SSL4Eco: A Global Seasonal Dataset for Geospatial Foundation Models in Ecology

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.

An Evaluation of \DATAx for Pretraining Geospatial Foundation Models in Ecology

The paper introduces \DATAx, a novel global, phenology-informed, seasonal dataset designed specifically for pretraining geospatial foundation models to assist macroecological studies. The primary endeavor is to address biases commonly found in existing datasets, such as overrepresentation of human-centric areas and ignorance of significant ecological zones. This study is grounded in the need for more accurate representation capturing the spatiotemporal dynamics essential for understanding and mitigating biodiversity crises.

Research Context and Dataset Construction

In the context of remote sensing and Earth observation, datasets are often biased towards regions with high human activity, such as urban and agricultural zones. These biases detract from the potential of geospatial foundation models to generalize to diverse ecological regions, which are critical for assessing biodiversity changes and environmental impact. Furthermore, typical temporal samplings of datasets fail to adapt to local phenological cycles, instead opting for calendar-based seasonal definitions.

The \DATAx dataset comprises multi-date Sentinel-2 satellite imagery, capturing global landmass with a sampling strategy that prioritizes phenology-informed local seasons. This sampling strategy diverges from prior approaches by utilizing the Enhanced Vegetation Index (EVI) trajectories to inform seasonal divisions. Consequently, \DATAx encompasses extensive ecological diversity, covering underserved regions such as tropical and Arctic biomes more comprehensively.

Model Pretraining and Evaluation

The authors pretrained the \SECO model using the \DATAx dataset with the seasonal contrastive learning framework, allowing for nuanced learning of both season-invariant and seasonal-specific representations. The results are benchmarked against a series of alternative geospatial foundation models (GFMs), such as SatMAE and SSL4EO, across numerous ecological and computer vision tasks.

\SECO's embeddings achieved state-of-the-art performance in 7 out of 8 diverse downstream tasks, notably improving regression results for BioMassters and surface-level climatic variables, such as temperature and evapotranspiration based on CHELSA data with improvements in R( ^2 ) scores by up to $+4.8$. These results emphasize the importance of dataset design, particularly global spatial uniformity and phenology-awareness in data sampling.

Discussion and Implications

The implications of this study are profound for ecological analytics. By more accurately reflecting the biodiversity and seasonal variations of global ecosystems, \DATAx affirms the critical role of dataset construction in enhancing model versatility and performance across tasks that simulate real-world ecological challenges. As the framework evolves, further exploration and integration of modalities, such as SAR and LiDAR, may yield even richer datasets for broader applications. Scaling this strategy to multimodal inputs could enable new surveillance capabilities across ecological sciences and biodata analytics.

Overall, the research outlines a comprehensive strategy to improve model generalization and predictive accuracy in environmental monitoring and ecological modeling. Given the rising importance of assessing biodiversity under climate change pressures, the advancement made by \DATAx represents a significant step forward in utilizing AI for ecological conservation efforts. Future developments may explore extending the methodology to different datasets and modalities, fostering multidisciplinary collaborations and applications spanning broader ecological and environmental arenas.