Geospatial Foundational Embedder: Top-1 Winning Solution on EarthVision Embed2Scale Challenge (CVPR 2025)

Fuxun Yu; Mike Bianco; Nikolaos Karianakis; Qi Zhang; Raphael Tang; Rishi Madhok; Zirui Xu

REVIEW 4 major objections 5 minor 9 references

Geospatial Foundational Embedder: Top-1 Winning Solution on EarthVision Embed2Scale Challenge (CVPR 2025)

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Ensemble of four models tops Embed2Scale geospatial leaderboard

desk verdict Top-1 challenge report: plausible recipe, but Section 2.3's full-rank linear-map claim is mathematically false and unvalidated. read the letter →

arxiv 2509.06993 v1 pith:OLHAUZSL submitted 2025-09-03 cs.CV

Zirui Xu , Raphael Tang , Mike Bianco , Qi Zhang , Rishi Madhok , Nikolaos Karianakis , Fuxun Yu This is my paper · ORCID

classification cs.CV

keywords geospatialembeddingsSSL4EO-S12hyperspectralCLIPDINOensemblecompressionSVDremotesensing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper reports the winning solution for the Embed2Scale geospatial challenge, which asks for embedding vectors from multiseasonal, 128-channel hyperspectral satellite cubes that work well across many downstream tasks. The authors claim the win came not from a single model but from an ensemble of four complementary embeddings: a CLIP-finetuned ConvNeXt-XXL and ViT-Huge, a DINO-finetuned ViT-Base, and a remote-sensing vision-language model applied separately to the four RGB seasons. They compress the concatenated embeddings with truncated SVD and, before compression, learn a linear map on the vision-language embeddings using pseudolabels to make them more linearly separable. If correct, the recipe shows that diverse pretrained models plus simple compression can beat larger single-model approaches on geospatial representation benchmarks.

What carries the argument

The load-bearing piece is the SVD-based compressive ensemble: four distinct embedding spaces are truncated to different dimensions with truncated SVD and concatenated, which keeps the final vector compact while preserving linear-probe performance. Supporting it are two pretraining modifications: CLIP models whose first layer is expanded from 3 to 128 channels by duplicating weights, and text supervision enriched with latitude/longitude plus four regression attributes (forest cover, elevation, nightlights, population density). A third mechanism is the GeoRSCLIP unsupervised finetuning step—agglomerative clustering produces pseudolabels, and a full-rank square linear map is fit end-to-end with

What would settle it

Run the challenge's downstream evaluation with the GeoRSCLIP linear-map transform removed from the submitted embeddings; if the remaining ensemble no longer beats the runner-up on the task-balanced score, the claimed win depends entirely on that unvalidated transform.

Watch

Extended reading notes

Core claim

The central claim is that the Top-1 result came from a compressive ensemble of four foundation models, each covering a different inductive bias, combined with a linear map learned from pseudolabels. On the test leaderboard, the team's vanilla mean score was second, but the challenge's task-balanced score, which weights tasks by participant spread, put the ensemble first; the paper interprets this as evidence of stronger cross-task generalization. The contribution is a composition recipe: take CLIP and DINO models fine-tuned on 128-channel SSL4EO data, add GeoRSCLIP embeddings from each of four RGB seasons, learn a square linear map on the GeoRSCLIP embeddings using agglomerative-clustering p

Load-bearing premise

The linear map that transforms the GeoRSCLIP embeddings is fit on the same development-set embeddings it transforms, and the paper gives no evidence that the learned transformation keeps improving linear separability on unseen test data rather than merely memorizing the development set.

Editorial extensions

If this is right

Combining a CNN and a ViT, and mixing CLIP, DINO, and vision-language objectives, gives a stronger geospatial embedding than any single architecture or objective.
SVD compression can reduce embedding dimension substantially with only minor loss for linear downstream tasks, making large ensembles practical.
Adding location and regression attributes as text during CLIP fine-tuning improves the usefulness of the embedding for geospatial tasks.
Processing the same scene in separate seasons through a natural-language remote-sensing model adds information complementary to the hyperspectral stack.
A linear map tuned on pseudolabels can make off-the-shelf embeddings more linearly separable, matching the constraints of linear-probe test setups.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The pseudolabel linear map is a generic post-processing step: the same trick could be applied to any self-supervised embedding before linear probing, though the paper never evaluates it on a held-out set.
If the task-balanced score rewards consistent performance, the ensemble's win suggests that diversity of inductive biases matters more than peak accuracy on a single task; this could generalize to other foundation-model benchmarks.
The seasonal decomposition of GeoRSCLIP means the method multiplies inference cost by four; a testable extension is to distill the four seasonal embeddings into a single model to retain the gain at lower cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The paper reports a Top-1 win on the Embed2Scale challenge with an ensemble of four geospatial embedding models (CLIP on hyperspectral with metadata text, DINO, and four-season GeoRSCLIP) compressed via SVD and concatenated. The competition result is credible, and the recipe is the kind of practical combination that the community can use. Metadata-as-text supervision and SVD compression with quality checks are sensible engineering choices.

The soft spots are real, and one is load-bearing. Section 2.3 claims that a full-rank square linear map on GeoRSCLIP embeddings improves linear separability for the test-time linear models. That is false: for any full-rank M, the set of achievable linear classifiers on Mx is exactly the same as on x, since w^T(Mx) = (M^T w)^T x and M^T is a bijection. If M is not full rank, the hypothesis class shrinks, so optimal linear loss cannot improve. The authors state the intuition explicitly, so this is an internal inconsistency, not a subtle generalization failure. The pseudolabel clustering is also fitted on the same development embeddings, and there is no ablation or held-out validation showing any benefit. It may be harmless if the learned M is near-identity, but the claimed mechanism doesn't hold.

The report also lacks what would make it reproducible: no code, no embedding dimensions per model beyond the table, no quantitative scores, no ablations. The 'stronger generalization' conclusion built on the task-balanced weighting is a stretch, since that weighting depends on all participants' scores and is not a direct property of the model.

That said, the central claim—winning the challenge—is externally anchored and probably real. The ensemble is a legitimate combination of existing components, and the SVD+concat approach is reasonable. For the geospatial practitioner, the paper is useful as a recipe, but as a technical report it is thin.

If this were submitted to a peer-reviewed venue, I would send it to reviewers: the competition result deserves scrutiny and the recipe is worth documenting, but Section 2.3 should be corrected or removed, and the authors need to share code and dev/test validation. As it stands, treat the linear-map component as unproven and possibly a no-op.

Referee Report

4 major / 5 minor

Summary. The paper describes the authors' Top-1 winning solution to the EarthVision Embed2Scale challenge at CVPR 2025. The method consists of four embedding models — a ConvNeXt-XXL and a ViT-Huge finetuned with CLIP using geographic/regression text metadata, a ViT-Base finetuned with DINO, and an off-the-shelf GeoRSCLIP ViT-Huge applied to four seasonal RGB views. The embeddings are then SVD-compressed and concatenated. Before ensembling, the GeoRSCLIP embeddings are transformed by a linear map fitted on development-set pseudolabels. The central claim is that the final embedding ranks first on the task-balanced test leaderboard.

Significance. The practical achievement — ranking first in an externally judged competition — is a genuine strength, and the paper describes a plausible high-level recipe (diverse model families, multi-modal geographic text conditioning, and SVD-based compressive ensembling). The external leaderboard anchors the claim in an independent evaluation. However, the scientific value is limited by the almost complete absence of quantitative scores, ablations, error bars, and implementation details. More seriously, one described component (Section 2.3) is theoretically problematic. If the result is taken at face value, the paper is a useful competition report, but it does not meet the evidentiary standard of a archival paper without substantial revision.

major comments (4)

[Section 2.3] The proposed full-rank square linear map M cannot improve linear separability for a subsequent linear classifier. For any test-time linear model on transformed features, w^T(Mx) = (M^T w)^T x; if M is invertible, the set of achievable linear classifiers on Mx is identical to that on x. If M is rank-deficient, the hypothesis class is strictly smaller, so the optimal linear loss cannot decrease. Thus the stated mechanism ('distort the embedding space to improve learnability') is theoretically a no-op. No ablation isolates this component or shows any benefit on a held-out split. The authors should either remove the claim, restrict it to a nonlinear/regularized training effect, or provide rigorous evidence that the finetuning step helps.
[Section 2.3] The pseudolabel linear map and auxiliary linear model are fitted on the development set using agglomerative clustering of the same GeoRSCLIP embeddings that are later transformed. This is self-referential: the labels and the transformation come from the same data distribution, and no independent validation is reported. Any apparent improvement on the development set could be an overfitting artifact. To support the claim that the transformation generalizes, the authors must evaluate on a held-out subset of the development data or on the official validation set, and compare with/without the linear map.
[Section 3] The central claim rests entirely on a leaderboard screenshot (Figure 2). The paper does not report any numeric scores: neither per-task scores, nor the unweighted q-mean, nor the task-balanced q-mean. This makes the result impossible to audit or compare with future work. The authors should include a table with the official scores for the top few teams, the submission identifier, and the date of the leaderboard snapshot. Since leaderboards can be dynamic, a timestamp and competition-phase identifier are necessary.
[Section 3, Figure 2] The sentence 'we are ranked Top-1 on the task-balanced q mean score ... denoting stronger generalization capability' is not logically justified. Task-balanced weighting assigns higher weight to tasks where participants differ more, so a top task-balanced rank reflects performance on a particular weighted aggregate, not a direct measure of generalization across tasks. To support the generalization claim, the authors should report per-task scores and show that their method is robust across tasks, e.g., by ranking first or near-first on most tasks rather than relying on one weighting scheme.

minor comments (5)

[Throughout] There are many typos and formatting issues: 'faciliatetes', 'Geospatialtial', 'Latitute', 'Longtitute', 'comptition', 'vanillaunweighted', 'hypersepctral', and the author address line 'Redmond, W A' (should be WA). A copy-edit is needed.
[Table 1] The asterisk in Table 1 is confusing: it marks the same ViT-Huge model used four times, but the table lists four entries. Clarify that the same GeoRSCLIP model is applied to the four seasonal RGB channels separately.
[Section 2.1] The initialization of the 128-channel first layer by duplicating 3-channel weights is stated without details. Report how the weights are duplicated and whether the duplicated weights are normalized; otherwise the scale of the input activations may change.
[Section 2.4] The hyperparameter search over SVD target dimensions is described qualitatively. Provide the evaluated compression rates, the MSE/silhouette values, and the final selection criterion, so the reader can reproduce the choice.
[References] Reference [1] (Baranwal et al.) is used to justify improved linear separability for graphical embeddings, but the connection to the proposed linear map is not explained. Either expand this justification or remove the citation.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential pseudolabel finetuning; central Top-1 claim rests on external leaderboard and is not circular.

other [Section 2.3, paragraph 1]
"we constructed a set of pseudolabels using agglomerative clustering over the four-season concatenated GeoRSCLIP embeddings. Then, we applied a linear map ... fitted a linear model ... to predict the pseudolabels. Both the linear map and linear model were optimized end-to-end on the development set, and we used the fully optimized linear map to transform and produce our final GeoRSCLIP embeddings."

The pseudolabels are generated from the exact same GeoRSCLIP embeddings that the linear map M transforms. Fitting M and a linear probe to predict those self-generated labels can trivially reduce training loss on the clustering objective; any apparent 'improvement' is by construction rather than evidence of better linear separability on independent test data. This is a self-referential fitting loop, but the paper's headline claim (Top-1 on the TEST leaderboard) is anchored to an external benchmark, so this step is not load-bearing for the main result.

full rationale

The paper's central claim is an externally measured competition result: 'Our ensemble embedding method ranks Top-1 in the TEST phase leaderboard.' That claim is validated by the challenge backend, not by the paper's own construction, so it cannot be circular in the sense of deriving its conclusion from its inputs. The only self-referential element is Section 2.3, where pseudolabels are derived from the same GeoRSCLIP embeddings that are later transformed by a fitted linear map. However, the paper does not use this to 'predict' the leaderboard score; it merely applies the transformation as one component of an ensemble that is then externally evaluated. No load-bearing self-citations are present: references [7] and [8] are external prior work, not the authors' own. The mathematical concern raised by a skeptic (a full-rank linear map cannot change the hypothesis class of linear models, so the claimed mechanism is a no-op at best) is a correctness/validity issue, not a circularity issue. Under the circularity pass rules, that concern should not inflate the circularity score. Overall, the derivation is self-contained against external benchmarks, with minor self-referential fitting keeping the score slightly above zero.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. Its central claim rests on the effectiveness of existing models and SVD, and on several ad hoc assumptions about weight duplication and pseudolabel transfer.

free parameters (4)

SVD target dimension per model = ConvNeXt 128; ViT-H CLIP 256; ViT-B DINO 128; GeoRSCLIP 4x128 (Table 1)
Chosen by empirical search over compression rates using reconstruction error and silhouette score on the development set; these sizes directly determine the final embedding layout.
Number of agglomerative clusters
Determines pseudolabel granularity in Section 2.3 and therefore the trained linear map; no value is given.
Linear map matrix
Square matrix fitted end-to-end on the development set to predict pseudolabels; its entries are optimized but not disclosed.
CLIP/DINO finetuning hyperparameters
Learning rate, epochs, batch size, and other training choices are omitted, though they affect all finetuned model embeddings.

assumptions (5)

standard math Truncated SVD yields a low-rank approximation that preserves enough information for downstream tasks
Used in Section 2.4; the paper relies on the standard properties of SVD without proof.
domain assumption The SSL4EO-S12 cube's first three channels are RGB for each season
Section 2.3 applies GeoRSCLIP to the RGB channels of the four seasons; the report assumes this channel correspondence holds.
domain assumption TorchSpatial attributes (forest cover, elevation, nightlights, population) are accurate at the given lat/lon
Section 2.2 uses these values as text supervision; if they are noisy, the CLIP text encoder learns from incorrect metadata.
ad hoc to paper Duplicating 3-channel weights to 128 channels is a valid initialization for the hyperspectral first layer
Section 2.1 states that the first layer is modified by duplicating weights; this is a heuristic with no analysis.
ad hoc to paper The pseudolabel linear map improves linear separability on the test set
Section 2.3 hypothesizes this; the report offers no evidence that the transformation transfers beyond the development set.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Geospatial Foundational Embedder: Top-1 Winning Solution on EarthVision Embed2Scale Challenge (CVPR 2025)." pith.science (2026). https://pith.science/paper/OLHAUZSL

@misc{pith2026250906993,
  author       = {Pith},
  title        = {Pith review of: Geospatial Foundational Embedder: Top-1 Winning Solution on EarthVision Embed2Scale Challenge (CVPR 2025)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OLHAUZSL}},
  note         = {Machine review of arXiv:2509.06993}
}

read the original abstract

EarthVision Embed2Scale challenge (CVPR 2025) aims to develop foundational geospatial models to embed SSL4EO-S12 hyperspectral geospatial data cubes into embedding vectors that faciliatetes various downstream tasks, e.g., classification, regression, etc. In this technical report, we introduce our proposed method for the Top-1 winning solution on the Embed2Scale Challenge.

Figures

Figures reproduced from arXiv: 2509.06993 by the authors.

**Figure 1.** SSL4EO dataset [2]: Each embedding target in the challenge is one novel 4-season hypersepctral geospatial data cube with 4 x 17 = 128 channels. • State-of-the-Art Pretraining Models: We adopt two families of powerful pretraining methods: CLIP [4], DINO [3]. With adapted input channel numbers, we finetuned them on the EO4SSL dataset; • Geospatial Semantic Enrichment: By incorporating multiple auxiliary regression si… view at source ↗

**Figure 2.** The TEST Phase Leaderboard. Interestingly, although [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 6 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...
[2]

Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization, 2022

Aseem Baranwal, Kimon Fountoulakis, and Aukosh Jagannath. Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization, 2022

work page 2022
[3]

Ssl4eo-s12 v1

Benedikt Blumenstiel, Nassim Ait Ali Braham, Conrad M Albrecht, Stefano Maurogiovanni, and Paolo Fraccaro. Ssl4eo-s12 v1. 1: A multimodal, multiseasonal dataset for pretraining, updated. arXiv preprint arXiv:2503.00168 , 2025

arXiv 2025
[4]

Emerging properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv \'e J \'e gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650--9660, 2021

work page 2021
[5]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , pages 8748--8763. PmLR, 2021

work page 2021
[6]

Cluster quality analysis using silhouette score

Ketan Rajshekhar Shahapure and Charles Nicholas. Cluster quality analysis using silhouette score. In 2020 IEEE 7th international conference on data science and advanced analytics (DSAA) , pages 747--748. IEEE, 2020

work page 2020
[7]

Fast approximate truncated svd

Serge L Shishkin, Arkadi Shalaginov, and Shaunak D Bopardikar. Fast approximate truncated svd. Numerical Linear Algebra with Applications , 26(4):e2246, 2019

work page 2019
[8]

Torchspatial: A location encoding framework and benchmark for spatial representation learning

Nemin Wu, Qian Cao, Zhangyu Wang, Zeping Liu, Yanlin Qi, Jielu Zhang, Joshua Ni, Xiaobai Yao, Hongxu Ma, Lan Mu, et al. Torchspatial: A location encoding framework and benchmark for spatial representation learning. arXiv preprint arXiv:2406.15658 , 2024

arXiv 2024

Show all 9 references

[9]

Rs5m and georsclip: A large scale vision-language dataset and a large vision-language model for remote sensing

Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. Rs5m and georsclip: A large scale vision-language dataset and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing , 2024

2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

[2] [2]

Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization, 2022

Aseem Baranwal, Kimon Fountoulakis, and Aukosh Jagannath. Graph convolution for semi-supervised classification: Improved linear separability and out-of-distribution generalization, 2022

work page 2022

[3] [3]

Ssl4eo-s12 v1

Benedikt Blumenstiel, Nassim Ait Ali Braham, Conrad M Albrecht, Stefano Maurogiovanni, and Paolo Fraccaro. Ssl4eo-s12 v1. 1: A multimodal, multiseasonal dataset for pretraining, updated. arXiv preprint arXiv:2503.00168 , 2025

arXiv 2025

[4] [4]

Emerging properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv \'e J \'e gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650--9660, 2021

work page 2021

[5] [5]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , pages 8748--8763. PmLR, 2021

work page 2021

[6] [6]

Cluster quality analysis using silhouette score

Ketan Rajshekhar Shahapure and Charles Nicholas. Cluster quality analysis using silhouette score. In 2020 IEEE 7th international conference on data science and advanced analytics (DSAA) , pages 747--748. IEEE, 2020

work page 2020

[7] [7]

Fast approximate truncated svd

Serge L Shishkin, Arkadi Shalaginov, and Shaunak D Bopardikar. Fast approximate truncated svd. Numerical Linear Algebra with Applications , 26(4):e2246, 2019

work page 2019

[8] [8]

Torchspatial: A location encoding framework and benchmark for spatial representation learning

Nemin Wu, Qian Cao, Zhangyu Wang, Zeping Liu, Yanlin Qi, Jielu Zhang, Joshua Ni, Xiaobai Yao, Hongxu Ma, Lan Mu, et al. Torchspatial: A location encoding framework and benchmark for spatial representation learning. arXiv preprint arXiv:2406.15658 , 2024

arXiv 2024