[预印本]面向时序与地理偏移的作物分类地-星知识蒸馏

Ground-to-Satellite Knowledge Distillation for Crop
Classification under Temporal and Geographic Shift
Miro Miranda1,, Rocco Sedona1,, Valerio Marsocci2,
Gabriele Cavallaro1,3,Claudia Paris 4
1Jülich Supercomputing Centre, Forschungszentrum Jülich, Germany
2European Space Agency (ESA/ESRIN), Italy
3University of Iceland, Iceland
4University of Twente (ITC), the Netherlands
Abstract
Continuous and up-to-date crop type mapping is essential for agricultural monitor-
ing but is often hampered by limited reference data. The growing availability of
ground-level field photographs offers a promising way to address this challenge.
When combined with satellite time series, these images provide complementary
perspectives on the same parcel, capturing crop characteristics from distinctly
different perspectives. However, field photographs are not always available for
each satellite pixel at inference time, limiting their potential in scalable crop type
mapping. We investigate whether ground photographs can provide privileged su-
pervision for crop classification built on pretrained satellite embeddings. A photo
teacher based on frozen CLIP features supervises classifiers trained on AlphaEarth
and TESSERA representations; photographs are used only during training. On a
paired field photo and satellite imagery dataset with 17 crop classes, knowledge
distillation consistently improves over satellite-only classification; for frozen em-
beddings the margin grows under temporal shift, reaching up to +5.5 macro-F1
points, and widens in geographic holdouts (+3.7 vs +3.0 in-distribution). These
results show that ground-view supervision can complement satellite embeddings
and improve their robustness without changing inference requirements.
1 Introduction
Ground photographs offer a complementary perspective to satellite data and are becoming increasingly
available for crop monitoring through both crowdsourcing platforms [Garcin et al., 2021] and field
data collection protocols [Nakalembe et al., 2025, d’Andrimont et al., 2022]. For instance, the Land
Use/Cover Area Frame survey (LUCAS) pairs georeferenced survey points across Europe with ground
photographs, providing fine-grained crop characteristics such as leaf shape, structure, and texture that
are difficult to capture from satellite data alone [d’Andrimont et al., 2022]. Such information can
remain informative when satellite signature change across space or time and thus contributes to robust
crop type mapping [Abbas et al., 2025]. However, the drawback is obvious: ground photographs
are limited to surveyed locations and are rarely available on demand, which makes them impractical
when rapid, scalable inference is required. This is a learning-using-privileged-information setting
[Vapnik and Vashist, 2009], for which Knowledge Distillation (KD) is a practical implementation
[Hinton et al., 2015, Lopez-Paz et al., 2016]. We hypothesize that paired ground photographs provide
a complementary semantic view of crop type that can improve robustness when scaling across regions
and years [Wang et al., 2025]. From an operational perspective, crop type maps require frequent
*Equal contribution.
40th Conference on Neural Information Processing Systems (NeurIPS 2026). Workshop: 2nd Workshop on Advances in
Representation Learning for Earth Observation.
1
arXiv:2610.37021v1 [cs.CV] 27 Oct 2026

updating to account for the highly dynamic nature of agricultural systems.
Recent advances in geospatial foundation-models (GFMs) offer the potential to streamline data
preparation and substantially reduce Earth Observation (EO) data requirements by leveraging frozen
embeddings for continuous mapping. However, limited reference data remains a key constraint,
as classifiers trained on labels from specific geographic and temporal domains can suffer from
distribution shifts, leading to misclassification [Wang et al., 2025]. Such shifts arise because satellite
observations are influenced not only by crop characteristics but also by climate, phenology, land
management, and acquisition conditions [Wang et al., 2025], causing the same crop to exhibit different
satellite signatures across domains. Ground photographs provide a complementary view of the same
parcel that is less affected by these factors. Distilling the ground view into the classifier may therefore
anchor predictions to crop-relevant cues that are more consistent across geographic and temporal
domains. This is particularly relevant when classifiers are trained on frozen GFM embeddings, where
the downstream model has limited ability to compensate for region- or time-specific biases encoded
in the representation.
Building on the hypothesis that ground photographs provide complementary semantic information
about crops, we investigate whether this information can be transferred to satellite-based classifiers
to improve their robustness across geographic and temporal domains. Because ground photographs
are not consistently available across space and time, we use a ground photo teacher only during
training. Ground photographs are encoded with CLIP ViT-B/32 [Radford et al., 2021], pretrained on
approximately 400 million web-collected image–text pairs; a small photo teacher classifier is trained
per seed and then frozen. Satellite students learn from the labels together with the teacher’s logits and
embeddings. At inference, the student receives satellite data only. We keep the teacher and loss fixed
while changing the satellite branch, considering: (i) frozen embeddings from AlphaEarth [Brown
et al., 2025] and TESSERA [Feng et al., 2025], (ii) sequence transformers trained from scratch on
monthly satellite time series, and (iii) a partially fine-tuned TESSERA encoder.
To assess the effectiveness of the proposed multimodal framework, we make three controlled compar-
isons. First, under the same teacher and training budget, we compare the KD–Cross Entropy (CE)
difference across satellite models based on frozen and from-scratch representations. Second, we evalu-
ate these models under temporal transfer (2018→2022) and, for AlphaEarth, held-out ecoregions, i.e.,
geographically distinct ecological regions [Dinerstein et al., 2017]. Third, we test partial fine-tuning
of TESSERA while explicitly quantifying the mismatch introduced by our monthly-composite pro-
cessing of satellite data. Across the main frozen and from-scratch regimes, KD consistently improves
in-domain macro-F1 across the satellite branches, by up to 5.1 points on frozen TESSERA, and the
frozen-embedding margins grow further under temporal shift and in the Mediterranean geographic
holdout.
2 Method
Photo Teacher.A frozen CLIP ViT-B/32 encoder [Radford et al., 2021] maps ground photographs
to a feature vector v. Each views is individually L2-normed and averaged to create a single
512-dimensional representation over the location. For each seed, a small photo teacher head is
trained on the training split and then frozen: a two-layer projection to the shared 128-dimensional
embedding et(v) and a one-hidden-layer classifier producing the crop logits lt(et(v)). The CLIP
encoder itself remains frozen; the head is trained because frozen CLIP provides neither the shared
embedding nor crop-class logits, the two distillation targets. Teacher targets are cached only for the
student-training points.
Satellite Student.Each satellite student maps the input z to a shared dimensional embedding es(z)
and logitsl s(es(z)).
We use (a) a small MLP head on frozen GFM embeddings, mirroring the teacher: a two-layer
projection (width 256, LayerNorm, GELU) to es(z) and a one-hidden-layer classifier on concatenated
z and es(z); (b) a sequence transformer trained from scratch on monthly composites; and (c)
TESSERA v1.1 [Feng et al., 2025], unfreezing the last two transformer blocks per modality and the
fusion reducer. For (c), the encoder and head learning rates are 10−4 and 10−3, respectively; earlier
layers remain frozen. Following standard knowledge distillation [Hinton et al., 2015], with class label
y, temperatureT, and cross-entropy CE, the loss is:
2

Photo Teacher
CLIP ViT-B/32
AlphaEarth/ Tessera v1.1 /
S2/S1 Scratch
Patch (30 x 30 m)
Satellite Student
Knowledge Distillation
Field Photos
Satellite Data
(a) (b) (c) (d)
Figure 1: Dataset and model overview. (a) Spatial distribution of the selected agricultural LUCAS
2018 points, including the Alpine, Mediterranean, and Boreal held-out ecoregions. (b) Example of
the Sentinel-2 annual composite (S2), thein-situphotograph with the four cardinal views (E, W, N,
S) and a close-up photograph (C). (c) KD workflow including photo teacher and satellite distilled
student model. (d) Distribution of the 17 CLMS crop classes.
L= CE(l s, y) +λKLT 2 KL(σ(lt/T)∥σ(l s/T)) +λ cos (1−cos(e s, et)).(1)
The CE control uses an identical student, optimization, and class-weighted CE term, with both
distillation weights set to zero. Thus each reported KD–CE contrast changes only the privileged
supervision. More implementation details are provided in Appendix A.
3 Experimental setup
Data.To test the proposed approach, we use LUCAS 2018 survey points with ground photographs
[d’Andrimont et al., 2022] for training, focusing on the 17 classes of the Copernicus High Resolution
Layer Croplands (CLMS) products [Fahl et al., 2026]. The paired dataset contains 53,885 points
with a label, photo, and time series of satellite data. The canonical split follows 70/15/15 for
train/validation/test and is fixed across the main comparisons. The 2 × 2 km LUCAS grid ensures
spatially separated, non-overlapping observations, reducing potential spatial dependence between
splits. Satellite inputs are either released 64-D AlphaEarth [Brown et al., 2025] and TESSERA
[Feng et al., 2025] embeddings, or 12 monthly composites of Sentinel-2 (10 bands) and Sentinel-1
(VV/VH) over a 3×3 pixel patch (30m×30m) centered around the LUCAS point. Figure 1 provides
an overview of the spatial distribution of crop types (a), Sentinel-2 composites, thein-situphotograph,
and the crop type distribution (d).
Temporal transfer.Models trained on LUCAS 2018 are evaluated without adaptation on LUCAS
2022. The class mapping and normalization constants fitted on 2018 are reused, and geographical
locations belonging to the 2018 train or validation sets are excluded from the 2022 transfer set. The
resulting transfer set contains 36,428 points with a satellite feature and one of the 17 classes, at
locations unseen during training.
Geographic transfer.Geographic transfer is evaluated entirely within 2018, isolating spatial from
temporal domain shift. EEA ecoregions define fixed geographic test domains [European Environment
Agency, 2016]: Mediterranean, Alpine, and Boreal are held out in turn, while the remaining 2018
points are split into training and validation sets. CE and KD are paired on the same split. Our input
construction does not follow TESSERA’s inference pattern: irregular acquisitions are reduced to 12
monthly composites, missing months become mean-valued observations with a valid mid-month
day-of-year, and orbit-merged S1 uses averaged ascending/descending normalization statistics. A
linear probe scores 36.2±1.6 macro-F1 on embeddings from this frozen monthly-input path and
43.6±1.6 on the released TESSERA embeddings, a −7.4 point difference. We therefore treat
the released embedding and our monthly-input setting as separate regimes rather than using their
difference in macro-F1 as an estimate of the effect of fine-tuning. The reported main experiments use
five seeds. Macro-F1 is computed on the same evaluation set for each paired CE/KD run, using the
fixed 17-label taxonomy.
3

Table 1: Macro-F1 (%), mean over five seeds, for matched CE and KD students. All students
are trained on 2018 only; the transfer block evaluates them on the 2022 survey (unseen locations;
Appendix A). ∆ is the KD−CE mean of per-seed paired differences, computed before rounding. ∆T
is the difference to teacher performance (69.7 macro-F1).
2018 test 2018→2022 transfer
Satellite branch CE+KD∆ ∆T CE+KD∆
Frozen AlphaEarth 63.5 66.5+2.9−3.252.6 56.0+3.4
Frozen TESSERA 66.7 71.7+5.1 +2.057.6 63.1+5.5
S2 SITS (scratch) 62.0 65.2+3.2−4.554.0 56.4+2.4
S1+S2 SITS (scratch) 66.1 68.7+2.6−1.057.5 58.7+1.2
TESSERA fine-tuned 71.272.1+0.8 +2.463.063.6+0.7
Figure 2: Left: Overall paired KD−CE macro-F1 margins per branch from 2018→2022. Margins
increase for frozen models, decrease for scratch models, and remain below one point for fine-
tuned models. Right: AlphaEarth macro-F1 gains in held-out ecoregions. Only the Mediterranean
exceeds the in-distribution gain (+3.7 vs. +2.9); Alpine gains +1.1, while Boreal shows unstable
improvements (+0.5, SD1.4, mixed signs). Error bars indicate±1SD, paired for margins.
4 Results and Discussion
Table 1 shows a consistent in-domain improvement of 2.6–5.1 macro-F1 points for the frozen and
from-scratch students, largest on frozen TESSERA. Partial TESSERA fine-tuning (Sec. 2) gives the
best absolute result, 72.1 macro-F1, with up to +2.4 macro-F1 over the teacher, but the paired KD
margin is smaller at +0.8 macro-F1. Because this branch uses the monthly input path above, this
margin is not a controlled estimate of the effect of fine-tuning alone.
Transfer Experiments.Figure 2 (left) shows the paired KD −CE margin for every branch under
temporal transfer: the frozen margins grow (+2.9 to +3.4 macro-F1 points for AlphaEarth, +5.1 to
+5.5 for TESSERA), the scratch margins shrink (+3.2 to +2.4 and +2.6 to +1.2), and the partially
fine-tuned TESSERA branch stays below one point. Geographic holdout (right) widens the margin in
Mediterranean (+3.7 vs +3.0 in-distribution); Alpine is positive but smaller (+1.1), and Boreal is
unstable (+0.5, sd 1.4, mixed signs); Mediterranean is thus the only clear geographic gain.
Discussion.Ground-view KD helps most clearly when the satellite representation is fixed, and its
margin can grow under domain shift. Operationally, the photo teacher is needed only during training;
deployment requires a satellite-only head on the same embedding product. KD also outperforms
alternative cross-modal objectives such as contrastive learning (Appendix C). The comparison has
limits. First, TESSERA fine-tuning uses the dataset’s monthly composites, for input consistency
with the scratch transformers, whereas the frozen TESSERA model uses released embeddings
produced through a different input path, a mismatch the −7.4 probe difference quantifies. Second,
the geographic results are clear in the Mediterranean holdout, Boreal shows how quickly a small
holdout can become noisy. Lastly, the transfer performance depend on the class itself, especially for
the frozen TESSERA (Appendix D).
4

5 Conclusion
Ground photographs can provide privileged supervision for satellite crop classification without
being required at inference. Their benefit is strongest for fixed pretrained satellite representations
and persists, or increases, under temporal and geographic shift. Our ablations show that this gain
is driven primarily by distillation of the teacher’s predictive structure rather than by cross-modal
feature alignment. The gain is consistent across all satellite branches with frozen-embedding margins
grow further under temporal shift and in the Mediterranean holdout. This knowledge-distillation
framework is particularly powerful from an operational perspective: field observations can be used to
improve satellite models without imposing field-data requirements on large-scale deployment. This is
especially relevant for scaling agricultural mapping and monitoring, where satellite imagery enables
broad and repeated coverage but ground observations are costly to collect.
Acknowledgements
This work was supported by the European Space Agency (ESA) Φ-Lab as part of the FAST-EO
project, under Contract No. 4000143501/23/I-DT and the the AgroVision project, under contract No.
4000147606. The authors gratefully acknowledge the computing time granted on the supercomputer
JUWELS Booster at Jülich Supercomputing Centre (JSC).
References
Adel Abbas, Claudia Paris, and Andrew Nelson. Multimodal satellite-field data integration for
scalable crop-type mapping with limited labels.IEEE Transactions on Geoscience and Remote
Sensing, 63:1–12, 2025.
Christopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, William J. Rucklidge, Masha
Samsikova, Chenhui Zhang, Evan Shelhamer, Estefania Lahera, Olivia Wiles, Simon Ilyushchenko,
Noel Gorelick, Lihui Lydia Zhang, Sophia Alj, Emily Schechter, Sean Askay, Oliver Guinan,
Rebecca Moore, Alexis Boukouvalas, and Pushmeet Kohli. AlphaEarth Foundations: An embed-
ding field model for accurate and efficient global mapping from sparse label data.arXiv preprint
arXiv:2507.22291, 2025.
Raphaël d’Andrimont, Momchil Yordanov, Laura Martinez-Sanchez, Peter Haub, Oliver Buck,
Carsten Haub, Beatrice Eiselt, and Marijn van der Velde. LUCAS cover photos 2006–2018 over
the EU: 874,646 spatially distributed geo-tagged close-up photos with land cover and plant species
label.Earth System Science Data, 14:4463–4477, 2022. doi: 10.5194/essd-14-4463-2022.
Eric Dinerstein, David Olson, Anup Joshi, Carly Vynne, Neil D Burgess, Eric Wikramanayake,
Nathan Hahn, Suzanne Palminteri, Prashant Hedao, Reed Noss, et al. An ecoregion-based
approach to protecting half the terrestrial realm.BioScience, 67(6):534–545, 2017.
European Environment Agency. Biogeographical regions, Europe 2016. Dataset,
version 1, 2016. https://www.eea.europa.eu/data-and-maps/data/
biogeographical-regions-europe-3.
Fernando Fahl, Marijn van der Velde, Martin Claverie, Momchil Yordanov, Laura Martinez Sanchez,
and Emanuele Lugato. Europe wide agricultural land use systems by harmonizing the copernicus
high resolution layers on grassland and croplands.Scientific Data, 2026.
Zhengpeng Feng, Clement Atzberger, Sadiq Jaffer, Jovana Knezevic, Silja Sormunen, Robin Young,
Madeline C. Lisaius, Markus Immitzer, David A. Coomes, Anil Madhavapeddy, Andrew Blake, and
Srinivasan Keshav. TESSERA: Temporal embeddings of surface spectra for earth representation
and analysis.arXiv preprint arXiv:2506.20380, 2025.
Camille Garcin, alexis joly, Pierre Bonnet, Antoine Affouard, Jean-Christophe Lombardo, Mathias
Chouet, Maximilien Servajean, Titouan Lorieul, and Joseph Salmon. Pl@ntnet-300k: a plant
image dataset with high label ambiguity and a long-tailed distribution. InThirty-fifth Conference
on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. URL
https://openreview.net/forum?id=eLYinD0TtIt.
5

Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv
preprint arXiv:1503.02531, 2015.
Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron
Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. InAdvances in Neural
Information Processing Systems (NeurIPS), volume 33, pages 18661–18673, 2020.
David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. Unifying distillation and
privileged information. InInternational Conference on Learning Representations, 2016.
Catherine Nakalembe, Ivan Zvonkov, Hannah Kerner, Diana Botchway Frimpong, Kenneth Mwangi,
Jane Kioko, Bhanu Tokas, Kartik Jawanjal, Iman Akhtar Smith, Anjali Paliyam, et al. Helmets
labeling crops: Kenya crop type dataset created via helmet-mounted cameras and deep learning.
Scientific Data, 12(1):1496, 2025.
Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,
Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever.
Learning transferable visual models from natural language supervision. InInternational Conference
on Machine Learning, 2021.
Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive
coding.arXiv preprint arXiv:1807.03748, 2018.
Vladimir Vapnik and Akshay Vashist. A new learning paradigm: Learning using privileged informa-
tion.Neural Networks, 22(5–6):544–557, 2009.
Hengbin Wang, Yu Yao, Junyi Liu, Xindan Zhang, Yuanyuan Zhao, Shaoming Li, Zhe Liu, Xiaodong
Zhang, and Yelu Zeng. Unsupervised cross-regional and cross-year adaptation by climate indicator
discrepancy for crop classification.Journal of Remote Sensing, 5:0439, 2025.
A Additional experimental details
Sequence Transformer.The time series of satellite images is encoded using a lightweight sequence
transformer model. Monthly spectral features are linearly projected into a shared latent space. The
sequence is processed by pre-normalized Transformer blocks with multi-head self-attention, feed-
forward layers, residual connections, and dropout. Mean pooling across all months, followed by
LayerNorm, produces a single sequence representation.
Photo TeacherPrecomputed photo features are mapped into a shared latent space using a two-layer
projection head with LayerNorm and GELU activation. The resulting embedding is passed to an
MLP classifier for crop type prediction. The trained photo teacher is frozen during student training
and provides both embedding- and logit-level supervision.
Satellite student.Precomputed satellite features are transformed into the teacher’s latent space
using an equivalent two-layer projection head. The projected embedding is concatenated with the
original satellite feature vector and passed to an MLP classifier. Training combines supervised
cross-entropy with logit- and embedding-level KD from the frozen photo teacher.
Main KD configuration.The main KD runs use temperature T= 2 , logit-KD weight 1.0,
embedding-cosine weight 0.25, and the same class-weighted CE term as the matched CE control.
The student architecture, optimizer, split, and seed are paired within each KD–CE comparison.
Transfer controls.The 2022 evaluation reuses the 2018 class map and scaler and excludes points
used for 2018 training or validation. In the geographic experiments the held-out ecoregion is fixed as
test data; only the train/validation partition of the remaining points changes with seed.
TESSERA comparison.The frozen released-embedding experiment and the partially fine-tuned
monthly-input experiment are deliberately reported as different regimes. A direct test of the effect of
fine-tuning would require both CE and KD with the encoder frozen on the same monthly inputs.
6

Figure 3: Temporal representation stability across five training seeds, computed on the 1,046 label-
stable locations observed in both surveys. Squares: relative reduction from CE to KD, ∆site =
100(DCE −D KD)/DCE, where D is the matched-location dissimilarity of Eq. (2), i.e. the cosine
dissimilarity between the 2018 and 2022 representations of the same location, averaged within each
crop class and then across classes. Diamonds: relative reduction ∆ratio = 100(RCE −R KD)/RCE
of the matched-to-nonmatched ratio R of Eq. (3), i.e. the same matched-location dissimilarity divided
by the mean cross-year dissimilarity between a 2018 location and the other 2022 locations of its crop,
so that a reduction inRindicates location-specific stability beyond increased class compactness.
B Temporal representation stability
We further examined whether ground-view distillation changes the temporal stability of the learned
satellite representation. This analysis was performed on the partially fine-tuned TESSERA student,
using the encoder features before the projection and classification heads. We matched the 2018 and
2022 LUCAS surveys by point identifier and retained only locations observed in both years whose
crop label was unchanged. This resulted in 1,046 repeated locations from 10 crop classes; locations
whose crop label changed between surveys were excluded because a change in representation is
expected in that case.
For a retained locationi, we quantify the change between years by the cosine dissimilarity
dcos

h2018i ,h 2022i
= 1− h2018i ·h 2022i∥h2018i ∥2 ∥h2022i ∥2 ,(2)
where hyi denotes the student representation of location i in year y. We first average this quantity
within each crop class and then macro-average over classes, such that frequent crops do not dominate
the measurement. As shown in Fig. 3, KD reduced the matched-location cross-year dissimilarity by
12.2% on average across the five training seeds, with a reduction for every seed.
Because lower matched-location dissimilarity can also arise from increased within-class compactness,
this quantity alone does not isolate a location-specific temporal effect. We therefore used a same-class
control. For each crop class c, the matched dissimilarity was compared with the mean cross-year
dissimilarity between a 2018 location and allotherretained 2022 locations of the same crop:
Rc =
1
|Ic|
P
i∈Ic dcos

h2018i ,h 2022i

1
|Ic|
P
i∈Ic
1
|Ic|−1
P
j∈Ic
j̸=i
dcos

h2018i ,h 2022j
,(3)
where Ic is the set of repeated, label-stable locations belonging to class c. The ratios Rc are again
macro-averaged across classes. KD reduced this matched-to-nonmatched ratio by 2.6% on average,
with lower values in four of five seeds. Thus, most of the reduction in cross-year dissimilarity
co-occurs with increased intra-class compactness, while a smaller additional same-location stability
effect remains. We therefore regard the result as consistent with the semantic-anchoring interpretation
of ground-view distillation with strong regularization by the photo teacher’s predictive structure.
7

Table 2: Macro F1 and Weighted F1 for the frozen AlphaEarth CE baseline, contrastive objectives
(InfoNCE, SupCon), photo probe, and fusion, mean ± std over the same five seeds, split, and
evaluation protocol as Table 1. Photographs are used only during training for the satellite-only KD
and contrastive models; the photo probe and fusion model require photographs at inference. The first
two rows repeat the matched CE and KD students of Table 1 as anchors. The CE control row is the
in-harness baseline of the contrastive comparison (identical backbone, data, and optimizer as the CL
rows); its head differs architecturally from the student, which accounts for the small offset between
the two CE rows.
Model MacroF1% WeightedF1%
Frozen AlphaEarth CE student [z] 63.5±0.6 67.0±0.4

• KD [z] 66.5±0.3 68.6±0.4
  CE control [z] 63.8±0.7 67.4±0.5
• CLIP (InfoNCE) 63.7±0.7 67.2±0.6
• SupCon 63.4±0.4 67.1±0.4
  Photo probe [CLIP, v] 69.7±0.4 69.2±0.3
  Fusion [z+v] 77.7±0.3 77.0±0.3
  C Contrastive Learning & Modality Fusion
  We further tested alternative cross-modal objectives, such as contrastive learning (CL), including a
  supervised contrastive (SupCon) [Khosla et al., 2020] and CLIP-style instance contrastive (InfoNCE)
  [van den Oord et al., 2018, Radford et al., 2021]. SupCon adds a supervised contrastive loss across
  modalities with positives defined by shared crop class. In contrast, InfoNCE adds an instance-
  contrastive loss with positives defined by the same site, aligning each satellite point with its own
  photo against the rest of the batch. We tested the CL approaches on the frozen AlphaEarth embeddings,
  coupled with the CLIP photo embeddings and compared against the AlphaEarth and photo probe.
  We further include the performance for combining both modalities. For the fusion model both
  modalities are concatenated at the input level and fed into the classifier. Results for 2018 are shown in
  Table 2. Importantly, the CL objectives do not improve over the matched CE control (per-seed paired
  differences are small and mixed in sign), in contrast to the KD objective repeated in the table (+2.9
  macro-F1 paired over the CE student). The same dissociation holds within the distillation harness:
  adding only the logit-KL term to CE reproduces the full gain ( +2.9 macro-F1, positive on every
  seed), whereas adding only the embedding-cosine term is null (+0.2 macro-F1, mixed signs). The
  CL objectives shape embeddings rather than predictive distributions, so their null result is consistent
  with the cosine-only control. In contrast, the macro-F1 of the photo probe is consistently about 6 pp
  higher than the CL objectives and both CE rows, which use the AlphaEarth embeddings only. The
  combination of both modalities provides consistently the best results with approximately 8 pp higher
  macro-F1 than the photo probe model. This is expected, since this model receives complementary
  information that boosts crop classification performance but it cannot be used for inference in locations
  where no field photos are available.
  D Per-class analysis
  We further show the per-class results of the marcro-F1 score on the 2018 test set in Table 3 and for
  the temporal transfer in Table 4. Notably, KD improves crop classification for most classes, except
  individual cases such as rice (89.0 to 88.9) for the frozen AlphaEarth embeddings. Moreover, paired
  KD−CE macro-F1 margins per crop class for the temporal transfer is shown in Figure 4, evaluated at

2022. Notably, performance gains depends on crop type and model branch, with +38.9 macro-F1
      points for Flax/Hemp on the frozen TESSERA model and −1.6 on Rice for the TESSERA fine-tuned
      branch. To demonstrate that the reported KD gain is not only driven by single class results, we show
      the macro-F1 score for a held-one-class-out validation in Figure 5 with stable gains, except for frozen
      TESSERA. For frozen TESSERA under temporal transfer, excluding the combined flax, cotton, and
      hemp class reduces the gain from 5.5 to 3.4 percentage points.
      8

Table 3: Per-class F1 (%) on the 2018 test set, mean± standard deviation over five seeds, for matched
CE and KD students. Classes are sorted by 2018 test support.
Frozen AlphaEarth Frozen TESSERA S2 SITS (scratch) S1+S2 SITS (scratch) TESSERA fine-tuned
Class CE +KD CE +KD CE +KD CE +KD CE +KD
Wheat 70.0±0.9 71.3±0.9 74.0±0.2 76.0±0.7 71.4±1.3 73.7±0.8 73.4±2.2 76.3±1.0 77.2±1.1 78.3±0.6
Maize 87.8±0.4 89.0±0.2 89.0±0.3 89.5±0.2 88.6±0.8 90.0±0.3 89.9±0.4 91.1±0.3 91.4±0.4 91.3±0.4
Barley 55.6±1.1 55.6±0.9 64.3±1.0 63.6±0.6 57.4±1.3 57.9±1.2 59.6±0.6 60.8±0.8 65.5±0.8 64.8±0.9
Other cereals 46.6±0.5 46.6±0.4 51.4±0.9 49.8±1.2 44.1±0.8 43.3±1.1 48.6±0.6 49.3±0.9 53.5±1.3 53.6±1.0
Rapeseed 77.4±1.6 79.4±0.6 80.6±1.0 83.3±0.3 79.0±0.1 79.8±0.5 82.1±0.9 84.0±0.5 82.1±0.5 83.4±0.4
Sunflower 74.5±1.3 77.4±1.0 83.2±1.5 86.2±0.5 79.0±1.3 80.0±2.1 85.2±1.0 86.1±0.7 86.5±1.1 87.2±1.3
Fruits 60.8±1.0 64.9±1.3 61.4±1.2 64.0±1.3 53.1±1.7 58.4±2.7 55.7±3.7 60.7±1.3 60.6±0.9 63.6±1.8
Dry pulses 38.3±1.5 42.6±1.3 59.1±3.3 64.0±0.8 44.2±1.0 48.7±2.5 54.3±2.6 58.3±1.5 62.2±1.0 62.6±2.8
Sugar Beet 81.8±1.1 84.5±1.6 84.2±0.9 86.3±0.7 83.6±1.2 86.4±1.6 86.9±2.0 87.6±1.4 87.7±0.7 88.0±0.8
Olives 71.3±1.9 74.3±0.4 74.4±1.7 76.0±1.3 67.5±1.4 69.7±1.4 69.0±1.2 69.1±2.1 73.4±1.9 74.0±1.0
Potatoes 60.3±1.7 67.2±1.6 68.0±1.9 71.8±1.3 62.2±1.3 66.6±1.8 69.8±2.5 72.6±1.1 75.6±2.3 78.0±1.3
Grapes 60.3±1.0 61.1±1.1 65.7±1.4 67.1±2.0 52.0±3.4 55.7±0.9 56.5±0.8 57.1±0.8 62.0±2.2 62.7±2.6
Fresh vegetables 30.6±2.0 34.4±1.6 39.3±1.9 44.5±0.5 30.6±3.6 30.9±3.1 33.7±2.7 34.7±2.9 39.1±5.2 39.8±4.0
Nuts 41.3±3.8 42.9±2.1 39.3±2.2 44.2±4.0 33.9±3.3 40.3±3.0 38.3±1.0 41.7±1.5 41.5±3.9 41.5±2.7
Soybeans 56.5±3.7 69.1±3.3 76.4±1.7 82.5±0.6 59.0±3.8 70.2±4.2 64.3±2.7 71.9±2.4 76.7±1.2 78.4±2.1
Flax, cotton and hemp 77.8±1.2 80.9±2.0 35.8±0.5 80.7±2.6 70.4±4.6 73.4±1.2 73.2±4.9 77.2±2.6 85.8±1.5 84.5±2.4
Rice 89.0±3.8 88.9±2.2 87.4±2.9 90.1±2.5 77.9±5.6 82.7±4.3 83.2±4.6 90.1±4.2 90.4±1.8 93.1±1.9
Macro F1 63.5±0.6 66.5±0.3 66.7±0.2 71.7±0.2 62.0±1.1 65.2±1.0 66.1±1.0 68.7±0.6 71.2±0.7 72.1±0.5
Table 4: Per-class F1 (%) under 2018→2022 temporal transfer, mean ± standard deviation over five
seeds: the same trained students evaluated on the 2022 survey (unseen locations, train-year scaler and
class map). Classes are sorted by 2018 test support.
Frozen AlphaEarth Frozen TESSERA S2 SITS (scratch) S1+S2 SITS (scratch) TESSERA fine-tuned
Class CE +KD CE +KD CE +KD CE +KD CE +KD
Wheat 64.2±0.4 65.8±0.9 65.9±1.2 70.8±1.1 64.0±1.4 66.4±0.6 66.6±2.5 69.7±0.9 71.2±1.5 72.3±0.5
Maize 78.5±0.9 83.3±0.2 86.3±0.4 88.3±0.2 83.8±0.7 85.7±0.9 85.8±0.6 87.2±0.2 88.1±0.3 88.6±0.3
Barley 48.0±1.9 48.3±0.2 55.5±0.5 57.3±1.1 49.5±1.0 50.1±0.3 52.3±0.8 53.1±0.8 57.7±1.4 58.5±0.4
Other cereals 42.2±1.5 42.7±1.9 47.6±1.6 48.3±1.5 41.9±1.7 41.9±1.8 46.0±1.1 46.2±0.5 50.3±1.4 49.0±1.9
Rapeseed 68.4±1.1 69.6±0.4 71.8±0.9 73.5±0.5 67.8±1.2 70.1±0.8 73.3±0.2 74.2±0.1 72.7±0.5 74.3±0.5
Sunflower 45.7±2.2 51.0±3.5 73.2±1.4 74.2±1.1 65.7±1.3 68.4±1.6 73.5±0.9 75.1±1.5 75.8±1.2 76.2±2.5
Fruits 59.9±1.1 62.4±1.1 53.9±1.6 57.4±1.0 54.3±0.6 55.8±1.6 54.2±1.9 56.8±1.3 58.4±0.8 59.8±0.8
Dry pulses 20.8±2.3 24.9±2.2 38.3±3.3 45.0±1.1 29.2±1.5 33.7±1.1 37.3±2.7 40.1±2.0 41.8±0.8 42.3±2.4
Sugar Beet 64.6±5.2 72.1±1.5 64.9±5.5 73.6±4.5 76.2±2.1 80.1±2.3 79.1±3.2 81.6±1.4 78.7±1.3 80.8±2.2
Olives 70.0±1.7 70.3±1.4 67.4±1.5 69.8±1.8 63.6±1.8 66.4±1.0 64.0±1.6 65.8±0.7 68.4±0.7 70.2±1.1
Potatoes 41.0±2.0 48.1±2.5 56.9±2.8 63.5±2.1 51.9±1.5 55.0±1.9 57.5±1.7 57.8±1.5 63.7±1.6 64.4±1.1
Grapes 54.7±3.0 54.8±5.0 49.6±4.7 52.3±3.0 46.8±3.1 46.6±3.9 53.3±1.9 51.8±2.4 55.7±3.5 55.2±2.6
Fresh vegetables 22.7±1.6 28.5±1.5 30.2±1.6 35.3±1.7 28.3±1.0 29.4±3.2 29.1±2.0 30.7±1.1 34.3±1.7 36.1±1.2
Nuts 32.1±2.8 31.6±1.6 29.5±3.4 32.9±1.7 25.2±1.6 28.0±0.9 28.9±1.4 29.2±1.1 30.1±0.9 30.2±0.8
Soybeans 34.3±2.6 42.0±1.8 62.3±3.4 64.9±1.8 41.2±1.5 45.6±3.1 45.2±1.7 45.5±4.2 60.2±1.9 58.8±1.7
Flax, cotton and hemp 65.6±3.8 70.7±3.1 39.8±0.6 78.7±2.3 58.6±3.6 61.9±2.9 56.7±1.8 58.4±2.8 74.9±2.6 78.4±1.3
Rice 81.7±3.1 86.5±2.1 85.9±1.2 86.8±1.6 70.3±4.9 73.7±3.1 75.3±3.9 75.0±3.0 88.1±1.5 86.5±1.6
Macro F1 52.6±1.2 56.0±0.6 57.6±0.7 63.1±0.6 54.0±1.0 56.4±0.7 57.5±0.7 58.7±0.1 63.0±0.6 63.6±0.4
AlphaEarth
TESSERA frozen
S2 scratch
S1+S2 scratch TESSERA-FT
Wheat (N=9605)
Maize (N=6402)
Barley (N=4138)
Other cereals (N=3899)
Rapeseed (N=2781)
Fruits (N=1707)
Sunflower (N=1460)
Olives (N=1284)
Grapes (N=1069)
Potatoes (N=826)
Sugar beet (N=771)
Nuts (N=598)
Dry pulses (N=591)
Fresh veg (N=556)
Soybeans (N=419)
Flax/hemp (N=111)
Rice (N=63)
Crop type, ordered by 2022 support
+1.7 +4.9 +2.4 +3.1 +1.1
+4.8 +2.0 +1.9 +1.3 +0.5
+0.3 +1.8 +0.6 +0.8 +0.8
+0.6 +0.7 +0.0 +0.3 -1.3
+1.2 +1.7 +2.4 +1.0 +1.6
+2.4 +3.5 +1.5 +2.6 +1.4
+5.3 +1.1 +2.7 +1.6 +0.4
+0.3 +2.4 +2.9 +1.8 +1.8
+0.0 +2.7 -0.2 -1.5 -0.5
+7.1 +6.6 +3.1 +0.3 +0.7
+7.5 +8.7 +3.9 +2.5 +2.1
-0.4 +3.4 +2.8 +0.3 +0.1
+4.1 +6.7 +4.4 +2.7 +0.5
+5.8 +5.1 +1.1 +1.6 +1.9
+7.7 +2.6 +4.4 +0.3 -1.5
+5.2 +38.9* +3.4 +1.7 +3.5
+4.8 +0.9 +3.4 -0.3 -1.6
−10.0
−7.5
−5.0
−2.5
0.0
2.5
5.0
7.5
10.0
KD − CE per-class F1 (percentage points)
Per-class KD performance gain under temporal transfer (2018 → 2022)
Cell values: mean F1(KD) − F1(CE) over 5 paired seeds; full aligned 2022 transfer population
Exact values are printed; * indicates color clipped at ±10 pp. N is mean class support across branches and seeds.
cereals
vegetables
pulses
root crops
oilseed and fibre crops
permanent crops
Figure 4: Heatmap for paired KD −CE macro-F1 margins per crop class for the temporal transfer
(2022). Colorbar is cropped to 10.0 macro-F1 points.
9

Frozen AlphaEarth
Frozen TESSERA
S2 SITS (scratch)
S1+S2 SITS (scratch)
TESSERA fine-tuned
+3.1 +3.1 +3.1 +3.1 +3.0 +3.0 +2.9 +2.9 +3.0 +3.0 +2.7 +3.1 +2.9 +3.0 +2.4 +2.9 +3.1
+5.3 +5.3 +5.4 +5.5 +5.2 +5.2 +5.2 +5.1 +5.2 +5.3 +5.1 +5.3 +5.1 +5.1 +5.0 +2.6 +5.2
+3.2 +3.3 +3.3 +3.4 +3.3 +3.3 +3.0 +3.1 +3.2 +3.2 +3.1 +3.1 +3.3 +3.0 +2.7 +3.2 +3.1
+2.6 +2.7 +2.7 +2.8 +2.7 +2.7 +2.5 +2.6 +2.8 +2.8 +2.6 +2.8 +2.7 +2.6 +2.3 +2.6 +2.4
+0.8 +0.9 +0.9 +0.8 +0.8 +0.8 +0.7 +0.8 +0.8 +0.8 +0.7 +0.8 +0.8 +0.8 +0.7 +0.9 +0.7
(a) 2018 test
66.5
(+3.0)
71.7
(+5.1)
65.2
(+3.2)
68.7
(+2.6)
72.0
(+0.8)
Overall
+KD F1
Wheat (29.5%)Maize (15.9%)Barley (14.4%)
Other cereals (10.0%)
Rapeseed (7.6%)Sun
flower (3.7%)Fruits (3.3%)
Dry pulses (2.5%)Sugar beet (2.4%)
Olives (2.3%)
Potatoes (2.2%)Grapes (1.8%)
Fresh vegetables (1.5%)
Nuts (1.2%)
Soybeans (0.9%)
Flax / cotton / hemp (0.6%)
Rice (0.3%)
Excluded crop class (ordered by 2018 support)
Frozen AlphaEarth
Frozen TESSERA
S2 SITS (scratch)
S1+S2 SITS (scratch)
TESSERA fine-tuned
+3.5 +3.3 +3.6 +3.6 +3.6 +3.3 +3.5 +3.4 +3.2 +3.6 +3.2 +3.6 +3.3 +3.7 +3.2 +3.3 +3.3
+5.5 +5.7 +5.7 +5.8 +5.7 +5.8 +5.6 +5.4 +5.3 +5.7 +5.4 +5.7 +5.5 +5.6 +5.7 +3.4 +5.8
+2.4 +2.4 +2.5 +2.5 +2.4 +2.4 +2.4 +2.3 +2.3 +2.4 +2.3 +2.5 +2.5 +2.4 +2.3 +2.3 +2.3
+1.1 +1.2 +1.2 +1.2 +1.2 +1.2 +1.1 +1.1 +1.1 +1.1 +1.2 +1.4 +1.2 +1.2 +1.2 +1.2 +1.3
+0.6 +0.7 +0.7 +0.8 +0.6 +0.7 +0.6 +0.7 +0.6 +0.6 +0.7 +0.7 +0.6 +0.7 +0.8 +0.5 +0.8
(b) 2022 transfer
56.0
(+3.4)
63.1
(+5.5)
56.4
(+2.4)
58.7
(+1.2)
63.6
(+0.7)
Overall
+KD F1
0
1
2
3
4
5
Leave-one-class-out KD − CE macro-F1 (pp)
Leave-one-class-out sensitivity of KD gainsHeatmap: KD − CE macro-F1 after excluding one class from the macro average (no retraining). Overall: all-class +KD F1; parentheses: gain.
Orange outline: Flax/cotton/hemp excluded.
Figure 5: Heatmap for paired macro-F1 KD−CE margins fore a held-one-class-out validation over the
model branches and yearsRre. Constantly stable performance gains, except for the frozen TESSERA.
10