SSL4Eco: A Global Seasonal Dataset for Geospatial Foundation Models in Ecology
Elena Plekhanova, Damien Robert, Johannes Dollinger, Emilia Arens, Philipp Brun, Jan Dirk Wegner, Niklaus Zimmermann · 2025
Read SSL4Eco: A Global Seasonal Dataset for Geospatial Foundation Models in Ecology by Elena Plekhanova, Damien Robert, Johannes Dollinger, Emilia Arens, Philipp Brun, Jan Dirk Wegner, Niklaus Zimmermann (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-18256.txt.
Introduction
Seven authors build SSL4Eco in 2025, a world photo set for earth-watch base models in ecology work. Earth photos hold rich data for green study, yet tagged sets stay scarce for that work. Past sets lean to busy human zones and use fixed calendar dates, which miss local green cycles at each spot.
Their pick plan keys sample dates to plant growth turns at each site in the set. A season-split self-teach run on the set beat photo sets on seven of eight green tasks tried. The set, code, and model weights are shared for wide use in earth study.
Catalog record
| CREATOR | Elena Plekhanova, Damien Robert, Johannes Dollinger, Emilia Arens, Philipp Brun, Jan Dirk Wegner, Niklaus Zimmermann |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-18256
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
SSL4Eco: A Global Seasonal Dataset for Geospatial Foundation Models in Ecology Elena Plekhanova, Damien Robert, Johannes Dollinger, Emilia Arens, Philipp Brun, Jan Dirk Wegner, Niklaus Zimmermann Abstract With the exacerbation of the biodiversity and climate crises, macroecological pursuits such as global biodiversity mapping become more urgent. Remote sensing offers a wealth of Earth observation data for ecological studies, but the scarcity of labeled datasets remains a major challenge. Recently, self-supervised learning has enabled learning representations from unlabeled data, triggering the development of pretrained geospatial models with generalizable features. However, these models are often trained on datasets biased toward areas of high human activity, leaving entire ecological regions underrepresented. Additionally, while some datasets attempt to address seasonality through multi-date imagery, they typically follow calendar seasons rather than local phenological cycles. To better capture vegetation seasonality at a global scale, we propose a simple phenology-informed sampling strategy and introduce corresponding SSL4Eco, a multi-date Sentinel-2 dataset, on which we train an existing model with a season-contrastive objective. We compare representations learned from SSL4Eco against other datasets on diverse ecological downstream tasks and demonstrate that our straightforward sampling method consistently improves representation quality, highlighting the importance of dataset construction. The model pretrained on SSL4Eco reaches state of the art performance on 7 out of 8 downstream tasks spanning (multi-label) classification and regression. We release our code, data, and model weights to support macroecological and computer vision research at https://github.com/PlekhanovaElena/ssl4eco . (a) Spatial distribution of SSL4EO-S12 [ 89 ] (b) Spatial distribution of SSL4Eco (c) Copernicus land cover [ 55 ] distribution for SSL4Eco (upwards) and SSL4EO-S12 [ 89 ] (downwards) Figure 1 : We propose SSL4Eco, a multi-date Sentinel-2 dataset for pretraining foundation models targeted for macroecological applications. Unlike comparable datasets (a), SSL4Eco uniformly covers the entire landmass (b), thus capturing all environment types without favoring urban and agricultural areas, or ignoring entire ecoregions (c). 1 Introduction Biodiversity is essential for ecosystem stability and human well-being, yet it faces an unprecedented crisis due to habitat loss and climate change [ 8 ] . Recognized as a global priority (SDG 15) [ 63 ] , biodiversity loss ranks among the most severe risks of the next decade [ 28 ] . This intensifying crisis calls for macroecological studies to understand spatiotemporal biodiversity patterns and identify priority areas for conservation [ 8 ] . Mapping changes in biodiversity, habitats, and land-use ( e.g . deforestation, urban or agricultural expansion) over time is essential for conservation planning [ 36 , 37 ] . Central to these efforts is monitoring vegetation change, as vegetation forms the primary structure of most terrestrial ecosystems and shapes biodiversity patterns and ecosystem functions [ 85 ] . Remote sensing is a powerful tool for monitoring vegetation change at broad spatial and temporal scales [ 92 ] . It provides consistent, repeated, global observations, enabling the detection of subtle shifts in vegetation health, species composition, and phenology—insights often unattainable through ground-based methods [ 25 , 27 ] . Several open-access satellite products support vegetation monitoring, each with distinct strengths and limitations (see Appendix A-1 ). This work focuses on Sentinel-2 due to its widespread use for large-scale vegetation monitoring [ 52 , 80 , 46 ] , but our conclusions remain applicable and may be extended to other satellite products. To extract ecological insights from remote sensing data, initial approaches relied on handcrafted features and classical machine learning [ 34 , 6 ] . Deep learning has since revolutionized the field by automating feature extraction for tasks with annotated datasets [ 97 , 56 ] . Recently, self-supervised learning (SSL) has gained traction for learning rich representations from large, unlabeled datasets [ 35 ] , with successful applications in the analysis of natural language [ 19 ] , natural images [ 66 ] , and remote sensing data [ 4 ] . The resulting pretrained models produce representations that generalize to downstream tasks, making these so-called foundation models (FMs) particularly suitable for applications where labeled data is scarce or costly, such as large-scale ecological studies [ 88 ] . The size and diversity of the pretraining dataset largely influences the generalizability of the learned representations [ 72 , 96 ] . While research on geospatial FMs operating on georeferenced data (GFMs) is an active field of study, most effort is currently geared towards new model architectures and SSL pretraining tasks, and little attention is given to the design of pretraining datasets. This oversight is critical, as the geographical distribution of training data significantly influences model performance [ 74 , 71 ] . For biodiversity applications in particular, existing GFMs are often trained on datasets which fail to capture important spatiotemporal ecological patterns, as summarized in Table 1 . First, the geographic sampling is often biased towards human activity, hence over-representing urban and agricultural areas while neglecting entire biomes. Second, multi-temporal datasets are typically sampled following calendar seasons, failing to account for local phenological cycles, essential to biodiversity monitoring. In this work, we propose a dataset construction recipe targeted towards the development of foundation models for ecology. Specifically, we propose to sample locations uniformly across the landmass, rather than around large urban areas [ 58 , 89 ] , and sample dates based on local phenological cycles, rather than calendar seasons [ 58 , 89 ] . Following this protocol, we introduce SSL4Eco, a pretraining dataset of multispectral, multi-date Sentinel-2 patches of 256\times 256 pixels, uniformly sampled across 250 k locations around the globe and capturing local phenology, as shown in Figure 1 and Figure 2 . From SSL4Eco, we derive SeCo-Eco, a seasonality-aware SeCo [ 58 ] model, and compare its embeddings against off-the-shelf GFMs on diverse macroecological tasks. We show that SeCo-Eco equals or exceeds the performance of all other baselines on 7 out of 8 downstream tasks spanning (multi-label) classification and regression, with larger gaps of +2 mAP on BigEarthNet-10% [ 82 ] and +3 to +4 R 2 in regression of climatic variables and biomass. Far from claiming a new SSL training or backbone, this work stresses the importance of dataset design, and how a straightforward spatiotemporal sampling protocol may consistently benefit GFMs downstream applications. We publicly release our datasets, code, and weights at https://github.com/PlekhanovaElena/ssl4eco , hoping to foster both downstream macroecological studies and methodological computer vision research with a concern for environmental applications. The contributions of this work are as follows: • SSL4Eco: a novel multi-temporal Sentinel-2 pre-training dataset with uniform global distribution and vegetation phenology-based seasonal sampling. • SeCo-Eco: a seasonality-aware geospatial foundation model pretrained on SSL4Eco. • New macroecological downstream tasks for benchmarking geospatial foundation models. 2 Related Work In this section, we provide an overview of existing remote sensing datasets used for pretraining geospatial foundation models, with a focus on their spatiotemporal distribution. We then introduce several such foundation models relevant to this work. Dataset Locations Seasons Number Distribution BigEarthNet [ 82 ] 600 k Europe - SEN12MS [ 77 ] 280 k Around cities Calendar SeCo [ 58 ] 200 k Around cities Random S2-100k [ 49 ] 100 k Lat-lon uniform - Planted [ 68 ] 3.0 M Semi global - SatlasPretrain [ 5 ] 3.0 M Semi global - SSL4EO-S12 [ 89 ]