipzig, Germany
3 Helmholtz Centre for Environmental Research (UFZ), Leipzig, Germany
4 Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI), Leipzig, Germany
5 Chair of Sensor-based Geoinformatics (geosense), University of Freiburg, Freiburg, Germany
Email addresses
Clemens Mosig clemens.mosig@uni-leipzig.de
David Montero david.montero@uni-leipzig.de
Miguel D. Mahecha miguel.mahecha@uni-leipzig.de
Janusch Vajna-Jehle janusch.vajna-jehle@geosense.uni-freiburg.de
Teja Kattenborn teja.kattenborn@geosense.uni-freiburg.de

• Corresponding author: Clemens Mosig ( clemens.mosig@uni-leipzig.de)
  This manuscript is a preprint submitted to EarthArXiv. It has not been peer reviewed. It
  has been submitted for peer review to the ISPRS Journal of Photogrammetry and Remote
  Sensing. Subsequent versions of this manuscript may have different content. If accepted, the
  final version will be available via the “Peer-reviewed Publication DOI” link on the EarthArXiv
  page of this preprint.

AerialStamp: dating aerial orthophotos from visual cues and Sentinel-2
time series
Clemens Mosiga,∗, David Monteroa,b, Miguel D. Mahechaa,b,c,d, Janusch Vajna-Jehlee and
Teja Kattenborne
aInstitute for Earth System Science and Remote Sensing, Leipzig University, Leipzig, Germany
bGerman Centre for Integrative Biodiversity Research (iDiv), Leipzig, Germany
cHelmholtz Centre for Environmental Research (UFZ), Leipzig, Germany
dCenter for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI), Leipzig, Germany
eChair of Sensor-based Geoinformatics (geosense), University of Freiburg, Freiburg, Germany
A R T I C L E I N F O
Keywords:
Aerial orthophoto
Drone
Acquisition date
Sentinel-2
Deep learning
Crowd-sourced data
A B S T R A C T
Aerial imagery from drones and airplanes is increasingly used as reference data for satellite remote
sensing. Linking an orthophoto to the satellite observations of the same time requires its acquisition
date. In crowd-sourced aerial archives, this reported acquisition date can be incorrect or incomplete.
Yet no method exists to verify or estimate it from the data. Here, we develop AerialStamp, a method
that estimates the acquisition date of an aerial orthophoto, from a drone or an aircraft, as a calibrated
probability distribution over the days of the flight year. The model combines DINOv2 embeddings
of 10 cm patches of the orthophoto with the Sentinel-2 time series of the site. For every week of
that series, spatial correlation features describe how well the orthophoto, averaged to 10 m, correlates
across pixels with the Sentinel-2 image of that week. We used 6,984 publicly available orthophotos
from 129 countries in the crowd-sourced database deadtrees.earth. Spatial block cross-validation
yields a median absolute error of 18.5 days (mean 30.3 days), where for 7% of orthophotos the
estimated and submitted dates differed by more than 90 days. Within the dataset of crowd-sourced
aerial images, at least 1.3% of the acquisition dates were detected to be wrong. A further 10.3% fall
on the 1st or 15th of a month, days that are submitted about 1.6 times as often as expected by chance
and therefore likely serve as placeholders. The resulting framework is available as a Python package
(github.com/cmosig/aerialstamp) and is being integrated into the platform deadtrees.earth.

1. Introduction1
   Aerial imagery from drones and airplanes is used across2
   domains as reference for satellite-based remote sensing, for3
   example in land use and land cover mapping and in envi-4
   ronmental monitoring. For vegetation monitoring, a range of5
   studies showed that features such as invasive species ( Kat-6
   tenborn et al., 2019), mortality ( Schiefer et al., 2023; Mosig7
   et al. , 2026a), vegetation cover ( Riihimäki et al. , 2019),8
   and burned areas ( Fernández-Guisuraga et al. , 2018) can9
   be readily identified in drone images and serve as scalable10
   reference for training and validating satellite-based models11
   (Alvarez-Vanhard et al., 2021; Kattenborn et al. , 2021).12
   The acquisition date of an orthophoto makes it possi-13
   ble to synchronize these reference observations with Earth14
   observation satellite data ( Mosig et al. , 2026a; Kattenborn15
   et al. , 2019). Errors in the acquisition date would pair the16
   orthophoto with non-corresponding satellite observations,17
   limiting model training or validation. For applications such18
   as phenological monitoring, date shifts of several weeks19
   already exceed the required accuracy.20
   In crowd-sourced databases such as deadtrees.earth21
   (Mosig et al. , 2026b), OpenAerialMap ( OpenAerialMap,22
   2026), or GeoNadir (GeoNadir, 2026), the date is entered by23
   the contributor, and as with other volunteered geographic24
   ∗Corresponding author
   clemens.mosig@uni-leipzig.de (C. Mosig)
   ORCID (s):
   information, its quality is heterogeneous and rarely verified 25
   (Goodchild, 2007; Senaratne et al. , 2017). For example, 26
   in deadtrees.earth, the 1st and the 15th of the month are 27
   submitted about 1.6 times more often than other days of the 28
   month, and 1 January more than twice as often as expected 29
   by chance, which suggests that these are often placeholder 30
   days. Further, the order of day and month can be confused, 31
   as the established standard depends on the country (e.g. 32
   mm-dd-yyyy vs. dd-mm-yyyy). Verifying such dates by 33
   eye is hardly possible at scale, and missing dates cannot 34
   be checked at all. A method that estimates the acquisition 35
   date from the data would allow checking the plausibility of 36
   submitted dates and proposing dates where they are entirely 37
   missing. Such an approach would also allow recovering 38
   acquisition dates within large aerial mosaics, such as those 39
   of national aerial surveys, where flight lines often differ in 40
   acquisition date. 41
   Estimating when an image was taken has so far been 42
   studied almost exclusively for ground-level photographs. 43
   Early work dated historical photographs ( Palermo et al. , 44
   2012; Müller et al., 2017) and exploited the regular temporal 45
   variation of outdoor webcam scenes ( Jacobs et al. , 2007). 46
   More recent work learns joint representations of location 47
   and time from ground-level images and predicts the month 48
   and hour of capture as cyclical targets ( Zhai et al. , 2018; 49
   Salem et al., 2020; Shatwell et al., 2025), or checks whether 50
   claimed time metadata are consistent with image content and 51
   location (Kakar and Sudha , 2012; Padilha et al. , 2022). For 52
   C. Mosig et al.: Preprint submitted to Elsevier Page 1 of 13

AerialStamp: dating aerial orthophotos
overhead imagery, seasonality has been used as a signal for53
self-supervised pre-training ( Mañas et al. , 2021), and drone54
and satellite observations of phenology have been compared55
across scales ( Klosterman et al. , 2018; Berra et al. , 2019;56
Thapa et al., 2021). To our knowledge, however, no method57
estimates the acquisition day of year of georeferenced aerial58
imagery.59
To fill this gap, we present AerialStamp, a method that60
estimates the day of the year on which an aerial orthophoto61
was acquired based on circular distributions, as is com-62
mon in phenological studies ( Morellato et al., 2010; Pabon-63
Moreno et al. , 2020). It uses the orthophoto, its location64
and submitted year, and the weekly Sentinel-2 time series65
of the site, and returns a probability for every day of the66
year. Specifically, we aim to (i) establish acquisition-date67
estimation of crowd-sourced aerial orthophotos as a task,68
with a spatially blocked evaluation on a dataset of 6,98469
orthophotos worldwide, (ii) quantify how much of the ac-70
curacy comes from Sentinel-2, by comparing a model with71
and without the Sentinel-2 time series, and from the spatial72
correlation features that compare the orthophoto with each73
week of that series, and (iii) demonstrate the utility of the74
method using the example of deadtrees.earth, by measuring75
how many of its submitted dates are potentially wrong or76
placeholders.77
2. Data78
2.1. Orthophotos79
We used the 6,984 orthophotos of deadtrees.earth80
(Mosig et al. , 2026b) that are publicly available under81
a CC BY license and have a Sentinel-2 series, that is,82
6,689 drone and 295 airborne acquisitions with a median83
ground sampling distance of 3.3 cm (10th to 90th percentile84
1.3–7.7 cm). They cover 129 countries on all continents85
except Antarctica (Fig. 1, Table 1). Most orthophotos86
lie in Europe (39%) and North America (20%), followed87
by South America (15%), Asia (14%), Africa (9%) and88
Oceania (2%), and 25% lie in the southern hemisphere.89
The flights date from 2015 to 2026, with a median of 202390
and 81% since 2020. For 5,180 orthophotos, the manual91
audit of deadtrees.earth ( Mosig et al. , 2026b) provides an92
area of interest (AOI), a multi-polygon of the part of the93
orthophoto without distortions; for the other 1,804, the94
whole orthophoto served as AOI.95
2.2. Date labels96
The acquisition date of each orthophoto is provided by97
the contributor at upload. We refer to it as the submitted98
date and to its year as the submitted year , and used both99
as provided. Of the 6,984 orthophotos, 6,936 have a full100
date (year, month and day) and 48 only a year and month;101
orthophotos with only a submitted year were not included.102
The 1st and the 15th of the month are over-represented103
among the full dates, with 360 and 351 orthophotos against104
about 228 expected for any day of the month, as is 1 January105
with 51 against about 19. We assume that these days were106
Figure 1: Locations of the 6,984 orthophotos, one translucent
point each, in the Equal Earth projection (EPSG:8857).
Table 1
Orthophotos by continent and biome (6,984 in total; share
in %). Biome separation is based on Olson et al. (2001)
and consists of boreal and montane, dryland, temperate and
(sub) tropical sites: 50% are temperate, 33% (sub) tropical,
11% drylands and 6% boreal and montane. Continents follow
the 1:110m countries of Natural Earth ( Natural Earth , 2026);
remote islands are sites more than 200 km from any of these
country polygons.
Continent 𝑛 % Biome 𝑛 %
Europe 2,740 39 Temperate 3,493 50
North America 1,369 20 (Sub) Tropical 2,313 33
South America 1,014 15 Drylands 743 11
Asia 984 14 Boreal and Montane 433 6
Africa 622 9
Oceania 174 2
Remote islands 81 1
often entered when only the month or the year was known 107
and refer to them as placeholder days. The 711 full dates on 108
placeholder days (10.3%) were excluded from training and 109
evaluation. They comprise 51 dates on 1 January (1 January 110
placeholders) and 660 dates on the 1st or 15th of a month 111
other than 1 January ( 1st/15th placeholders). 112
The remaining 6,225 full dates are referred to as clean 113
dates and form the evaluation set. They were used for both 114
training and evaluation in a five-fold spatial block cross- 115
validation (Section 2.4), in which each clean date is pre- 116
dicted by models trained on the other four folds. The 48 117
year-and-month labels were added to the training folds only, 118
which gives 6,273 training labels in total. 119
2.3. Sentinel-2 120
For each orthophoto location, a Sentinel-2 (Drusch et al., 121
2012) Level-2A time series of 7-day mean composites of the 122
12 bands at 10 m resolution in the local Universal Transverse 123
Mercator (UTM) zone was prepared with sentle, an open- 124
source Python package for Sentinel-1 and -2 data cube 125
generation, as in Mosig et al. (2026a). Clouds were masked 126
with the CloudSEN12 cloud classifier ( Aybar et al. , 2022) 127
and snow with the snow test of Fmask ( Zhu and Woodcock, 128
2012), both as implemented in sentle. The series start in July 129
2015 and end when they were downloaded, between July 130
2025 and May 2026. All 12 bands enter the S2 model as 131
C. Mosig et al.: Preprint submitted to Elsevier Page 2 of 13

AerialStamp: dating aerial orthophotos
image-only model
added in the S2 model
Orthophoto
16 random 10 cm
patches from the area
of interest (AOI)
Site facts
geolocation, submitted
year, ground sampling
distance
Orthophoto at 10 m
on the Sentinel-2 grid
Sentinel-2 series
weekly, in the window
(submitted year ±35 d)
DINOv2 ViT-B/14,
frozen
→ patch embeddings
MLP
Week tokens
one per week: bands,
indices, chromaticity,
clear-sky share, position,
12 correlation features
projection to 128,
attention pooling
trunk MLP
→ context ℎ
Seasonal head
MLP on sine–cosine (𝑑)
and context ℎ
→ 𝓁season𝑑 for 365 days
Sentinel-2 head
2 convolutions over weeks,
Gaussian kernel to days
→ 𝓁S2𝑑
+
softmax
𝑝𝑑 , 𝑑 = 1, … , 365
day of year
ℎ
Figure 2: Overview of the two models. The image-only model consists of the components on the orange background and predicts
the distribution from the seasonal head alone. The S2 model adds the components on the blue background and predicts the
softmax of the sum of the seasonal and the Sentinel-2 scores (Eq. 1). In the S2 model, the orthophoto enters twice: as 16 random
10 cm patches embedded by DINOv2, and averaged to 10 m in the correlation features, which compare it with the Sentinel-2
image of each week (Table 2). Gray boxes: inputs; white boxes: network modules.
time series; the visible bands (B02, B03, B04) and the near-132
infrared band (B08) are additionally compared with the RGB133
orthophoto (Section 3.2).134
2.4. Spatial blocking135
To prevent spatial autocorrelation between training and136
test data from inflating the accuracy (Kattenborn et al., 2022;137
Nowosad et al., 2026), we split the data by spatial blocks. We138
assigned each orthophoto to a 50 km cell in the Equal Earth139
projection (EPSG:8857). Orthophotos of one campaign are140
often uploaded together with the same date, so we merged141
cells whenever two orthophotos within 500 km share the142
uploading account and flight date, or the authors listed in the143
metadata and flight date. The resulting 1,308 spatial blocks144
were assigned to five folds by group 𝑘-fold with the blocks145
as groups, stratified by biome ( Roberts et al. , 2017; Ploton146
et al., 2020).147
3. Methods148
AerialStamp consists of two models that estimate the ac-149
quisition date of an orthophoto as a probability distribution150
over the days of the submitted year, which we assume to be151
the flight year (Fig. 2):152

1. The image-only model uses only the orthophoto and153
   the site facts (geolocation, submitted year and ground154
   sampling distance).155
2. The S2 model (S2: Sentinel-2) is the same model with156
   the weekly Sentinel-2 series of the site as additional157
   input. For each week, spatial correlation features (in158
   the following, correlation features) describe how well159
   the orthophoto matches the Sentinel-2 image of that160
   week.161
   The inputs are described in Section 3.2, the model in Sec-162
   tion 3.3, and training and calibration in Sections 3.4 and 3.5.163
   3.1. Target 164
   We divided the year into 365 day bins on a circle, so 165
   that 31 December is adjacent to 1 January; leap years lose 166
   one bin. The model returns a probability 𝑝𝑑 for every day 167
   𝑑. Although Sentinel-2 enters as weekly composites, the 168
   estimate has daily resolution: the seasonal head scores every 169
   day directly, and the Sentinel-2 head spreads its weekly 170
   scores to days (Section 3.3). A full date becomes a wrapped 171
   Gaussian target 𝑦 (𝜎 = 3 days) centered on the submitted 172
   day, and a year-and-month label a uniform target over that 173
   month. A distribution on the circle is required because many 174
   orthophotos visually fit several seasons, for example spring 175
   and autumn, or two wet seasons in the tropics. 176
   3.2. Inputs 177
   The S2 model receives three inputs (Fig. 2): 10 cm 178
   patches of the orthophoto, week tokens built from the 179
   Sentinel-2 series of the site, and site facts , i.e. the 180
   geolocation, the submitted year and the ground sampling 181
   distance. The image-only model receives the 10 cm patches 182
   and the site facts. 183
   10 cm patches. We randomly drew 16 patches of 224 × 184
   224 pixels at 10 cm resolution (22.4 m side) inside the AOI 185
   and embedded each with a frozen DINOv2 vision trans- 186
   former with registers (ViT-B/14: base size, 14-pixel tokens; 187
   Oquab et al., 2024; Darcet et al., 2024). The class token and 188
   the mean patch token together form the patch embedding of 189
   1,536 values. DINOv3 backbones (Siméoni et al., 2025) did 190
   not improve our results (Appendix A). 191
   Week tokens. One token for each weekly Sentinel-2 192
   composite in the window from 35 days before to 35 days 193
   after the submitted year, totaling up to 64 weeks. Each token 194
   holds the AOI mean reflectance of the 12 bands (square- 195
   root transformed, which spreads the low reflectance values 196
   typical of vegetation), the normalized difference vegetation 197
   C. Mosig et al.: Preprint submitted to Elsevier Page 3 of 13

AerialStamp: dating aerial orthophotos
Table 2
Orthophoto-to-Sentinel-2 correlation features of a week, com-
puted over the matched pixels. R, G, B: orthophoto bands
averaged to 10 m; B02, B03, B04, B08: Sentinel-2 blue, green,
red and near-infrared bands. Brightness, chromaticity (chrom.)
and the indices are defined in Section 3.2.
Feature Orthophoto Sentinel-2
Pearson correlation across matched pixels
red, green, blue R, G, B B04, B03, B02
brightness brightness brightness
green chromaticity green chrom. green chrom.
NGRDI NGRDI NGRDI
NGRDI vs. NDVI NGRDI NDVI
brightness vs. near infrared brightness B08
Spearman correlation across matched pixels
brightness brightness brightness
Other
green chromaticity difference mean mean
red chromaticity difference mean mean
matched share matched / orthophoto-covered pixels
index (NDVI; Rouse et al., 1974), the normalized difference198
moisture index (NDMI; Wilson and Sader , 2002), the nor-199
malized green–red difference index (NGRDI;Tucker, 1979),200
the green and red chromaticity ( Gillespie et al. , 1987), the201
clear-sky share (the share of clear AOI pixels), the position202
of the week in the year, and 12 orthophoto-to-Sentinel-2203
correlation features (Table 2). NDVI, (𝐵08 − 𝐵04)∕(𝐵08 +204
𝐵04), and NDMI, (𝐵08− 𝐵11)∕(𝐵08+ 𝐵11), are commonly205
used indices of canopy greenness and water content. NGRDI206
and the chromaticities describe greenness and color in the207
visible bands alone and can therefore also be computed from208
the RGB orthophoto. With the brightness defined as the sum209
of the visible bands ( 𝑅 + 𝐺 + 𝐵 for the orthophoto, 𝐵02 +210
𝐵03+𝐵04 for Sentinel-2), the green chromaticity is the share211
of the green band in the brightness, e.g. 𝐺∕(𝑅 + 𝐺 + 𝐵), the212
red chromaticity is defined analogously, and NGRDI is (𝐺 −213
𝑅)∕(𝐺+𝑅) for the orthophoto and(𝐵03−𝐵04)∕(𝐵03+𝐵04)214
for Sentinel-2.215
To compute the correlation features, we averaged the216
RGB orthophoto onto the 10 m Sentinel-2 grid and com-217
pared it with the Sentinel-2 image of the week over the218
matched pixels: those that the orthophoto covers to at least219
80% and that are clear in Sentinel-2. Nine features are220
correlations across the matched pixels between one quantity221
of the orthophoto and one of Sentinel-2, for example the red222
channel and band B04 (Table 2): eight Pearson correlations223
and, for brightness, also the Spearman rank correlation.224
The other three are the differences in mean green and red225
chromaticity between Sentinel-2 and the orthophoto and226
the share of matched pixels. A correlation across pixels227
measures whether the two images share a spatial pattern,228
for example which fields are green and which are bare.229
The orthophoto values were used as stored, as 8-bit digital230
numbers without radiometric calibration. A correlation does231
not change when a quantity is linearly rescaled, and the chro-232
maticities and NGRDI do not depend on the overall scale233
of the values, so orthophotos from uncalibrated consumer234
cameras can be compared with Sentinel-2 reflectance.235
Site facts. The geolocation as latitude and longitude in 236
a sine–cosine encoding, which maps a cyclic variable 𝑥 237
with period 𝑃 to sin(2𝜋𝑥∕𝑃 ) and cos(2𝜋𝑥∕𝑃 ) so that both 238
ends of the cycle are close ( Mac Aodha et al. , 2019) (here 239
𝑃 = 360 ◦); in addition, the latitude itself, the hemisphere, 240
the submitted year and the ground sampling distance. The 241
submitted year also selects the window of the week tokens. 242
The position of the week in the week tokens is encoded in 243
the same way ( 𝑃 = one year). 244
3.3. Model 245
The S2 model scores every day 𝑑 of the year with two 246
heads (Fig. 2); the image-only model uses only the first. 247
The seasonal head returns a score 𝓁season𝑑 for how well day 248
𝑑 fits the appearance of the orthophoto at its location, and 249
the Sentinel-2 head a score 𝓁S2𝑑 for how well the orthophoto 250
matches the Sentinel-2 images of the weeks around day 𝑑. 251
The predicted distribution is the softmax of their sum, 252
𝑝𝑑 = sof tmax𝑑
(𝓁season𝑑 + 𝓁S2𝑑 ). (1)
Both heads receive a context vector ℎ that summarizes 253
the orthophoto and its site. The embeddings of the 10 cm 254
patches are layer-normalized, projected to 128 dimensions 255
and pooled by multi-head attention with one learned query; 256
a trunk multilayer perceptron (MLP) combines the pooled 257
embedding with an embedding of the site facts into ℎ (256 258
dimensions). The seasonal head applies a two-layer MLP 259
to the sum of linear projections of ℎ and of a sine–cosine 260
encoding of 𝑑 (first four harmonics). The Sentinel-2 head 261
adds a projection of ℎ to each projected week token, applies 262
two 1-D convolutions of width 5 weeks over the sequence of 263
weeks, in which missing weeks are set to zero, and returns a 264
score 𝑔𝑤 per week 𝑤. A Gaussian kernel 𝑘𝑑𝑤 = exp ( − (𝑑 − 265
𝑡𝑤)2∕2𝑏2) with a learned width 𝑏 spreads the week scores to 266
days, 267
𝓁S2𝑑 =
∑
𝑤 𝑘𝑑𝑤 𝑔𝑤∑
𝑤 𝑘𝑑𝑤 + 0.1 , (2)
where 𝑡𝑤 is the position of week 𝑤 in days from 1 January of 268
the submitted year, so that the weeks of the±35-day margins 269
inform the first and last days of the year. The sums run over 270
all weeks of the window, cloudy ones included, whose clear- 271
sky share is part of their token. The constant 0.1 lets the 272
term fade out where the series has no week near day 𝑑, for 273
example after its end. 274
3.4. Training 275
To cope with potentially wrong submitted dates, we 276
trained both models with a noise-robust negative log- 277
likelihood of the target 𝑦 as the loss, 278
 = − log
( 𝜀
365 + (1 −𝜀)
∑
𝑑
𝑝𝑑 𝑦𝑑
)
, 𝜀 = 0.05, (3)
which mixes the predicted distribution with a uniform dis- 279
tribution and thus caps the loss of a single wrong label 280
at − log(𝜀∕365). We trained with AdamW ( Loshchilov and 281
C. Mosig et al.: Preprint submitted to Elsevier Page 4 of 13

AerialStamp: dating aerial orthophotos
Hutter, 2019) (learning rate 5⋅10−4, weight decay 0.05, one-282
cycle schedule (Smith and Topin, 2019), batch size 128) for283
32 epochs without early stopping.284
We used two data augmentations. In each step, 20% of285
the 16 patch embeddings were dropped. In addition, in 30%286
of the training samples the Sentinel-2 series was cut after287
a uniformly random day between 1 March of the submit-288
ted year and the end of the window ( window truncation ),289
because new uploads are recent and their Sentinel-2 series290
often ends inside the submitted year (Appendix B).291
We trained five models with different random initializa-292
tions and averaged their predicted distributions (ensemble);293
in cross-validation, this gives five models per fold.294
The image-only model was trained with the same recipe295
and on the same orthophotos as the S2 model, but without296
the Sentinel-2 head.297
3.5. Calibration298
A predicted distribution is defined as calibrated when its299
prediction intervals contain the true date as often as their300
probability states, e.g. the 80% intervals for 80% of the or-301
thophotos (Section 3.7). We calibrated the model ensemble302
with two parameters (Guo et al., 2017): a temperature 𝑇 that303
flattens the distribution for 𝑇 > 1, and a circular Gaussian304
smoothing of width 𝑠 days,305
𝑝cal ∝ smooth 𝑠 (𝑝1∕𝑇 ). (4)
For the orthophotos of each test fold,𝑇 and 𝑠 were chosen on306
a grid to minimize the negative log-likelihood (NLL) of the307
submitted full dates, placeholder days included, in the out-308
of-fold predictions of the other four folds. The calibration309
thus never sees the test labels. For the S2 model, this gave310
𝑇 = 2 .0 and 𝑠 = 0 in every fold, and for the image-only311
model 𝑇 = 1 .7 and 𝑠 = 10 days in four folds and 𝑇 = 2 .0312
and 𝑠 = 0 in one; values beyond the grid lowered the NLL313
by less than 0.001.314
3.6. Naïve baselines315
We additionally compare our models against three naïve316
baselines that predict the flight date from the typical flight317
dates of the region, from the phenology of the site, or from318
the most common flight date overall. The free parameters319
of Climatology and Phenology were chosen by maximum320
likelihood on the training folds. Most common date predicts321
for every orthophoto the most common flight day of the year322
in the training folds, the mode of a kernel density ( 𝜎 = 7323
days) of the training flight dates (3–5 July, depending on the324
fold), and uses that density as its distribution. Climatology325
is a kernel density ( 𝜎 = 7 days) of the training flight dates,326
weighted by a Gaussian of the distance to the site, and mixed327
with the flight-date distribution of the same hemisphere.328
Phenology assumes that flights are more likely in leaf, 𝑝𝑑 ∝329
(𝑐𝑑 + 𝑎)𝛾 , with 𝑐𝑑 the phenology curve of the site, its long-330
term probability of being in the maximum greenness period331
on day 𝑑 from satellite land surface phenology (Zhang et al.,332
2018), as in Mosig et al. (2026a).333
3.7. Evaluation 334
All numbers are out-of-fold on the five spatially blocked 335
folds. The predicted date is the circular median of 𝑝, which 336
minimizes the expected absolute error. The 𝑞 prediction 337
interval is the smallest set of days that together hold the 338
probability 𝑞 (highest-density region; Hyndman, 1996). The 339
error of an orthophoto is the circular distance between the 340
predicted and the submitted date, the shorter way around the 341
year (at most 182 days), so that 28 December and 3 January 342
are 6 days apart. We report the mean absolute error (MAE) 343
and the median error in days, and the share of orthophotos 344
within ±30 days of the submitted date. Because the spatial 345
blocks were very uneven (the ten largest hold 29% of the 346
training labels), we also report the block-weighted MAE 347
(MAE bw), with each spatial block weighted equally; all 348
other metrics weight every orthophoto equally. 349
3.8. Ablations 350
To quantify the contribution of the individual inputs, we 351
retrained the S2 model with one component changed at a 352
time: the image input to DINOv2, the patch embedding, and 353
the content of the week tokens. Each variant was trained on 354
the same folds as an ensemble of three models and was not 355
calibrated (see Section 3.5). The variants are compared by 356
the block-weighted MAE. 357
3.9. Validation with dense multitemporal drone 358
data 359
To assess the consistency and the temporal biases of the 360
models along dense time series, and to demonstrate them at 361
well-known sites, we examined three sites flown repeatedly: 362
the Leipzig floodplain forest (Germany), the ECOSENSE 363
research station near Freiburg (Germany) and the 50 ha 364
forest dynamics plot on Barro Colorado Island (Panama). We 365
used the out-of-fold predictions; since the folds are grouped 366
by spatial block, no flight was predicted by a model that 367
saw its site. In addition to the MAE, we report Spearman’s 368
𝜌 between submitted and predicted day of year to reflect 369
whether the model was successful in ordering the flights 370
correctly, the share of correctly ordered pairs of flights of 371
the same year by their time apart, and, for flights in July and 372
August, whether the predicted distribution puts more proba- 373
bility into the submitted month than into the other. Since the 374
flights of a series fall within one calendar year, we computed 375
𝜌 on a linear time axis, after shifting each predicted date 376
by a full year where this brought it within half a year of 377
the submitted date. We did not use a circular correlation 378
coefficient ( Jammalamadaka and SenGupta , 2001), as it is 379
measured around the mean direction of the dates, which is 380
poorly defined for flights spread over the whole year. 381
3.10. Plausibility of submitted dates 382
For each full submitted date, we computed its plausibil- 383
ity 𝜋: the total probability of all days that the out-of-fold 384
prediction of the S2 model finds at most as likely as the 385
submitted day. Like a𝑝-value, 𝜋 is small when the submitted 386
day is among the least likely days, and for a calibrated model 387
and correct dates, a share 𝛼 of the dates has 𝜋 < 𝛼 by 388
C. Mosig et al.: Preprint submitted to Elsevier Page 5 of 13

AerialStamp: dating aerial orthophotos
Table 3
Acquisition-date estimation on the 6,225 evaluation orthophotos (out-of-fold, spatial block cross-validation). All methods except
Most common date also use the location of the site. MAE and median absolute error in days; bw: block-weighted; ±30: share of
orthophotos dated within 30 days of the submitted dates.
Method Input MAE MAE bw median ±30
Baselines
Most common date training flight dates 62.5 66.3 51.5 .32
Climatology training flight dates 55.0 59.1 41.0 .39
Phenology phenology curve 56.7 59.6 43.5 .36
Learned models
Image-only model 10 cm patches, site facts 38.1 41.2 25.5 .56
S2 model 10 cm patches, site facts, week tokens 30.3 33.2 18.5 .66
Table 4
The methods of Table 3 by biome. MAE and median absolute error in days; 𝑛: evaluation orthophotos of the biome.
Temperate Boreal and Montane (Sub) Tropical Drylands
𝑛 = 3,148 𝑛 = 388 𝑛 = 2,031 𝑛 = 656
Method MAE median MAE median MAE median MAE median
Baselines
Most common date 45.8 36.5 37.9 23.5 86.6 81.6 82.3 76.6
Climatology 43.2 34.5 30.0 19.5 72.3 59.5 73.1 65.5
Phenology 41.9 34.5 37.5 23.7 74.8 62.5 83.1 78.1
Learned models
Image-only model 28.6 20.5 23.6 13.5 50.8 36.5 52.5 36.5
S2 model 24.0 15.9 18.7 10.5 37.4 22.5 45.6 32.3
Table 5
Ablation of the S2 model: block-weighted MAE (days) on
the evaluation set when one component is changed, and
the difference Δ to the S2 model. Uncalibrated three-seed
ensembles, hence 32.9 and 41.9 instead of 33.2 and 41.2
days as in Table 3; seed-to-seed standard deviation 0.1–0.5
days. Color statistics: 34 per-patch statistics (means, standard
deviations and quantiles of color, brightness and the excess
green index ( Woebbecke et al. , 1995), shares of color classes,
and a hue histogram).
Variant MAE bw Δ
S2 model 32.9
Image input to DINOv2 (S2 model: 16 random 10 cm patches)
none (orthophoto only in the correlation features) 39.4 +6.4
overview: AOI as one 224 × 224 image 40.5 +7.6
four 224 × 224 tiles at 1 m 37.5 +4.6
one 10 cm patch 36.0 +3.1
16 patches, mean instead of attention pooling 33.6 +0.7
Patch embedding (S2 model: DINOv2)
34 color statistics instead of DINOv2 40.3 +7.4
DINOv2 and the 34 color statistics 33.8 +0.8
Week tokens (S2 model: bands, indices, correlation features)
correlation features only 33.1 +0.2
bands and indices only 40.8 +7.9
none (image-only model) 41.9 +8.9
chance. Equivalently, a date with 𝜋 < 𝛼 lies outside the389
1 − 𝛼 prediction interval. If 𝑛𝛼 of 𝑁 dates have 𝜋 < 𝛼 ,390
𝑛𝛼 − 𝛼𝑁 therefore estimates the number of wrong dates. The391
Table 6
Submitted full dates with a plausibility 𝜋 below 0.05, 0.01
and 0.001 under the S2 model (out-of-fold); in brackets, the
number expected by chance for a calibrated model if all dates
were correct. Clean: dates not on placeholder days.
Full dates 𝑛 𝜋 < 0.05 𝜋 < 0.01 𝜋 < 0.001
Clean dates 6,225 341 (311) 107 (62) 26 (6.2)
1st/15th placeholders 660 65 (33) 30 (6.6) 4 (0.7)
1 January placeholders 51 33 (2.6) 23 (0.5) 11 (0.05)
All 6,936 439 (347) 160 (69) 41 (6.9)
estimate is a lower bound, because a wrong date that falls 392
into a plausible season cannot be detected, and because the 393
calibration was fitted on submitted dates that include wrong 394
ones, which widens the intervals. How well the month of a 395
date fits is measured by the share of the predicted probability 396
that falls into the submitted month. To test for confused day 397
and month, we swapped them in the implausible dates ( 𝜋 < 398
0.01) where this gives a different valid date, and compared 399
the number of swapped dates with 𝜋 ≥ 0.05 with the number 400
expected for a random day of the year. 401
C. Mosig et al.: Preprint submitted to Elsevier Page 6 of 13

AerialStamp: dating aerial orthophotos
4. Results402
4.1. Accuracy of the estimated acquisition dates403
The S2 model estimated the acquisition date with a me-404
dian error of 18.5 days and an MAE of 30.3 days (Table 3).405
66% of the orthophotos were dated within ±30 days and 23%406
within a week (Appendix C). The best baseline, Climatol-407
ogy, had a median error of 41.0 days, and Phenology 43.5408
days. The image-only model, which uses the 10 cm patches409
and the site facts but no Sentinel-2 data, reached a median410
error of 25.5 days (MAE 38.1 days).411
The prediction intervals were calibrated: the 50, 80 and412
95% intervals contained the submitted date for 53, 80 and413
95% of the orthophotos (Table C.1). Before calibration, the414
ensemble was over-confident: its 80% intervals were 55 days415
wide and contained the submitted date for only 61% of416
the orthophotos. After calibration, the 80% intervals were417
96 days wide on average, against 157 and 197 days for418
Climatology and Phenology.419
Performance depended on the biome (Table 4). Temper-420
ate and boreal and montane sites have sharp seasons and421
were dated with median errors of 15.9 and 10.5 days (MAE422
24.0 and 18.7 days). (Sub) tropical sites (median 22.5, MAE423
37.4 days) and drylands (32.3 and 45.6 days) were harder,424
and so was the southern hemisphere (MAE 39.5 against 27.4425
days in the northern hemisphere), which has fewer training426
labels. In every biome, the S2 model reduced the MAE by427
37 to 48% against the better of Climatology and Phenology,428
whose MAE ranged from 30.0 to 73.1 days across biomes.429
Coverage of the 80% intervals ranged from 0.73 (drylands)430
to 0.86 (boreal and montane).431
The errors had a long tail, where for 7% of the orthopho-432
tos, the predicted and the submitted dates differed by more433
than 90 days, and these orthophotos accounted for 30% of434
the summed absolute error. The tail occurred in every biome435
(3% of the temperate and 14% of the dryland orthophotos)436
but was concentrated at (sub) tropical and dryland sites.437
The ablations (see Table 5) showed that without the438
correlation features, the Sentinel-2 bands and indices of the439
window improved the block-weighted error by 1.0 days over440
the image-only model (40.8 against 41.9 days), whereas the441
correlation features alone gave almost the full gain (33.1442
against 32.9 days). The 10 cm patches reduced the error443
from 39.4 days without an image input to 32.9 days. An444
overview image of the whole AOI (40.5 days) was no better445
than no image input, and four 1 m tiles (37.5 days) gave less446
than a third of the gain of the 10 cm patches; the S2 model447
without an image input still sees the orthophoto through the448
correlation features. One patch versus 16 increased the error449
by 3.1 days, mean instead of attention pooling by 0.7 days.450
The 34 simple color statistics were weaker than the DINOv2451
embeddings (40.3 days) and also did not improve them (+0.8452
days).453
4.2. Plausibility 𝜋 of the submitted dates454
Of the 6,984 orthophotos, 6,936 have a full submitted455
date. Of these, 160 had a plausibility 𝜋 < 0.01 (Sec-456
tion 3.10), where 69 are expected by chance, and 41 had457
𝜋 < 0.001, where 7 are expected (Table 6). At least 91 458
dates (1.3%) are therefore wrong, and about 34 of the 41 459
dates with 𝜋 < 0.001 are likely genuine errors. For the 460
711 dates on placeholder days (10.3%) that were excluded 461
from the experiments (see Section 2.2), the results differed 462
depending on the date. For the 1st/15th placeholders, the 463
month was almost as plausible as for clean dates: 28% of 464
the probability fell into the submitted month, against 30% 465
for clean dates. Still, these dates had 𝜋 < 0.01 more than 466
four times as often as expected by chance. The 1 January 467
placeholders were mostly wrong altogether: 75% had 𝜋 < 468
0.2, against 20% expected, their predicted date was off by 469
120 days on average, and only 6% of the probability fell 470
into January. Confused day and month explained few of the 471
implausible dates: of the 160 dates with 𝜋 < 0.01, 58 give 472
a different valid date when day and month are swapped, and 473
the swapped date was plausible ( 𝜋 ≥ 0.05) for 30 of them, 474
against 26 expected for a random day of the year. 475
Fig. 3 shows four examples. In the two cases with an 476
implausible submitted date (Fig. 3b and 3c), the S2 model 477
places the flight five and eight months after the submitted 478
date, in a season in which the Sentinel-2 composite resem- 479
bles the orthophoto visually more closely. In Fig. 3d, the 480
distribution has two modes of similar size and the submitted 481
date lies in one of them, so it is plausible (𝜋 = 0.86) although 482
it lies 128 days from the predicted date, which falls next to 483
the other mode. 484
4.3. Validation with dense multitemporal drone 485
data 486
At the Leipzig floodplain forest (Germany), flown 43 487
times at near-weekly intervals from February to December 488
2025, the predicted dates followed the submitted ones with 489
Spearman’s 𝜌 = 0 .99 and an MAE of 13.3 days for the S2 490
model and 17.8 days for the image-only model (Fig. 4a). 491
With the S2 model, flights 7–14 days apart were ordered 492
correctly in 76% of the pairs, 14–30 days apart in 91% and 493
30–60 days apart in 97%, and six of the seven July and 494
August flights were assigned to the right month. Since the 495
image-only model alone reached 𝜌 = 0 .99, within a site 496
the signal lies in the orthophoto itself. The S2 model pulled 497
flights towards midsummer: flights from late May to mid 498
June were predicted two to six weeks late, and those in the 499
second half of September four to six weeks early. For the 500
nine flights from November to February, both models were 501
uncertain: the range between the 10th and 90th percentiles 502
of the distribution spanned 124 days on average for the S2 503
model and 131 days for the image-only model. 504
At the ECOSENSE research station (Werner et al., 2024) 505
near Freiburg (Germany), the S2 model dated the 27 flights 506
of 2024–2026 with an MAE of 15.4 days, against 21.6 days 507
for the image-only model (Fig. 4b): 16 flights in 2024 with 508
14.2 days ( 𝜌 = 0 .99) and 10 flights in 2025 with 18.5 days 509
(𝜌 = 0 .93), against 19.9 and 21.3 days for the image-only 510
model. 511
The 50 ha forest dynamics plot on Barro Colorado Island 512
(Panama) is a semi-deciduous tropical forest, in which many 513
C. Mosig et al.: Preprint submitted to Elsevier Page 7 of 13

AerialStamp: dating aerial orthophotos
27 Jun
27 Jun
J F M A M J J A S O N D
submitted 23 Jun 2024
predicted 26 Jun 2024
80% interval
Germany
id 6192
orthophoto sample 10 cm patch
Sentinel-2 RGB,
submitted date
Sentinel-2 RGB,
predicted date predicted distribution
(a) Plausible submitted date ( 𝜋 = 0.91).
12 Mar
20 Aug
J F M A M J J A S O N D
submitted 16 Mar 2020
predicted 19 Aug 2020
80% interval
T urkey
id 10819
(b) Implausible submitted date ( 𝜋 = 0.003): submitted in March, predicted for late August.
24 Mar
24 Nov
J F M A M J J A S O N D
submitted 31 Mar 2022
predicted 23 Nov 2022
80% interval
Mozambique
id 7460
(c) Implausible submitted date ( 𝜋 < 0.001): dry vegetation, submitted date at the end of the wet season.
28 Mar
18 Jul
J F M A M J J A S O N D
submitted 26 Mar 2019
predicted 31 Jul 2019
80% interval
T aiwan
id 8226
(d) Bimodal distribution at a (sub) tropical site; the submitted date lies in one of two modes of similar size ( 𝜋 = 0.86).
Figure 3: Four evaluation orthophotos with out-of-fold predictions of the S2 model and the plausibility 𝜋 of the submitted date
(Section 3.10). Columns from left to right: the orthophoto over its AOI, a sample 10 cm patch, the Sentinel-2 RGB composite
(bands B04, B03, B02) of the weekly composite nearest the submitted date (red frame) and nearest the predicted date (black
dashed frame), each with at least 90% clear AOI pixels and labeled with its date, and the predicted distribution over the submitted
year. In the distribution, the red line marks the submitted date, the dashed line the predicted date, and the shaded area the 80%
prediction interval. Dataset identifiers are accessible at deadtrees.earth/dataset/ ⟨ID⟩.
trees lose some or all of their leaves in the dry season from514
mid-December to mid-April (Park et al., 2019). It was flown515
about monthly from 2018 to 2023 and weekly in the 2023 dry516
season (90 flights; Vásquez et al. , 2023) and was dated less517
accurately than the two temperate sites. The S2 model dated518
these flights with an MAE of 29.1 days ( 𝜌 = 0 .97 over all519
years), and the image-only model with 69.9 days (Fig. 4c).520
With the image-only model, flights from January to March,521
the dry season, were dated two to five months off on average522
(MAE 64–168 days), and all 20 were placed in September to523
December. With the S2 model, the errors dropped to 53 days 524
in January, 33 days in February and 7 days in March, and 525
six of seven March flights fell into March, whereas flights 526
from May to August were predicted three to six weeks late. 527
Among the weekly flights of 2023, the S2 model dated flights 528
with an MAE of 18.7 days and ordered pairs 7–14 days apart 529
correctly in 84% of cases, against 88.5 days and 53% for the 530
image-only model. 531
C. Mosig et al.: Preprint submitted to Elsevier Page 8 of 13

AerialStamp: dating aerial orthophotos
J F M A M J J A S O N D
submitted date
J
F
M
A
M
J
J
A
S
O
N
D
predicted date
image-only model (MAE 17.8 d)
S2 model (MAE 13.3 d)
1 2 3 4 5 6
1
17 Feb
6 Feb
2
7 Apr
24 Apr
3
12 May
15 May
4
27 Jun
12 Jun
5
8 Sep
4 Sep
6
14 Nov
13 Nov
orthophotoSentinel-2
(a) Leipzig floodplain forest, 2025 (43 flights)
J F M A M J J A S O N D
submitted date
J
F
M
A
M
J
J
A
S
O
N
D
predicted date
image-only model (MAE 21.6 d)
S2 model (MAE 15.4 d)
1 2 3 4 5 6
1
20 Mar 2025
20 Mar
2
30 Apr 2024
16 May
3
26 May 2024
16 May
4
2 Jul 2025
3 Jul
5
3 Sep 2024
29 Aug
6
15 Oct 2024
24 Oct
orthophotoSentinel-2
(b) ECOSENSE, Freiburg, 2024–2026 (27 flights)
J F M A M J J A S O N D
submitted date
J
F
M
A
M
J
J
A
S
O
N
D
predicted date
image-only model (MAE 69.9 d)
S2 model (MAE 29.1 d)
1 2 3 4 5 6
1
24 Jan 2020
9 Jan
2
24 Feb 2022
24 Feb
3
25 Mar 2019
21 Mar
4
25 Apr 2023
13 Apr
5
24 Jul 2019
18 Jul
6
26 Oct 2020
12 Nov
orthophotoSentinel-2
(c) Barro Colorado Island 50 ha plot, 2018–2023 (90 flights)
Figure 4: Three sites with dense multitemporal drone data. Left: predicted against submitted date (out-of-fold; blue: S2 model,
orange: image-only model, drawn 3 days to the right of the S2 model; MAE over all flights of the site in brackets). Points show
the predicted date, lines the range between the 10th and 90th percentiles of the predicted distribution; predictions are unwrapped
to within half a year of the submitted date, so the axis extends beyond the calendar year. The dotted diagonal marks a perfect
prediction. Triangles on the x-axis mark the submitted dates of the six numbered flights on the right. Right: these flights over
the same 200 × 200m cut-out ( 20 × 20 Sentinel-2 pixels): the orthophoto with its submitted date above, and below the Sentinel-2
RGB composite (B04, B03, B02) nearest the submitted date with at least 95% clear pixels, labeled with its date.
C. Mosig et al.: Preprint submitted to Elsevier Page 9 of 13

AerialStamp: dating aerial orthophotos
5. Discussion532
5.1. Relevance of Sentinel-2 time series533
Most of the improvement from Sentinel-2 came from the534
correlation features, and little from the bands and indices535
of the Sentinel-2 series itself (Table 5). The Sentinel-2536
series of a site describes its seasons, but not which of them537
the orthophoto shows. The correlation features provide this538
information: the spatial arrangement of green and bare fields,539
of mown meadows, harvested crops or flooded areas changes540
from week to week, and the orthophoto matches only the541
weeks that share its arrangement. Because a correlation542
depends only on which pixels are brighter or greener than543
others, the features also work with the uncalibrated colors544
of consumer drone cameras (Section 3.2). The S2 model545
combines the correlation features of all weeks with the 10 cm546
patches, which show the seasonal state of the vegetation in547
detail. We hypothesize that the 10 cm patches carry pheno-548
logical detail of individual crowns, such as flowering, leaf549
color and leaf-off branches, that is averaged out in coarser550
views.551
The value of Sentinel-2 differed between biomes (Ta-552
ble 4). At temperate sites, the image-only model dated the553
evaluation set with a median error of 20.5 days (MAE 28.6554
days), and Sentinel-2 reduced it to 15.9 days (MAE 24.0555
days). At (sub) tropical sites, the median error dropped from556
36.5 to 22.5 days (MAE 50.8 to 37.4 days). The dense557
time series show the same pattern: in the Leipzig floodplain558
forest, the image-only model ordered the flights of a year559
almost perfectly (MAE 17.8 against 13.3 days with Sentinel-560
2), whereas on Barro Colorado Island it dated dry-season561
flights two to five months off (MAE 69.9 against 29.1 days;562
Section 4.3). Temperate deciduous forests follow a leaf cycle563
that is visible in the orthophoto itself. In tropical forests,564
leaf exchange is often asynchronous among species and the565
canopy changes little between seasons ( Park et al. , 2019;566
Albert et al. , 2019). Sentinel-2 is therefore most important567
for dating orthophotos in the tropics.568
5.2. Limitations569
Both models date the day within a submitted year and570
rely on that year: in the S2 model, a wrong year pairs the571
orthophoto with the wrong Sentinel-2 series and degrades572
the estimate to about the level of the image-only model573
(Appendix D). The year is partly recoverable from the data.574
In an exploratory analysis that is not part of the model, a575
small network ranked the ten candidate years 2016–2025 by576
the correlation features of each year and picked the submitted577
year for 50% of 6,335 orthophotos (73% within ±1 year),578
against 19% for the most common year; the other orthopho-579
tos were flown in 2026 or have fewer than two candidate580
years. Estimating year and day jointly would remove the581
dependence on the submitted year.582
The estimated numbers of wrong dates are lower bounds.583
A wrong date in a plausible season, a wrong day within the584
right season, or a date that matches the second mode of a585
bimodal distribution cannot be detected. Evergreen tropical586
forests, urban areas with little vegetation and dryland stands587
that remain unchanged through a long dry season are the 588
typical cases where the distribution stays broad or bimodal 589
(Fig. 3d). The manual check marked only five of the dates 590
used here as invalid, all of them on 1 January, too few 591
to measure how well the S2 model detects known errors. 592
Moreover, the 1st and 15th of the month are genuine flight 593
days for some orthophotos, such as the eight flights on these 594
days in the dense time series (Section4.3), so excluding them 595
from training discards correct labels. 596
5.3. Further applications 597
Linking drone or aerial imagery with satellite data is 598
a growing field ( Alvarez-Vanhard et al., 2021; Kattenborn 599
et al., 2021), and all of its applications depend on the acquisi- 600
tion date, which links the image to the satellite observations 601
of the same time. AerialStamp checks this link and supplies a 602
date where none is recorded. Because it returns a probability 603
for every day, the date uncertainty can be carried into the 604
validation of satellite products, which removes one source 605
of uncertainty that is otherwise unknown, in particular for 606
crowd-sourced reference data. 607
Aerial surveys of large areas are often mosaics of flights 608
in different months. Applied to tiles of such a mosaic, 609
AerialStamp could estimate the flight date of each region 610
and thus reconstruct the acquisition dates within the survey. 611
Since the season can be read from an orthophoto with an 612
error of a few weeks, the relation could also be used in 613
reverse. Generative models of remote sensing imagery are 614
conditioned on metadata such as location and time ( Khanna 615
et al. , 2023); a date model can test whether a generated 616
or augmented image matches its target season, or guide a 617
seasonal augmentation of training data. We have not tested 618
these applications. 619
6. Conclusion 620
We introduce AerialStamp, a method to estimate the 621
acquisition date of crowd-sourced aerial orthophotos as a 622
calibrated distribution over the days of the flight year, from 623
the orthophoto and the Sentinel-2 time series of its site. In 624
spatial block cross-validation on 6,225 publicly available 625
orthophotos, the S2 model dated orthophotos with a median 626
error of 18.5 days and calibrated prediction intervals. We 627
show that most of the improvement from Sentinel-2 comes 628
from simple correlation features, and that the 10 cm patches 629
add to it. Sentinel-2 time series reduced the median error 630
by 5 days at temperate and by 14 days at (sub) tropical 631
sites. In deadtrees.earth, at least 1.3% of the submitted dates 632
were detected to be wrong, and a further 10.3% fell on 633
likely placeholder days. AerialStamp is being integrated into 634
deadtrees.earth to check every upload and propose dates for 635
review where the submitted date is missing or contradicts 636
the data, and the AerialStamp package makes both models 637
available to any user. 638
C. Mosig et al.: Preprint submitted to Elsevier Page 10 of 13

AerialStamp: dating aerial orthophotos
Code and data availability639
AerialStamp, a Python package that dates a drone640
orthophoto with the S2 model or the image-only641
model and downloads the Sentinel-2 series with sentle642
(https://github.com/cmosig/sentle), is available at https:643
//github.com/cmosig/aerialstamp. The data of this study644
(the model inputs of all 6,984 orthophotos, the spatial645
folds and the out-of-fold predictions of both models), the646
evaluation protocol and the code to train and evaluate647
the models are available as a benchmark at https:648
//github.com/cmosig/aerialstamp-benchmark. Orthophotos649
and metadata are available through deadtrees.earth ( Mosig650
et al., 2026b), into which the model is being integrated.651
Declaration of competing interest652
The authors declare that they have no known competing653
financial interests or personal relationships that could have654
appeared to influence the work reported in this paper.655
Declaration of generative AI and AI-assisted656
technologies in the manuscript preparation657
process658
During the preparation of this work the authors used659
Claude (Anthropic) for data analysis code, drafting and660
language editing. After using this tool, the authors reviewed661
and edited the content as needed and take full responsibility662
for the content of the published article.663
Acknowledgements664
C.M., T.K., M.D.M. and J.V .J. thank the European Space665
Agency (ESA) for funding the “DeepFeatures” project via666
the AI4SCIENCE activity and the “FORTRACK” project667
via the ESA CLIMATE SPACE: Climate-Biodiversity stud-668
ies. T.K. acknowledges funding from the German Re-669
search Foundation via the “PANOPS” project (grant no.670
504978936) and EXC 3127 Future Forests. T.K. and J.V .J.671
acknowledge support by the Ministry for Food, Rural Areas,672
and Consumer Protection, Baden-Württemberg, Germany673
(grant no. 52-8670.00) and the German Research Foun-674
dation via the NFDI4Earth pilot project GeoLabel (DFG675
grant no. 460036893). D.M. and M.D.M. acknowledge sup-676
port from ESA (“cube.ng — No Gap Earth System Data677
Cubes”, CCN1 of ARCEME — Adaptation and Resilience678
to Climate Extremes and Multi-hazard Events, ESA Contract679
No. 4000144482/24/I-KE). We thank all contributors who680
shared their orthophotos via deadtrees.earth.681
A. DINOv3 vs. DINOv2682
Table A.1 compares the DINOv2 embeddings of the683
10 cm patches with three DINOv3 backbones (Siméoni et al.,684
2025): ViT-B/16 (base size, 16-pixel tokens) and ViT-L/16685
(large size) trained on web images (LVD-1689M), and ViT-686
L/16 trained on satellite imagery (SAT-493M). The patches,687
the embedding procedure (class token and mean patch token 688
at 224 px) and the training setup were identical. All DINOv3 689
backbones gave a higher error than DINOv2, by 1.7 to 2.7 690
days for the S2 model and by 2.3 to 3.5 days for the image- 691
only model; in all six comparisons, the 95% interval of the 692
difference excluded zero. 693
Table A.1
Block-weighted MAE (days) on the evaluation set with dif-
ferent patch embeddings (calibrated five-seed ensembles). In
brackets, the difference to DINOv2 with a 95% bootstrap
interval over spatial blocks.
Backbone S2 model Image-only model
DINOv2 ViT-B/14 33.2 41.2
DINOv3 ViT-B/16 34.8 ( +1.7; +0.5, +2.8) 43.5 ( +2.3; +0.9, +3.7)
DINOv3 ViT-L/16 35.8 ( +2.7; +1.3, +4.0) 44.7 ( +3.5; +1.9, +5.1)
DINOv3 ViT-L/16 SAT 35.0 ( +1.8; +0.5, +3.3) 43.6 ( +2.4; +0.7, +4.1)
B. Forward in time and window truncation 694
To test the S2 model forward in time, we trained it on 695
flights up to 2024 and evaluated it on the 1,550 orthophotos 696
flown in 2025–2026. Without window truncation, the S2 697
model was much less accurate at sites whose Sentinel-2 698
series ended inside the submitted year. Where the series 699
ended between July and October of the submitted year, its 700
error was 47.3 days, against 35.8 days for the same sites 701
in spatial cross-validation, where the training folds contain 702
other 2025–2026 flights whose series also end inside the 703
submitted year. Window truncation during training reduced 704
the error at these sites to 37.5 days, the overall forward-in- 705
time error from 47.1 to 38.9 days and the standard deviation 706
of the forward error across seeds from 2.8 to 0.6 days, and it 707
also reduced the error in spatial cross-validation from 34.1 708
to 33.1 days (Table B.1). Removing the year from the site 709
facts made no consistent difference. 710
Table B.1
Block-weighted MAE of the S2 model in spatial cross-
validation (all 6,225 evaluation orthophotos) and forward in
time (trained on flights up to 2024, tested on the 1,550
orthophotos of 2025–2026), and the standard deviation of
the forward MAE across three seeds (calibrated three-seed
ensembles). Truncation: window truncation; the recipe with
it is the S2 model. Errors in days.
MAE bw
Training recipe spatial forward seed s.d.
Without truncation 34.1 47.1 2.8
With truncation (S2 model) 33.1 38.9 0.6
With truncation, no year input 32.8 37.9 1.1
C. Calibration and further accuracy metrics 711
Table C.1 complements Table 3 with the shares of or- 712
thophotos dated within 7 and 15 days and with the metrics 713
of the predicted distribution (Section 3.7). 714
C. Mosig et al.: Preprint submitted to Elsevier Page 11 of 13

AerialStamp: dating aerial orthophotos
Table C.1
Further metrics for the methods of Table 3. ±7/15: share
of orthophotos dated within that many days; coverage: share
whose submitted date lies inside the 50/80/95% prediction
interval; width: mean width of the 80% interval in days; NLL:
negative log-likelihood of the submitted day. The baselines are
fitted by maximum likelihood and not calibrated further.
Method ±7 ±15 coverage width NLL
Most common date .10 .18 .51/.80/.94 203 5.71
Climatology .10 .21 .44/.74/.91 157 5.60
Phenology .08 .19 .49/.81/.95 197 5.65
Image-only model .16 .33 .50/.82/.96 123 5.14
S2 model .23 .44 .53/.80/.95 96 4.89
D. Shifted year715
To test how much the S2 model relies on the submitted716
year, we gave it a shifted year, i.e. the week tokens and the717
year in the site facts of the previous or the next year, for718
the 5,180 evaluation orthophotos with Sentinel-2 data in all719
three years. The error increased by 8 to 9 days, to about the720
level of the image-only model (Table D.1).721
Table D.1
Error of the S2 model with the submitted and with a shifted
year (week tokens and the year in the site facts of the previous
or next year; 𝑛 = 5 ,180 orthophotos with Sentinel-2 data in
all three years; uncalibrated five-seed ensemble, hence the low
coverage). Cov. 80%: coverage of the 80% prediction interval.
Errors in days.
Year MAE MAE bw cov. 80%
Submitted year (S2 model) 29.7 32.1 .62
Submitted year −1 37.4 40.8 .51
Submitted year +1 38.2 41.7 .50
References722
Albert, L.P., Restrepo-Coupe, N., Smith, M.N., Wu, J., Chavana-Bryant,723
C., Prohaska, N., et al., 2019. Cryptic phenology in plants: case studies,724
implications, and recommendations. Global Change Biology 25, 3591–725
3608. doi: 10.1111/gcb.14759.726
Alvarez-Vanhard, E., Corpetti, T., Houet, T., 2021. UAV & satellite727
synergies for optical remote sensing applications: A literature review.728
Science of Remote Sensing 3, 100019. doi: 10.1016/j.srs.2021.100019.729
Aybar, C., Ysuhuaylas, L., Loja, J., Gonzales, K., Herrera, F., Bautista, L.,730
et al., 2022. CloudSEN12, a global dataset for semantic understanding731
of cloud and cloud shadow in Sentinel-2. Scientific Data 9, 782.732
doi:10.1038/s41597-022-01878-2 .733
Berra, E.F., Gaulton, R., Barr, S., 2019. Assessing spring phenology of734
a temperate woodland: A multiscale comparison of ground, unmanned735
aerial vehicle and Landsat satellite observations. Remote Sensing of736
Environment 223, 229–242. doi: 10.1016/j.rse.2019.01.010.737
Darcet, T., Oquab, M., Mairal, J., Bojanowski, P., 2024. Vision trans-738
formers need registers, in: The Twelfth International Conference on739
Learning Representations (ICLR). URL: https://openreview.net/forum?740
id=2dnO3LLiJ1.741
Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V ., Gascon,742
F., et al., 2012. Sentinel-2: ESA’s optical high-resolution mission for743
GMES operational services. Remote Sensing of Environment 120, 25–744
36. doi: 10.1016/j.rse.2011.11.026.745
Fernández-Guisuraga, J., Sanz-Ablanedo, E., Suárez-Seoane, S., Calvo, L., 746
2018. Using unmanned aerial vehicles in postfire vegetation survey 747
campaigns through large and heterogeneous areas: Opportunities and 748
challenges. Sensors 18, 586. doi: 10.3390/s18020586. 749
GeoNadir, 2026. GeoNadir: a repository for drone mapping data. https: 750
//geonadir.com. Accessed 13 October 2026. 751
Gillespie, A.R., Kahle, A.B., Walker, R.E., 1987. Color enhancement of 752
highly correlated images. II. Channel ratio and “chromaticity” trans- 753
formation techniques. Remote Sensing of Environment 22, 343–365. 754
doi:10.1016/0034-4257(87)90088-5 . 755
Goodchild, M.F., 2007. Citizens as sensors: the world of volunteered 756
geography. GeoJournal 69, 211–221. doi: 10.1007/s10708-007-9111-y . 757
Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q., 2017. On calibration 758
of modern neural networks, in: Proceedings of the 34th International 759
Conference on Machine Learning, PMLR. pp. 1321–1330. URL: https: 760
//proceedings.mlr.press/v70/guo17a.html. 761
Hyndman, R.J., 1996. Computing and graphing highest density regions. 762
The American Statistician 50, 120–126. doi: 10.1080/00031305.1996. 763
10474359. 764
Jacobs, N., Roman, N., Pless, R., 2007. Consistent temporal variations 765
in many outdoor scenes, in: IEEE Conference on Computer Vision and 766
Pattern Recognition (CVPR), pp. 1–6. doi: 10.1109/cvpr.2007.383258. 767
Jammalamadaka, S.R., SenGupta, A., 2001. Topics in Circular Statistics. 768
World Scientific, Singapore. doi: 10.1142/4031. 769
Kakar, P., Sudha, N., 2012. Verifying temporal data in geotagged images 770
via sun azimuth estimation. IEEE Transactions on Information Forensics 771
and Security 7, 1029–1039. doi: 10.1109/tifs.2012.2188796. 772
Kattenborn, T., Leitloff, J., Schiefer, F., Hinz, S., 2021. Review on 773
convolutional neural networks (CNN) in vegetation remote sensing. 774
ISPRS Journal of Photogrammetry and Remote Sensing 173, 24–49. 775
doi:10.1016/j.isprsjprs.2020.12.010. 776
Kattenborn, T., Lopatin, J., Förster, M., Braun, A.C., Fassnacht, F.E., 2019. 777
UAV data as alternative to field sampling to map woody invasive species 778
based on combined Sentinel-1 and Sentinel-2 data. Remote Sensing of 779
Environment 227, 61–73. doi: 10.1016/j.rse.2019.03.025. 780
Kattenborn, T., Schiefer, F., Frey, J., Feilhauer, H., Mahecha, M.D., Dor- 781
mann, C.F., 2022. Spatially autocorrelated training and validation 782
samples inflate performance assessment of convolutional neural net- 783
works. ISPRS Open Journal of Photogrammetry and Remote Sensing 784
5, 100018. doi: 10.1016/j.ophoto.2022.100018. 785
Khanna, S., Liu, P., Zhou, L., Meng, C., Rombach, R., Burke, M., et al., 786
2023. DiffusionSat: A generative foundation model for satellite imagery. 787
arXiv preprint doi: 10.48550/arXiv.2312.03606. 788
Klosterman, S., Melaas, E., Wang, J.A., Martinez, A., Frederick, S., 789
O’Keefe, J., et al., 2018. Fine-scale perspectives on landscape phenology 790
from unmanned aerial vehicle (UAV) photography. Agricultural and 791
Forest Meteorology 248, 397–407. doi: 10.1016/j.agrformet.2017.10. 792
015. 793
Loshchilov, I., Hutter, F., 2019. Decoupled weight decay regularization, in: 794
International Conference on Learning Representations (ICLR). URL: 795
https://openreview.net/forum?id=Bkg6RiCqY7. 796
Mac Aodha, O., Cole, E., Perona, P., 2019. Presence-only geographical 797
priors for fine-grained image classification, in: 2019 IEEE/CVF In- 798
ternational Conference on Computer Vision (ICCV), pp. 9595–9605. 799
doi:10.1109/ICCV.2019.00969. 800
Mañas, O., Lacoste, A., Giró-i Nieto, X., Vazquez, D., Rodríguez, P., 2021. 801
Seasonal contrast: Unsupervised pre-training from uncurated remote 802
sensing data, in: Proceedings of the IEEE/CVF International Conference 803
on Computer Vision (ICCV), pp. 9394–9403. doi: 10.1109/iccv48922. 804
2021.00928. 805
Morellato, L.P.C., Alberti, L.F., Hudson, I.L., 2010. Applications of circular 806
statistics in plant phenology: a case studies approach, in: Hudson, I.L., 807
Keatley, M.R. (Eds.), Phenological Research. Springer, Dordrecht, pp. 808
339–359. doi: 10.1007/978-90-481-3335-2_16 . 809
Mosig, C., Kattenborn, T., Montero Loaiza, D., Vajna-Jehle, J., Brandt, 810
J., Jacobs, N., et al., 2026a. Sub-pixel mapping of annual tree and 811
standing deadwood cover from Sentinel-2 time series across the globe. 812
EarthArXiv preprint. URL: https://eartharxiv.org/repository/view/ 813
C. Mosig et al.: Preprint submitted to Elsevier Page 12 of 13

AerialStamp: dating aerial orthophotos
11912/, doi:10.31223/X5B18W.814
Mosig, C., Vajna-Jehle, J., Mahecha, M.D., Cheng, Y., Hartmann, H.,815
Montero, D., et al., 2026b. deadtrees.earth – an open-access and in-816
teractive database for centimeter-scale aerial imagery to uncover global817
tree mortality dynamics. Remote Sensing of Environment 332, 115027.818
doi:10.1016/j.rse.2025.115027.819
Müller, E., Springstein, M., Ewerth, R., 2017. “when was this picture820
taken?” – image date estimation in the wild, in: Advances in Information821
Retrieval (ECIR 2017), Springer, Cham. pp. 619–625. doi: 10.1007/822
978-3-319-56608-5_57 .823
Natural Earth, 2026. Natural Earth: free vector and raster map data at 1:10m,824
1:50m, and 1:110m scales. https://www.naturalearthdata.com. Accessed825
14 October 2026.826
Nowosad, J., Bonannella, C., Görgen, D., Jemeljanova, M., Kattenborn, T.,827
Linnenbrink, J., Meyer, H., Nussbaum, M., Patelli, L., Simoes, R., et al.,828
2026. Navigating challenges in spatial machine learning: Validation,829
uncertainty, algorithms, and reproducibility. ERDKUNDE .830
Olson, D.M., Dinerstein, E., Wikramanayake, E.D., Burgess, N.D., Powell,831
G.V .N., Underwood, E.C., et al., 2001. Terrestrial ecoregions of the832
world: A new map of life on earth. BioScience 51, 933–938. doi: 10.833
1641/0006-3568(2001)051[0933:TEOTWA]2.0.CO;2.834
OpenAerialMap, 2026. OpenAerialMap: the open collection of aerial835
imagery. https://openaerialmap.org. Accessed 13 October 2026.836
Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V ., Szafraniec, M., Khalidov,837
V ., et al., 2024. DINOv2: Learning robust visual features without838
supervision. Transactions on Machine Learning Research URL: https:839
//openreview.net/forum?id=a68SUt6zFt.840
Pabon-Moreno, D.E., Musavi, T., Migliavacca, M., Reichstein, M., Römer-841
mann, C., Mahecha, M.D., 2020. Ecosystem physio-phenology revealed842
using circular statistics. Biogeosciences 17, 3991–4006. doi: 10.5194/843
bg-17-3991-2020 .844
Padilha, R., Salem, T., Workman, S., Andalo, F.A., Rocha, A., Jacobs,845
N., 2022. Content-aware detection of temporal metadata manipulation.846
IEEE Transactions on Information Forensics and Security 17, 1316–847
1327. doi: 10.1109/tifs.2022.3159154.848
Palermo, F., Hays, J., Efros, A.A., 2012. Dating historical color images,849
in: Computer Vision – ECCV 2012, Springer, Berlin, Heidelberg. pp.850
499–512. doi: 10.1007/978-3-642-33783-3_36 .851
Park, J., Muller-Landau, H., Lichstein, J., Rifai, S., Dandois, J., Bohlman,852
S., 2019. Quantifying leaf phenology of individual trees and species in853
a tropical forest using unmanned aerial vehicle (UAV) images. Remote854
Sensing 11, 1534. doi: 10.3390/rs11131534.855
Ploton, P., Mortier, F., Réjou-Méchain, M., Barbier, N., Picard, N., Rossi,856
V ., et al., 2020. Spatial validation reveals poor predictive performance857
of large-scale ecological mapping models. Nature Communications 11,858
4540. doi: 10.1038/s41467-020-18321-y .859
Riihimäki, H., Luoto, M., Heiskanen, J., 2019. Estimating fractional cover860
of tundra vegetation at multiple scales using unmanned aerial systems861
and optical satellite data. Remote Sensing of Environment 224, 119–862
132. doi: 10.1016/j.rse.2019.01.030.863
Roberts, D.R., Bahn, V ., Ciuti, S., Boyce, M.S., Elith, J., Guillera-Arroita,864
G., et al., 2017. Cross-validation strategies for data with temporal,865
spatial, hierarchical, or phylogenetic structure. Ecography 40, 913–929.866
doi:10.1111/ecog.02881.867
Rouse, Jr., J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1974. Monitoring868
vegetation systems in the Great Plains with ERTS, in: Third Earth Re-869
sources Technology Satellite-1 Symposium, Volume 1: Technical Pre-870
sentations, Section A, NASA. URL: https://ntrs.nasa.gov/citations/871
19740022614.872
Salem, T., Workman, S., Jacobs, N., 2020. Learning a dynamic map of873
visual appearance, in: Proceedings of the IEEE/CVF Conference on874
Computer Vision and Pattern Recognition (CVPR), pp. 12432–12441.875
doi:10.1109/cvpr42600.2020.01245.876
Schiefer, F., Schmidtlein, S., Frick, A., Frey, J., Klinke, R., Zielewska-877
Büttner, K., Junttila, S., Uhl, A., Kattenborn, T., 2023. UAV-based ref-878
erence data for the prediction of fractional cover of standing deadwood879
from Sentinel time series. ISPRS Open Journal of Photogrammetry and880
Remote Sensing 8, 100034. doi: 10.1016/j.ophoto.2023.100034.881
Senaratne, H., Mobasheri, A., Ali, A.L., Capineri, C., Haklay, M.M., 2017. 882
A review of volunteered geographic information quality assessment 883
methods. International Journal of Geographical Information Science 31, 884
139–167. doi: 10.1080/13658816.2016.1189556. 885
Shatwell, D.G., Dave, I.R., Swetha, S., Shah, M., 2025. GT-Loc: Unify- 886
ing when and where in images through a joint embedding space, in: 887
Proceedings of the IEEE/CVF International Conference on Computer 888
Vision (ICCV), pp. 1–11. doi: 10.1109/iccv51701.2025.00008. 889
Siméoni, O., Vo, H.V ., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., 890
et al., 2025. DINOv3. arXiv preprint. doi: 10.48550/arXiv.2508.10104, 891
arXiv:2508.10104. 892
Smith, L.N., Topin, N., 2019. Super-convergence: very fast training of 893
neural networks using large learning rates, in: Pham, T. (Ed.), Artifi- 894
cial Intelligence and Machine Learning for Multi-Domain Operations 895
Applications, SPIE. doi: 10.1117/12.2520589. 896
Thapa, S., Garcia Millan, V .E., Eklundh, L., 2021. Assessing forest 897
phenology: A multi-scale comparison of near-surface (UAV, spectral re- 898
flectance sensor, PhenoCam) and satellite (MODIS, Sentinel-2) remote 899
sensing. Remote Sensing 13, 1597. doi: 10.3390/rs13081597. 900
Tucker, C.J., 1979. Red and photographic infrared linear combinations for 901
monitoring vegetation. Remote Sensing of Environment 8, 127–150. 902
doi:10.1016/0034-4257(79)90013-0 . 903
Vásquez, V ., García, M., Hernández, M., Muller-Landau, H.C., 2023. Barro 904
Colorado Island 50-ha plot aerial photogrammetry orthomosaics and 905
digital surface models for 2018–2023: globally and locally aligned time 906
series. Smithsonian Tropical Research Institute, Smithsonian Figshare, 907
dataset. doi: 10.25573/data.24782016. 908
Werner, C., Wallrabe, U., Christen, A., Comella, L., Dormann, C., Göritz, 909
A., et al., 2024. ECOSENSE – multi-scale quantification and mod- 910
elling of spatio-temporal dynamics of ecosystem processes by smart au- 911
tonomous sensor networks. Research Ideas and Outcomes 10, e129357. 912
doi:10.3897/rio.10.e129357. 913
Wilson, E.H., Sader, S.A., 2002. Detection of forest harvest type using 914
multiple dates of Landsat TM imagery. Remote Sensing of Environment 915
80, 385–396. doi: 10.1016/S0034-4257(01)00318-2 . 916
Woebbecke, D.M., Meyer, G.E., Von Bargen, K., Mortensen, D.A., 1995. 917
Color indices for weed identification under various soil, residue, and 918
lighting conditions. Transactions of the ASAE 38, 259–269. doi: 10. 919
13031/2013.27838. 920
Zhai, M., Salem, T., Greenwell, C.S., Workman, S., Pless, R., Jacobs, 921
N., 2018. Learning geo-temporal image features, in: Proceedings 922
of the British Machine Vision Conference (BMVC), Newcastle, UK. 923
arXiv:1909.07499. 924
Zhang, X., Liu, L., Liu, Y., Jayavelu, S., Wang, J., Moon, M., Henebry, 925
G.M., Friedl, M.A., Schaaf, C.B., 2018. Generation and evaluation of the 926
VIIRS land surface phenology product. Remote Sensing of Environment 927
216, 212–229. doi: 10.1016/j.rse.2018.06.047. 928
Zhu, Z., Woodcock, C.E., 2012. Object-based cloud and cloud shadow 929
detection in Landsat imagery. Remote Sensing of Environment 118, 930
83–94. doi: 10.1016/j.rse.2011.10.028. 931
C. Mosig et al.: Preprint submitted to Elsevier Page 13 of 13