Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

This paper claims that using all 12 modalities available at test time as reconstruction targets—a method called TTT-MMR—improves performance over joint training on all five environmental tasks and both random and geographic test splits, wit

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 03:55 UTC pith:UZWCMGJO

load-bearing objection A useful benchmark, a plausible but overclaimed TTT method; the geographic claims need a missing transfer check. the 3 major comments →

arxiv 2602.06285 v3 pith:UZWCMGJO submitted 2026-02-06 cs.CV

MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training

Lucia Gordon , Serge Belongie , Christian Igel , Nico Lang This is my paper
classification cs.CV
keywords test-time trainingmultimodal reconstructionEarth observationgeographic generalizationbenchmark datasetremote sensingsatellite imageryself-supervised learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces MMEarth-Bench, a new benchmark of five globally distributed environmental tasks (biomass, soil nitrogen, soil organic carbon, soil pH, species occurrence) with 12 aligned modalities per task and both random and geographic (Africa) test splits. It benchmarked seven pretrained models and found that while multimodal pretraining helps in limited-data settings, geographic generalization is poor, and using multiple input modalities can even increase overfitting to the non-Africa training domain. To address this, the paper proposes test-time training with multimodal reconstruction (TTT-MMR), which repurposes all available modalities as auxiliary reconstruction targets at test time, providing a self-supervised adaptation signal for any encoder. The authors report that TTT-MMR improves performance over a joint-training baseline on all tasks, across all models, and on both test splits, with geographic batching (TTT-MMR-GEO) offering a beneficial trade-off between regularization and specialization.

Core claim

The central claim is that a model-agnostic test-time training procedure can use the full set of task modalities—whether or not the pretrained encoder accepts them as input—by treating reconstruction of those modalities as a self-supervised adaptation signal. During joint training, a lightweight linear decoder h is trained to reconstruct all 12 modalities from the encoder's embeddings; at test time, this decoder is frozen and the reconstruction losses are backpropagated into the encoder, with per-modality gradients normalized and averaged so each modality contributes equally. Across five tasks, seven diverse pretrained models, and both random and geographic splits, TTT-MMR outperforms the joi

What carries the argument

The task modality decoder h: a lightweight linear head trained jointly with the main task decoder and shared encoder to reconstruct all 12 modalities from the encoder's embeddings. At test time it is frozen, and the per-modality reconstruction losses provide a normalized, averaged gradient signal that adapts the encoder before the final prediction. Geographic batching, generated by recursive spatial partitioning, groups test tiles into contiguous regions to balance regularization and specialization.

Load-bearing premise

The reconstruction head trained on non-Africa tiles remains accurate enough on African tiles that minimizing reconstruction error there steers the encoder toward features that also improve the downstream task.

What would settle it

Take the same tasks but define a geographic split on a region whose modality statistics differ sharply from Africa, such as high latitudes or a region with a novel land-cover class. If TTT-MMR on that region does not improve over joint training, or actually degrades performance, the claim that a reconstruction head trained on non-target data transfers to the target region is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any pretrained encoder, regardless of its training modalities, can exploit modalities it has never seen by treating them as reconstruction targets during test-time adaptation.
  • Geographic batching of test data provides a regularization-specialization trade-off that is especially beneficial for long-tail distributions, with adaptation cost under one second per sample.
  • Performance gains are consistent on both random and held-out-geography splits, with larger improvements on the geographic split for biomass and soil organic carbon.
  • Multimodal finetuning can hurt geographic generalization, as learned fusion modules may overfit the non-Africa training domain, suggesting careful handling of modality fusion for out-of-distribution use.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The mechanism implies that test-time reconstruction could be applied beyond Earth observation, whenever a model is pretrained on one sensor subset but aligned auxiliary sensors are available at deployment, such as in climate or health monitoring.
  • A testable extension: ablating individual modality reconstruction targets would reveal which modalities drive the improvement; the paper uses all 12 jointly but does not isolate their contributions.
  • Since the reconstruction head is linear and frozen, the gains likely arise from re-aligning encoder features rather than from increasing representational capacity; this could be verified by checking whether gains persist with even lighter or random reconstruction heads.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces MMEarth-Bench, a multimodal Earth observation benchmark with five new global tasks (biomass, soil nitrogen, soil organic carbon, soil pH, species occurrence) sharing 12 aligned modalities and evaluated under both random and Africa-held-out geographic test splits. It benchmarks seven pretrained models and a randomly initialized baseline under 5/50/100% training-data regimes. The paper then proposes TTT-MMR: after joint training (JT) of the task head and a frozen linear task-modality decoder h on reconstruction of all 12 modalities, test-time training adapts the encoder by gradient descent on the averaged, per-modality normalized reconstruction losses; TTT-MMR-GEO additionally batches test tiles by geographic proximity. The headline claim is that TTT-MMR improves performance over JT on all models, all tasks, and both test splits.

Significance. If supported, the benchmark contribution is substantial: existing EO benchmarks rarely share the same 12 modalities across tasks, are not globally distributed, and typically lack a strict held-out-region geographic split. The experimental design is thorough in many respects — 7 models, 3 seeds, multiple training-set fractions, raw result tables, and public code/data links. TTT-MMR is also methodologically attractive because it is architecture-agnostic, avoids batch-norm/FiLM modifications, and turns unused modalities into reconstruction targets rather than requiring them as inputs. However, the geographic half of the central claim is not currently established: the linear reconstruction decoder h is trained only on non-Africa tiles and is frozen during TTT, and no evidence is provided that h transfers to Africa. The manuscript also states a stronger claim than the raw per-model-task numbers support. These are fixable with additional diagnostics and more careful statistical reporting, but they are load-bearing for the paper's main conclusions.

major comments (3)
  1. [Sec. 4, Eq. (2); Sec. 5.5, Fig. 7; Sec. 5.4] The geographic-split claim depends on the frozen task-modality decoder h transferring from non-Africa to Africa. h is trained only on non-Africa tiles (Sec. 3, S.1.6) and is frozen during TTT, so every gradient in Eq. (2) is computed through h's reconstruction error. The paper never reports reconstruction loss on the geographic split before or after TTT, nor compares h's Africa reconstruction error with its random-split error. This is not a formality: Sec. 5.4 states that additional modalities can 'lead to overfitting to the non-Africa training domain,' which is precisely the failure mode that would make Eq. (2) push African embeddings toward source-like features. The reported Africa gains could also be driven by tile-level proxies (geolocation, biome, ecoregion) in the reconstruction target set rather than by a transferable self-supervised signal. Please report reconstruction R²/MSE per
  2. [Sec. 5.5, Table 5; Tables S.16–S.30] The claim that TTT-MMR 'improves performance across all models and tasks' is stronger than the evidence. Table 5 reports average ranks over tasks and seeds, and Fig. 7 pools all models/seeds before applying a one-sided Wilcoxon test. The raw per-model-task tables contain many zero deltas (e.g., TerraMind species random split: JT = 0.95, TTT-MMR = 0.95, Tables S.28–S.30; DINOv3 Sat biomass random split: 0.40 to 0.40, Table S.16) and even negative deltas (e.g., SatlasNet soil OC geographic split: JT = −0.08, TTT-MMR = −0.25, Table S.22). Please either qualify the headline claim as averaged behavior or provide per-model-task significance testing and effect sizes with appropriate multiple-comparison correction.
  3. [Sec. 4, Eq. (2); Sec. 5.5, Fig. 7] The random-split gains are consistent with a 'more training' effect rather than with test-time domain adaptation. The TTT objective is exactly the reconstruction objective used in JT; applying additional gradient steps on random test batches is equivalent to continuing optimization on the same loss. To support the adaptation claim, include a control that performs the same number of additional reconstruction-gradient steps on random unlabeled data (or on the training set) with the same frozen h, and show that using the actual test batch or geographic batching matters beyond additional compute. This distinction is important because the random split is in-distribution by construction.
minor comments (5)
  1. [Table 5 vs. Table S.10] The main text says TTT-MMR-GEO is the best method for all models except SatlasNet, but Table S.10 states that by the RI metric TTT-MMR-GEO is best for all models and splits. Please reconcile the two ranking criteria or clarify which one supports the claim.
  2. [Fig. 7 and Fig. S.15] Fig. 7 pools raw ΔR² and ΔmAP values, which are not directly comparable across metrics or tasks. Consider making the relative-improvement metric of Fig. S.15 the primary display, or justify why raw deltas are appropriate.
  3. [Eq. (1), Table S.3] Sentinel-1 has roughly 60% missing values for most tasks. The text says missing modalities are excluded from the gradient average, but the exact mechanism is not specified. Please clarify how missing modalities are handled in Eq. (1) and in the per-modality gradient normalization.
  4. [Sec. 4, Table 3] The reconstruction target set includes geolocation encoding, month encoding, biome, and ecoregion. Because these are tile-level and highly informative about location, the linear decoder may exploit shortcuts. An ablation excluding these targets would strengthen the interpretation of the adaptation signal.
  5. [Sec. 5.2] The paragraph beginning 'Originally, Copernicus-FM...' appears to address a reviewer concern but breaks the flow of the results section. Consider moving it to the experimental setup.

Circularity Check

0 steps flagged

No significant circularity: TTT-MMR improvements are empirical, not derived from fitted parameters or self-citation chains.

full rationale

The paper's central claim is empirical: TTT-MMR and TTT-MMR-GEO improve over the joint-training (JT) baseline on held-out random and geographic test splits (Fig. 7, Tab. 5, Tabs. S.16–S.30). The reported metrics are R^2 and mAP computed from task labels, which are never used during TTT. The reconstruction loss used at test time is the same objective used in JT, but that is a methodological design choice, not a definitional reduction: the paper tests the hypothesis by comparing TTT against JT across seven pretrained models and a random baseline, and even notes that MPMAE, pretrained by reconstructing the same modalities, shows no special TTT gains, which would be unlikely if the result were forced by construction. Hyperparameters (I_max, learning rate, batch size) are fixed or selected on the validation set, not on the test set, so no fitted parameter is renamed as a prediction. The only self-citation (MMEarth [42]) supplies the shared modality set, the MPMAE baseline, and the cyclic encoding for geolocation/month; none is load-bearing for the claimed improvement. The geographic-split result relies on the empirical assumption that the frozen task-modality decoder h transfers to Africa, and the paper does not report reconstruction-loss diagnostics for that split; this is a validation gap, not circularity. No equation in the paper reduces a reported prediction to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

No new physical entities are introduced. The central methodological assumptions are the value of reconstruction as a test-time signal, geographic batching, and the significance of the Africa split. No formal proofs are provided.

free parameters (3)
  • TTT learning rate λ = 1e-2
    Chosen for all experiments without sensitivity analysis; affects adaptation magnitude.
  • TTT batch size |B| = 8
    Fixed hyperparameter; the paper does not ablate it.
  • I_max (max TTT iterations) = 5
    The best number of iterations per batch is selected on the validation set; the mean is then used at test time. This is a tuned hyperparameter.
axioms (3)
  • domain assumption Reconstruction of the 12 task modalities at test time provides an adaptation signal that improves downstream task performance.
    This is the core premise of TTT-MMR; it is validated empirically but not derived. The paper shows aggregate gains but no analysis of when reconstruction helps or hurts.
  • domain assumption Geographic proximity of test tiles implies sample similarity that improves the trade-off between regularization and specialization.
    TTT-MMR-GEO relies on this; the paper reports it works on average but no theoretical or per-task characterization.
  • domain assumption The Africa split represents a meaningful out-of-distribution shift for global generalization.
    Africa is held out as the test region, but several models (e.g., DINOv3) may have seen Africa during pretraining, and the tasks have different label distributions there (e.g., biomass mean 26 vs. 65).

pith-pipeline@v1.3.0-alltime-deepseek · 47103 in / 6759 out tokens · 67209 ms · 2026-08-03T03:55:43.589445+00:00 · methodology

0 comments
read the original abstract

Recent research in geospatial machine learning demonstrates that models pretrained with self-supervised learning on Earth observation data can perform well on downstream tasks with limited labeled data. However, most benchmark datasets have few data modalities and poor global representation, limiting the ability to evaluate multimodal pretrained models at global scales. In order to fill this gap, we introduce MMEarth-Bench, a collection of five new environmental tasks with 12 modalities, globally distributed data, and both random and geographic test splits. We benchmark a diverse set of pretrained models and find that while (multimodal) pretraining tends to improve model robustness in limited data settings, geographic generalization abilities remain poor. Moreover, a simple randomly initialized multimodal model is competitive given enough labeled data. Although data is abundant, models can currently only make use of the modalities on which they were pretrained. To solve this problem, we propose using all the modalities available at test time as auxiliary tasks for test-time adaptation. Our model-agnostic method for test-time training with multimodal reconstruction (TTT-MMR) can improve performance across all models and tasks on both test splits. Furthermore, geographic batching leads to a good trade-off between regularization and specialization during TTT, which is especially beneficial for long-tail distributions. Our dataset, code, and visualization tool are linked on the project page: lgordon99.github.io/mmearth-bench.

Figures

Figures reproduced from arXiv: 2602.06285 by Christian Igel, Lucia Gordon, Nico Lang, Serge Belongie.

Figure 1
Figure 1. Figure 1: Self-supervised multimodal pretraining promises to overcome the grand challenges in Earth observation. Crucial applications have to rely on limited and sparse and geographi￾cally biased training data. Furthermore, the ambiguities inherent to modeling biophysical quantities with remotely sensed data may be resolved by models conditioned on multiple modalities. in the case of abundant reference data [75], bu… view at source ↗
Figure 2
Figure 2. Figure 2: Data splits in MMEARTH-BENCH. Each of the 5 tasks consists of a geographic test split (“Africa”) and splits the rest of the world randomly into training (70%), validation (15%), and random test (15%). While the full training dataset is shown here, we also provide subsets with 50% and 5% of the training data for even lower-shot experiments. this a multi-label classification task. The minimum number of tiles… view at source ↗
Figure 3
Figure 3. Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Low-shot in-distribution performance. Finetuning on subsets of the training data. Symbology: •=RGB, ■=S2, ▲=multimodal, solid=random init., dashed=pretrained. R G 0.1 0.4 Performance Biomass R G −0.4 0.5 Soil N R G Test Split −2.4 0.4 Soil OC R G 0.0 0.6 Soil pH R G 0.4 0.9 Species [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Geographic generalization. Performance comparison on random (R) vs. geographic (G) test splits using all training data. ery task modality at the tile resolution. We group the chan￾nels by modality and perform global average pooling on the bands of the tile-level modalities, yielding our modality re￾constructions. In computing the reconstruction losses we use cross-entropy loss for the categorical-valued mo… view at source ↗
Figure 6
Figure 6. Figure 6: Unimodal vs. multimodal input data. Finetuning performance of S2-only (solid-circle) versus multimodal (dashed￾triangle) variants of TerraMind and Copernicus-FM. regions. All the soil tasks and species occurrence offer an opportunity for globally pretrained models to improve their methodology to better facilitate geographic generalization on downstream tasks. The DINOv3 models have likely seen unlabeled da… view at source ↗
Figure 7
Figure 7. Figure 7: Multimodal test-time training improvement per task. Improvement of TTT-MMR (random batching) and TTT-MMR￾GEO (geographic batching) over joint training. Boxplots show the distribution over all models and seeds. ♦=mean, −=median, whiskers are based on the 1.5 IQR value. ∆ Performance reflects the absolute change in the performance metric. pacity to fit the training data. 5.5. Multimodal test-time training pe… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Better Together: Evaluating the Complementarity of Earth Embedding Models

    cs.CV 2026-05 unverdicted novelty 7.0

    Fusing embeddings from four Earth models (AlphaEarth, Tessera, GeoCLIP, SatCLIP) outperforms the best single model on four of six tasks, with gains depending on task and location.

  2. Now We Know? A Systematic Comparison of TerraMind and THOR

    cs.LG 2026-07 conditional novelty 6.0

    In a controlled comparison of two geospatial foundation models, patch size and decoder type explain more of the performance difference than the choice of model itself.

  3. Scalable and Trustworthy Earth Observation Foundation Models

    cs.LG 2026-07 conditional novelty 3.0

    Remote-sensing foundation models need domain-specific design and evaluation around measurement physics and decision constraints; benchmark accuracy alone is insufficient for trustworthy EO deployment.

Reference graph

Works this paper leans on

78 extracted references · 2 linked inside Pith · cited by 3 Pith papers

  1. [1]

    The IUCN Red List of Threatened Species

    IUCN 2025. The IUCN Red List of Threatened Species. Version 2025-1, Downloaded on June 13, 2025. 4, 15

    2025

  2. [2]

    Grundy, Alfred E

    Dominique Arrouays, Michael G. Grundy, Alfred E. Hartemink, Jonathan W. Hempel, Gerard B.M. Heuvelink, S. Young Hong, Philippe Lagacherie, Glenn Lelyk, Alexan- der B. McBratney, Neil J. McKenzie, Maria d.L. Mendonca- Santos, Budiman Minasny, Luca Montanarella, Inakwu O.A. Odeh, Pedro A. Sanchez, James A. Thompson, and Gan- Lin Zhang. Chapter three - globa...

  3. [3]

    AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities

    Guillaume Astruc, Nicolas Gonthier, Cl ´ement Mallet, and Loic Landrieu. AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities. InCVPR, 2025. 1, 2

    2025

  4. [4]

    Geography-aware self-supervised learning.ICCV, 2021

    Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tan- may, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning.ICCV, 2021. 2

    2021

  5. [5]

    MultiMAE: Multi-modal Multi-task Masked Autoen- coders.ECCV, 2022

    Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. MultiMAE: Multi-modal Multi-task Masked Autoen- coders.ECCV, 2022. 6

    2022

  6. [6]

    SatlasPretrain: A Large- Scale Dataset for Remote Sensing Image Understanding

    Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdi- nando, and Aniruddha Kembhavi. SatlasPretrain: A Large- Scale Dataset for Remote Sensing Image Understanding. In ICCV, Los Alamitos, CA, USA, 2023. IEEE Computer So- ciety. 2, 6, 20

    2023

  7. [7]

    Batjes, Luis Calisto, and Luis M

    Niels H. Batjes, Luis Calisto, and Luis M. de Sousa. Pro- viding quality-assessed and standardised soil data to sup- port global mapping and modelling (WoSIS snapshot 2023). Earth System Science Data, 2024. 4, 14

    2023

  8. [8]

    Multidimensional binary search trees used for associative searching.Commun

    Jon Louis Bentley. Multidimensional binary search trees used for associative searching.Commun. ACM, 1975. 6

    1975

  9. [9]

    Ter- raMesh: A Planetary Mosaic of Multimodal Earth Observa- tion Data, 2025

    Benedikt Blumenstiel, Paolo Fraccaro, Valerio Marsocci, Johannes Jakubik, Stefano Maurogiovanni, Mikolaj Cz- erkawski, Rocco Sedona, Gabriele Cavallaro, Thomas Brun- schwiler, Juan Bernabe-Moreno, and Nicolas Long´ep´e. Ter- raMesh: A Planetary Mosaic of Multimodal Earth Observa- tion Data, 2025. 6, 20

    2025

  10. [10]

    FoMo: multi-modal, multi-scale and multi-task remote sensing foundation models for forest monitoring

    Nikolaos Ioannis Bountos, Arthur Ouaknine, Ioannis Pa- poutsis, and David Rolnick. FoMo: multi-modal, multi-scale and multi-task remote sensing foundation models for forest monitoring. InAAAI. AAAI Press, 2025. 2, 3

    2025

  11. [11]

    Brown, S.P

    C.F. Brown, S.P. Brumby, and B. Guzder-Williams et al. Dy- namic World, Near real-time global 10 m land use land cover mapping.doi:10.1038/s41597- 022- 01307- 4,

    doi:10.1038/s41597-

  12. [12]

    Dynamic world, near real-time global 10 m land use land cover mapping.Sci- entific data, 2022

    Christopher F Brown, Steven P Brumby, Brookie Guzder- Williams, Tanya Birch, Samantha Brooks Hyde, Joseph Mazzariello, Wanda Czerwinski, Valerie J Pasquarella, Robert Haertel, Simon Ilyushchenko, et al. Dynamic world, near real-time global 10 m land use land cover mapping.Sci- entific data, 2022. 4

    2022

  13. [13]

    Functional Map of the World

    Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional Map of the World. InCVPR, 2018. 6, 20

    2018

  14. [14]

    Lobell, and Stefano Ermon

    Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for tem- poral and multi-spectral satellite imagery. InNeurIPS, 2022. 2

    2022

  15. [15]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. InCVPR, 2009. 2

    2009

  16. [16]

    An ecoregion-based approach to protecting half the terrestrial realm.BioScience, 2017

    E Dinerstein, D Olson, A Joshi, C Vynne, ND Burgess, E Wikramanayake, N Hahn, S Palminteri, P Hedao, R Noss, et al. An ecoregion-based approach to protecting half the terrestrial realm.BioScience, 2017. 4, 12

    2017

  17. [17]

    The global ecosystem dynamics investigation: High-resolution laser ranging of the earth’s forests and topography.Science of remote sensing, 2020

    Ralph Dubayah, James Bryan Blair, Scott Goetz, Lola Fa- toyinbo, Matthew Hansen, Sean Healey, Michelle Hofton, George Hurtt, James Kellner, Scott Luthcke, et al. The global ecosystem dynamics investigation: High-resolution laser ranging of the earth’s forests and topography.Science of remote sensing, 2020. 4

    2020

  18. [18]

    Dubayah, J

    R.O. Dubayah, J. Armston, J.R. Kellner, L. Duncanson, S.P. Healey, P.L. Patterson, S. Hancock, H. Tang, J. Bruening, M.A. Hofton, J.B. Blair, and S.B. Luthcke. GEDI L4A Footprint Level Aboveground Biomass Density, Version 2.1. ORNL DAAC, Oak Ridge, Tennessee, USA, 2022. 12

    2022

  19. [19]

    Aboveground biomass density models for NASA’s Global Ecosystem Dynamics Investigation (GEDI) lidar mission

    Laura Duncanson, James R Kellner, John Armston, Ralph Dubayah, David M Minor, Steven Hancock, Sean P Healey, Paul L Patterson, Svetlana Saarela, Suzanne Marselis, et al. Aboveground biomass density models for NASA’s Global Ecosystem Dynamics Investigation (GEDI) lidar mission. Remote Sensing of Environment, 270, 2022. 14

    2022

  20. [20]

    World continents

    Esri. World continents. 15

  21. [21]

    Feeney, B.J

    C.J. Feeney, B.J. Cosby, and D.A. Robinson et al. Multi- ple soil map comparison highlights challenges for predicting topsoil organic carbon concentration at national scale.Sci Rep 12, 1379, 2022. 7

    2022

  22. [22]

    PhilEO Bench: Evaluat- ing Geo-Spatial Foundation Models

    Casper Fibaek, Luke Camilleri, Andreas Luyts, Nikolaos Dionelis, and Bertrand Le Saux. PhilEO Bench: Evaluat- ing Geo-Spatial Foundation Models. InIGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Sympo- sium, 2024. 2, 3

    2024

  23. [23]

    Test-Time Training with Masked Autoencoders

    Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros. Test-Time Training with Masked Autoencoders. In NeurIPS, 2022. 2, 3

    2022

  24. [24]

    Google 9 Earth Engine: Planetary-scale geospatial analysis for every- one.Remote Sensing of Environment, 2017

    Noel Gorelick, Matt Hancher, Mike Dixon, Simon Ilyushchenko, David Thau, and Rebecca Moore. Google 9 Earth Engine: Planetary-scale geospatial analysis for every- one.Remote Sensing of Environment, 2017. 4

    2017

  25. [25]

    Test-time adaptation via conjugate pseudo-labels

    Sachin Goyal, Mingjie Sun, Aditi Raghunanthan, and Zico Kolter. Test-time adaptation via conjugate pseudo-labels. NeurIPS, 2022. 2, 3

    2022

  26. [26]

    Using Pre-Training Can Improve Model Robustness and Uncer- tainty

    Dan Hendrycks, Kimin Lee, and Mantas Mazeika. Using Pre-Training Can Improve Model Robustness and Uncer- tainty. InICML, 2019. 2

    2019

  27. [27]

    iNaturalist Research-grade Obser- vations

    iNaturalist contributors. iNaturalist Research-grade Obser- vations. Occurrence datasethttps://doi.org/10. 15468/ab3s5xaccessed via GBIF.org on 2025-11-10,

    2025

  28. [28]

    TerraMind: Large-Scale Generative Multi- modality for Earth Observation.ICCV, 2025

    Johannes Jakubik, Felix Yang, Benedikt Blumenstiel, Erik Scheurer, Rocco Sedona, Stefano Maurogiovanni, Jente Bosmans, Nikolaos Dionelis, Valerio Marsocci, Niklas Kopp, et al. TerraMind: Large-Scale Generative Multi- modality for Earth Observation.ICCV, 2025. 1, 2, 6, 20

    2025

  29. [29]

    Synchronizing Task Behavior: Aligning Multiple Tasks during Test-Time Training

    Wooseong Jeong, Jegyeong Cho, Youngho Yoon, and Kuk- Jin Yoon. Synchronizing Task Behavior: Aligning Multiple Tasks during Test-Time Training. InICCV, 2025. 3

    2025

  30. [30]

    GEO-Bench: Toward Foundation Models for Earth Monitoring.NIPS, 2023

    Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan David Sherwin, Hannah Kerner, Bj ¨orn L ¨utjens, Jeremy Andrew Irvin, David Dao, Hamed Alemohammad, Alexandre Drouin, Mehmet Gunturkun, Gabriel Huang, David Vazquez, Dava Newman, Yoshua Bengio, Stefano Er- mon, and Xiao Xiang Zhu. GEO-Bench: Toward Foundation Models for Earth Monitoring.NIPS, 2023. 2, 3

    2023

  31. [31]

    A high-resolution canopy height model of the Earth

    Nico Lang, Walter Jetz, Konrad Schindler, and Jan Dirk Wegner. A high-resolution canopy height model of the Earth. Nature Ecology & Evolution, 2023. 4, 18

    2023

  32. [32]

    Ge- ographical knowledge-driven representation learning for re- mote sensing images.IEEE Transactions on Geoscience and Remote Sensing, 2021

    Wenyuan Li, Keyan Chen, Hao Chen, and Zhenwei Shi. Ge- ographical knowledge-driven representation learning for re- mote sensing images.IEEE Transactions on Geoscience and Remote Sensing, 2021. 2

    2021

  33. [33]

    Jay Kuo, Georges El Fakhri, and Jonghye Woo

    Xiaofeng Liu, Zhenhua Guo, Site Li, Fangxu Xing, Jane You, C.-C. Jay Kuo, Georges El Fakhri, and Jonghye Woo. Adversarial unsupervised domain adaptation with condi- tional and label shift: Infer, align and iterate. InICCV, 2021. 2

    2021

  34. [34]

    Deep un- supervised domain adaptation: A review of recent advances and perspectives.APSIPA Transactions on Signal and Infor- mation Processing, 2022

    Xiaofeng Liu, Chaehwa Yoo, Fangxu Xing, Hyejin Oh, Georges Fakhri, Je-Won Kang, and Jonghye Woo. Deep un- supervised domain adaptation: A review of recent advances and perspectives.APSIPA Transactions on Signal and Infor- mation Processing, 2022. 2

    2022

  35. [35]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InICLR, 2017. 6

    2017

  36. [36]

    Seasonal contrast: Un- supervised pre-training from uncurated remote sensing data

    Oscar Manas, Alexandre Lacoste, Xavier Gir ´o-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Un- supervised pre-training from uncurated remote sensing data. InICCV, 2021. 2

    2021

  37. [37]

    Pangaea: A global and inclusive benchmark for geospatial foundation models

    Valerio Marsocci, Yuru Jia, Georges Le Bellier, David Kerekes, Liang Zeng, Sebastian Hafner, Sebastian Gerard, Eric Brune, Ritu Yadav, Ali Shibli, et al. Pangaea: A global and inclusive benchmark for geospatial foundation models. arXiv preprint arXiv:2412.04204, 2024. 1, 2, 3, 6

    Pith/arXiv arXiv 2024

  38. [38]

    Maxar Technologies.maxar.com. 3

  39. [39]

    Search-tta: A multimodal test-time adaptation framework for visual search in the wild

    Derek Ming, Siang Tan, Shailesh, Boyang Liu, Alok Raj, Qi Xuan Ang, Weiheng Dai, Tanishq Duhan, Jimmy Chiun, Yuhong Cao, Florian Shkurti, and Guillaume Sartoretti. Search-tta: A multimodal test-time adaptation framework for visual search in the wild. InConference on Robot Learning. PMLR, 2025. 2, 3

    2025

  40. [40]

    4M: Massively Multimodal Masked Modeling

    David Mizrahi, Roman Bachmann, O ˘guzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4M: Massively Multimodal Masked Modeling. InNeurIPS,

  41. [41]

    Digital Elevation Model V003

    NASA/METI/AIST/Japan Spacesystems and U.S./Japan ASTER Science Team. Digital Elevation Model V003. https : / / doi . org / 10 . 5067 / ASTER / ASTGTM . 003, 2018. 4

    2018

  42. [42]

    MMEarth: Explor- ing multi-modal pretext tasks for geospatial representation learning

    Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke, Serge Belongie, Christian Igel, and Nico Lang. MMEarth: Explor- ing multi-modal pretext tasks for geospatial representation learning. InECCV. Springer, 2024. 1, 2, 4, 6, 20, 21

    2024

  43. [43]

    Towards Stable Test-Time Adaptation in Dynamic Wild World

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards Stable Test-Time Adaptation in Dynamic Wild World. InICLR,

  44. [44]

    Courville

    Ethan Perez, Florian Strub, Harm de Vries, Vincent Du- moulin, and Aaron C. Courville. Film: Visual reasoning with a general conditioning layer. InAAAI, 2018. 3

    2018

  45. [45]

    Claudio Persello, Jan Dirk Wegner, Ronny H ¨ansch, Devis Tuia, Pedram Ghamisi, Mila Koeva, and Gustau Camps- Valls. Deep learning and earth observation to support the sustainable development goals: Current approaches, open challenges, and future opportunities.IEEE Geoscience and Remote Sensing Magazine, 2022. 1

    2022

  46. [46]

    Geo- plant: Spatial plant species prediction dataset

    Lukas Picek, Christophe Botella, Maximilien Servajean, C´esar Leblanc, R ´emi Palard, Th ´eo Larcher, Benjamin Deneu, Diego Marcos, Pierre Bonnet, and Alexis Joly. Geo- plant: Spatial plant species prediction dataset. InNIPS. Cur- ran Associates, Inc., 2024. 2

    2024

  47. [47]

    Planet.planet.com. 3

  48. [48]

    Poggio, L

    L. Poggio, L. M. de Sousa, N. H. Batjes, G. B. M. Heuvelink, B. Kempen, E. Ribeiro, and D. Rossiter. Soilgrids 2.0: pro- ducing soil information for the globe with quantified spatial uncertainty.SOIL, 2021. 14

    2021

  49. [49]

    Google earthtm/kml files

    Earthquake Hazards Program. Google earthtm/kml files. 15

  50. [50]

    Using Multiple Input Modali- ties can Improve Data-Efficiency and O.O.D

    Arjun Rao and Esther Rolf. Using Multiple Input Modali- ties can Improve Data-Efficiency and O.O.D. Generalization for ML with Satellite Imagery. InTerraBytes - ICML 2025 workshop, 2025. 1, 2, 7

    2025

  51. [51]

    Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning

    Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Can- dido, Matt Uyttendaele, and Trevor Darrell. Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning. InICCV, Los Alamitos, CA, USA,

  52. [52]

    Taxabind: A unified embedding space for ecological applications

    Srikumar Sastry, Subash Khanal, Aayush Dhakal, Adeel Ah- mad, and Nathan Jacobs. Taxabind: A unified embedding space for ecological applications. In2025 IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV),

  53. [53]

    Parameter efficient self-supervised geospatial domain adap- tation

    Linus Scheibenreif, Michael Mommert, and Damian Borth. Parameter efficient self-supervised geospatial domain adap- tation. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  54. [54]

    MM-TTA: Multi-Modal Test-Time Adaptation for 3D Semantic Segmentation

    Inkyu Shin, Yi-Hsuan Tsai, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Sparsh Garg, In So Kweon, and Kuk- Jin Yoon. MM-TTA: Multi-Modal Test-Time Adaptation for 3D Semantic Segmentation. InCVPR, 2022. 3

    2022

  55. [55]

    Wegner, and Konrad Schindler

    Ghjulia Sialelli, Torben Peters, Jan D. Wegner, and Konrad Schindler. AGBD: A Global-scale Biomass Dataset.ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, X-G-2025, 2025. 2

    2025

  56. [56]

    Oriane Sim ´eoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timoth´ee Darcet, Th´eo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie,...

    2025

  57. [57]

    When does self-supervision improve few-shot learning? In ECCV, Berlin, Heidelberg, 2020

    Jong-Chyi Su, Subhransu Maji, and Bharath Hariharan. When does self-supervision improve few-shot learning? In ECCV, Berlin, Heidelberg, 2020. Springer-Verlag. 2

    2020

  58. [58]

    Unsupervised domain adaptation through self-supervision

    Yu Sun, Eric Tzeng, Trevor Darrell, and Alexei A Efros. Unsupervised domain adaptation through self-supervision. arXiv preprint arXiv:1909.11825, 2019. 2, 3

    Pith/arXiv arXiv 1909

  59. [59]

    Yu Sun, Xiaolong Wang, Liu Zhuang, John Miller, Moritz Hardt, and Alexei A. Efros. Test-Time Training with Self- Supervision for Generalization under Distribution Shifts. In ICML, 2020. 2, 3

    2020

  60. [60]

    SatBird: a Dataset for Bird Species Distribution Modeling using Remote Sensing and Citizen Science Data

    M ´elisande Teng, Amna Elmustafa, Benjamin Akera, Yoshua Bengio, Hager Radi, Hugo Larochelle, and David Rolnick. SatBird: a Dataset for Bird Species Distribution Modeling using Remote Sensing and Citizen Science Data. InNIPS,

  61. [61]

    Galileo: Learning Global & Local Features of Many Remote Sens- ing Modalities

    Gabriel Tseng, Anthony Fuller, Marlena Reil, Henry Her- zog, Patrick Beukema, Favyen Bastani, James R Green, Evan Shelhamer, Hannah Kerner, and David Rolnick. Galileo: Learning Global & Local Features of Many Remote Sens- ing Modalities. InICML, 2025. 1

    2025

  62. [62]

    Tent: Fully Test-Time Adapta- tion by Entropy Minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully Test-Time Adapta- tion by Entropy Minimization. InICLR, 2021. 3

    2021

  63. [63]

    Stewart, Thomas Dujardin, Nikolaos Ioannis Bountos, Angelos Za- vras, Franziska Gerken, Ioannis Papoutsis, Laura Leal-Taix´e, and Xiao Xiang Zhu

    Yi Wang, Zhitong Xiong, Chenying Liu, Adam J. Stewart, Thomas Dujardin, Nikolaos Ioannis Bountos, Angelos Za- vras, Franziska Gerken, Ioannis Papoutsis, Laura Leal-Taix´e, and Xiao Xiang Zhu. Towards a Unified Copernicus Foun- dation Model for Earth Vision, 2025. 2, 3, 6, 7, 21

    2025

  64. [64]

    Exploring uncertainty-based self-prompt for test-time adaptation se- mantic segmentation in remote sensing images.Remote Sensing, 16(7), 2024

    Ziquan Wang, Yongsheng Zhang, Zhenchao Zhang, Zhipeng Jiang, Ying Yu, Lei Li, and Lei Zhang. Exploring uncertainty-based self-prompt for test-time adaptation se- mantic segmentation in remote sensing images.Remote Sensing, 16(7), 2024. 3

    2024

  65. [65]

    Unlocking the potential of earth observation to ad- dress africa’s critical challenges

    WEF. Unlocking the potential of earth observation to ad- dress africa’s critical challenges. Technical report, World Economic Forum, 2021. In collaboration with Digital Earth Africa. 4

    2021

  66. [66]

    Con- vNeXt V2: Co-designing and scaling convnets with masked autoencoders

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Con- vNeXt V2: Co-designing and scaling convnets with masked autoencoders. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 20

    2023

  67. [67]

    Neural Plasticity-Inspired Foundation Model for Ob- serving the Earth Crossing Modalities.arXiv preprint arXiv:2403.15356, 2024

    Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J Stew- art, Jo ¨elle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural Plasticity-Inspired Foundation Model for Ob- serving the Earth Crossing Modalities.arXiv preprint arXiv:2403.15356, 2024. 2, 20

    arXiv 2024

  68. [68]

    Test-time adaptation against multi-modal reliability bias

    Mouxing Yang, Yunfan Li, Changqing Zhang, Peng Hu, and Xi Peng. Test-time adaptation against multi-modal reliability bias. InICLR, 2024. 3

    2024

  69. [69]

    Jones, Toby Pak Tsang, Shuang Xing, Richard T

    Eugene Yu Yau, Emily E. Jones, Toby Pak Tsang, Shuang Xing, Richard T. Corlett, Patrick Roehrdanz, David J. Lohman, Adam Kai Lee, Catherine Wai Hai, and Shawan et al. Chowdhury. Spatial occurrence records and distribu- tions of tropical asian butterflies.Scientific Data, 12(1004),

  70. [70]

    Lobell, and Stefano Ermon

    Christopher Yeh, Chenlin Meng, Sherrie Wang, Anne Driscoll, Erik Rozi, Patrick Liu, Jihyeon Lee, Marshall Burke, David B. Lobell, and Stefano Ermon. SustainBench: Benchmarks for Monitoring the Sustainable Development Goals with Machine Learning. InNeurIPS, 2021. 3

    2021

  71. [71]

    Rapid Network Adaptation: Learning to Adapt Neu- ral Networks Using Test-Time Feedback

    Teresa Yeo, O ˘guzhan Fatih Kar, Zahra Sodagar, and Amir Zamir. Rapid Network Adaptation: Learning to Adapt Neu- ral Networks Using Test-Time Feedback. InICCV, 2023. 3

    2023

  72. [72]

    Zanaga, R

    D. Zanaga, R. Van De Kerchove, W. De Keersmaecker, N. Souverijns, C. Brockmann, R. Quast, J. Wevers, A. Grosu, A. Paccini, S. Vergnaud, O. Cartus, M. Santoro, S. Fritz, I. Georgieva, M. Lesiv, S. Carter, M. Herold, Linlin Li, N.E. Tsendbazar, F. Ramoino, and O. Arino. ESA WorldCover 10 m 2020 v100.doi:10.5281/zenodo.5571936, 2021. 18

    doi:10.5281/zenodo.5571936 2020

  73. [73]

    ESA WorldCover 10 m 2020 v100, 2021

    Daniele Zanaga, Ruben Van De Kerchove, Wanda De Keers- maecker, Niels Souverijns, Carsten Brockmann, Ralf Quast, Jan Wevers, Alex Grosu, Audrey Paccini, Sylvain Vergnaud, Oliver Cartus, Maurizio Santoro, Steffen Fritz, Ivelina Georgieva, Myroslava Lesiv, Sarah Carter, Martin Herold, Linlin Li, Nandin-Erdene Tsendbazar, Fabrizio Ramoino, and Olivier Arino....

    2020

  74. [74]

    NanoAd- apt: Mitigating Negative Transfer in Test Time Adaptation with Extremely Small Batch Sizes

    Shiji Zhao, Shao-Yuan Li, and Sheng-Jun Huang. NanoAd- apt: Mitigating Negative Transfer in Test Time Adaptation with Extremely Small Batch Sizes. InIJCAI. International Joint Conferences on Artificial Intelligence Organization,

  75. [75]

    Deep learning in remote sensing: A comprehensive review and list of resources.IEEE Geoscience and Remote Sensing Maga- zine, 2017

    Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources.IEEE Geoscience and Remote Sensing Maga- zine, 2017. 1 11 MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training Supplementary Material S.1. MMEarth-Bench ...

    2017

  76. [78]

    circa 100 m,

    only show AGBD values up to≈2,000. Next, we randomly shuffle the GEDI points and select however many points are needed from the beginning of the list, possibly fewer depending on how many there are. We then export the collection of points to a bucket in Google Cloud Storage in GeoJSON format. Once it has been up- loaded, we download the file. If there is ...

    2023

  77. [2020]

    degrade flag = 0,

    We then extract the geographic range of the ecoregion cropped to the GEDI range as discussed earlier. We use this Table S.5.Split ranges for regression tasks.Values are [min, max]. Units are Mg/ha for biomass and g/kg for soil N and soil OC. Split Biomass Soil N Soil OC Soil pH Train 100% [0.00, 1991.21] [0.00, 38.81] [0.00, 779.00] [3.00, 10.70] Train 50...

    1991

  78. [2023]

    2, 6, 20

    IEEE Computer Society. 2, 6, 20