Ground-to-Satellite Knowledge Distillation for Crop Classification under Temporal and Geographic Shift

Miro Miranda; Rocco Sedona; Valerio Marsocci; Gabriele Cavallaro; Claudia Paris

cs.CVOct 27, 2026

Ground-to-Satellite Knowledge Distillation for Crop Classification under Temporal and Geographic Shift

Authors: Miro Miranda, Rocco Sedona, Valerio Marsocci, Gabriele Cavallaro, Claudia Paris

Organizations: Jülich Supercomputing Centre, Forschungszentrum Jülich, Germany · European Space Agency (ESA/ESRIN), Italy · University of Iceland, Iceland · University of Twente (ITC), the Netherlands

Abstract

Continuous and up-to-date crop type mapping is essential for agricultural monitoring 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 supervision 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 embeddings 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.

Figures & tables

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), the in-situ photograph 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.

Satellite branch	CE	$+$ KD	$\Delta$	$\Delta$ T	CE	$+$ KD	$\Delta$
	2018 test				2018 $\rightarrow$ 2022 transfer
Frozen AlphaEarth	63.5	66.5	$+2.9$	$-3.2$	52.6	56.0	$+3.4$
Frozen TESSERA	66.7	71.7	$+5.1$	$+2.0$	57.6	63.1	$+5.5$
S2 SITS (scratch)	62.0	65.2	$+3.2$	$-4.5$	54.0	56.4	$+2.4$
S1 $+$ S2 SITS (scratch)	66.1	68.7	$+2.6$	$-1.0$	57.5	58.7	$+1.2$
TESSERA fine-tuned	71.2	72.1	$+0.8$	$+2.4$	63.0	63.6	$+0.7$

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 ).

\Delta

is the KD

-

CE mean of per-seed paired differences, computed before rounding.

\Delta

T is the difference to teacher performance (69.7 macro-F1).

Figure 2: Left: Overall paired KD $-$ CE macro-F1 margins per branch from 2018 $\rightarrow$ 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 ( $+0.5$ , SD $1.4$ , mixed signs). Error bars indicate $\pm 1$ SD, paired for margins.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

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, $\Delta_{\mathrm{site}}=100(D_{\mathrm{CE}}-D_{\mathrm{KD}})/D_{\mathrm{CE}}$ , 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 $\Delta_{\mathrm{ratio}}=100(R_{\mathrm{CE}}-R_{\mathrm{KD}})/R_{\mathrm{CE}}$ 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 in $R$ indicates location-specific stability beyond increased class compactness.

Model	MacroF1%	WeightedF1%
Frozen AlphaEarth CE student [z]	63.5 $\pm$ 0.6	67.0 $\pm$ 0.4
+ KD [z]	66.5 $\pm$ 0.3	68.6 $\pm$ 0.4
CE control [z]	63.8 $\pm$ 0.7	67.4 $\pm$ 0.5
+ CLIP (InfoNCE)	63.7 $\pm$ 0.7	67.2 $\pm$ 0.6
+ SupCon	63.4 $\pm$ 0.4	67.1 $\pm$ 0.4
Photo probe [CLIP, v]	69.7 $\pm$ 0.4	69.2 $\pm$ 0.3

Appendix

Table 2: Macro F1 and Weighted F1 for the frozen AlphaEarth CE baseline, contrastive objectives (InfoNCE, SupCon), photo probe, and fusion, mean

\pm

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.

	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 $\pm$ 0.9	71.3 $\pm$ 0.9	74.0 $\pm$ 0.2	76.0 $\pm$ 0.7	71.4 $\pm$ 1.3	73.7 $\pm$ 0.8	73.4 $\pm$ 2.2	76.3 $\pm$ 1.0	77.2 $\pm$ 1.1	78.3 $\pm$ 0.6
Maize	87.8 $\pm$ 0.4	89.0 $\pm$ 0.2	89.0 $\pm$ 0.3	89.5 $\pm$ 0.2	88.6 $\pm$ 0.8	90.0 $\pm$ 0.3	89.9 $\pm$ 0.4	91.1 $\pm$ 0.3	91.4 $\pm$ 0.4	91.3 $\pm$ 0.4
Barley	55.6 $\pm$ 1.1	55.6 $\pm$ 0.9	64.3 $\pm$ 1.0	63.6 $\pm$ 0.6	57.4 $\pm$ 1.3	57.9 $\pm$ 1.2	59.6 $\pm$ 0.6	60.8 $\pm$ 0.8	65.5 $\pm$ 0.8	64.8 $\pm$ 0.9
Other cereals	46.6 $\pm$ 0.5	46.6 $\pm$ 0.4	51.4 $\pm$ 0.9	49.8 $\pm$ 1.2	44.1 $\pm$ 0.8	43.3 $\pm$ 1.1	48.6 $\pm$ 0.6	49.3 $\pm$ 0.9	53.5 $\pm$ 1.3	53.6 $\pm$ 1.0
Rapeseed	77.4 $\pm$ 1.6	79.4 $\pm$ 0.6	80.6 $\pm$ 1.0	83.3 $\pm$ 0.3	79.0 $\pm$ 0.1	79.8 $\pm$ 0.5	82.1 $\pm$ 0.9	84.0 $\pm$ 0.5	82.1 $\pm$ 0.5	83.4 $\pm$ 0.4

Appendix

Table 3: Per-class F1 (%) on the 2018 test set, mean

\pm

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	64.2 $\pm$ 0.4	65.8 $\pm$ 0.9	65.9 $\pm$ 1.2	70.8 $\pm$ 1.1	64.0 $\pm$ 1.4	66.4 $\pm$ 0.6	66.6 $\pm$ 2.5	69.7 $\pm$ 0.9	71.2 $\pm$ 1.5	72.3 $\pm$ 0.5
Maize	78.5 $\pm$ 0.9	83.3 $\pm$ 0.2	86.3 $\pm$ 0.4	88.3 $\pm$ 0.2	83.8 $\pm$ 0.7	85.7 $\pm$ 0.9	85.8 $\pm$ 0.6	87.2 $\pm$ 0.2	88.1 $\pm$ 0.3	88.6 $\pm$ 0.3
Barley	48.0 $\pm$ 1.9	48.3 $\pm$ 0.2	55.5 $\pm$ 0.5	57.3 $\pm$ 1.1	49.5 $\pm$ 1.0	50.1 $\pm$ 0.3	52.3 $\pm$ 0.8	53.1 $\pm$ 0.8	57.7 $\pm$ 1.4	58.5 $\pm$ 0.4
Other cereals	42.2 $\pm$ 1.5	42.7 $\pm$ 1.9	47.6 $\pm$ 1.6	48.3 $\pm$ 1.5	41.9 $\pm$ 1.7	41.9 $\pm$ 1.8	46.0 $\pm$ 1.1	46.2 $\pm$ 0.5	50.3 $\pm$ 1.4	49.0 $\pm$ 1.9
Rapeseed	68.4 $\pm$ 1.1	69.6 $\pm$ 0.4	71.8 $\pm$ 0.9	73.5 $\pm$ 0.5	67.8 $\pm$ 1.2	70.1 $\pm$ 0.8	73.3 $\pm$ 0.2	74.2 $\pm$ 0.1	72.7 $\pm$ 0.5	74.3 $\pm$ 0.5

Appendix

Table 4: Per-class F1 (%) under 2018

\rightarrow

2022 temporal transfer, mean

\pm

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.

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.

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.

Agricultural Remote Sensing

Agricultural remote sensing uses satellite and airborne observations to map crops and fields and monitor yield, soil, and crop health.

See all 84 papers

Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability
Oct 27, 2026Priyanka Goyal, Maitreyee Sar, Adway Mitra +1Agricultural Image Analysis Agricultural Remote Sensing
Department of Artificial Intelligence, Indian Institute of Technology Kharagpur, West Bengal, India · Centre for Atmospheric Sciences, Indian Institute of Technology Delhi, India · Tata Consultancy Services Research, Kolkata, India
PhenoStitch: Training-Free Panoptic Crop Mapping from Satellite Image Time Series
Aug 1, 2026Xuechen LiPanoptic Segmentation Agricultural Remote Sensing
University of Minnesota Minneapolis, MN, USA
Space2Ground 2.0: A Multi-Source Dataset and Framework for Agricultural Monitoring through Fusion of Street-Level and Satellite Imagery
Jul 30, 2026Iason Tsardanidis, Alkiviadis Koukos, George Choumos +2Agricultural Image Analysis Agricultural Remote Sensing
Operational Unit BEYOND Centre, IAASARS, National Observatory of Athens, Athens, Greece · DHI Water & Environment, Hørsholm, Denmark · Artificial Intelligence Group, Wageningen University & Research, The Netherlands

Oct 27, 2026J/K move · Enter open · S save

Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability

Priyanka Goyal, Maitreyee Sar, Adway Mitra, Manjira Sinha

Department of Artificial Intelligence, Indian Institute of Technology Kharagpur, West Bengal, India · Centre for Atmospheric Sciences, Indian Institute of Technology Delhi, India · Tata Consultancy Services Research, Kolkata, India

Agricultural Image Analysis Agricultural Remote Sensing Geospatial Foundation Models Cross-Domain Transfer Learning Temporal Representation Learning

Open paper AskPDF arXiv

Earth Observation (EO)-based agricultural monitoring is shifting from pixel-level spectral analysis toward representation-driven modeling, yet systematic evaluation of learned representations across spatial, temporal, and domain-generalization dimensions remains limited. Geospatial foundation models generate high-dimensional embeddings encoding spatio-temporal land-surface properties, but their utility for agricultural applications remains insufficiently understood. This study presents a representation-agnostic evaluation framework integrating geometric separability, temporal dynamics, spatial robustness, and cross-regional generalisation. A 64-dimensional AlphaEarth embedding space is benchmarked against Sentinel-2 reflectance, NDVI-derived phenology, and harmonic features using field-scale data from the Conterminous United States (CONUS; 2017-2023) and Northern India (2021). Traditional features achieve strong linear separability (Fisher ratio), while embeddings provide superior distribution-aware separation -- Bhattacharyya and Mahalanobis distances up to an order of magnitude higher for spectrally similar pairs (e.g. Corn/Soybeans: 3.9 vs. 0.10). Embeddings uniquely capture multi-year crop rotation dynamics, reaching a separability ratio of 1.784 versus a maximum of 1.003 across all engineered features, and exhibit lower residual spatial autocorrelation of errors (mean Moran's

I \approx 0.17

). However, harmonic and NDVI features show stronger cross-continental transferability (US

\to

India Macro F1: 0.493 vs. 0.326 for embeddings), reflecting better distributional alignment. The combined feature space yields balanced performance, supporting hybrid modeling. These findings reveal a fundamental trade-off between representational richness and cross-region transferability, providing a structured framework for feature selection in large-scale crop classification.

Abstract

Agricultural Remote Sensing

Explore similar work

Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability

PhenoStitch: Training-Free Panoptic Crop Mapping from Satellite Image Time Series

Space2Ground 2.0: A Multi-Source Dataset and Framework for Agricultural Monitoring through Fusion of Street-Level and Satellite Imagery