- The paper presents ChronoEarth-492K, the first large-scale hyperspectral dataset with long-term spatiotemporal coverage for advanced Earth observation modeling.
- It introduces a unified benchmark for static and temporal tasks with rigorous evaluation protocols, emphasizing generalization across out-of-distribution splits and cross-sensor transfers.
- Experiments reveal that self-supervised pretraining combined with explicit temporal adaptation significantly enhances the performance in land cover, crop, and forest change monitoring.
ChronoEarth-492K: A Unified Benchmark for Long-Horizon Spatiotemporal Hyperspectral Earth Observation
Motivation and Context
Hyperspectral imaging (HSI) has enabled material-level understanding of complex terrestrial and ecosystem processes by providing contiguous spectral signatures for every sensor pixel. Despite recent advances in hyperspectral self-supervised learning (SSL), model development has been constrained by the lack of large-scale datasets with sufficient temporal depth and diversity. Existing HSI datasets are mainly tailored for static scene analysis, lacking the temporal coverage required to advance long-horizon spatiotemporal modeling—a key capability for understanding trends such as land use change, crop dynamics, and forest cover trajectories. ChronoEarth-492K directly addresses this gap.
Dataset Construction and Characteristics
ChronoEarth-492K is constructed from NASA's EO-1 Hyperion mission (2001–2017), currently the longest running continuous spaceborne hyperspectral archive. The dataset comprises 492,354 radiometrically harmonized 128×128-pixel patches (30 m ground sampling distance), covering 185,398 globally distributed spatial locations. Notably, 28,786 of these locations are associated with multi-temporal sequences (≥3 time points), thus supporting both short-range aggregation and long-horizon forecasting. Spectral harmonization is achieved by discarding noisy or water absorption bands, retaining 155 precisely aligned spectral bands per observation.
The datacube's spatial diversity is ensured by partitioning the globe into nine continental-scale regions, suppressing oceanic dominance and optimizing landmass representation. For temporal consistency, geodetic patching based on UTM zone-aligned grids with stable unique identifiers (UIDs) guarantees precise spatial-temporal correspondence across years and satellites' variable acquisitions.
Evaluation Protocol: ChronoEarth-Benchmark
Built on ChronoEarth-492K, the associated ChronoEarth-Benchmark provides a unified suite of tasks for systematic spatiotemporal representation evaluation. It leverages six (plus one for generalization) open-access geospatial products, spanning land cover, crop type, forest change detection, and soil class labels. This enables coverage across segmentation, multi-label classification, and change detection tasks, and supports both temporal and spatial out-of-distribution (OOD) evaluations.
Task Taxonomy
- Static Tasks: Per-year segmentation/classification from a single HSI observation.
- Short-Horizon Temporal Tasks (SH): Multi-frame aggregation, predicting labels from multiple temporally adjacent observations with identical labels.
- Long-Horizon Temporal Tasks (LH): Forecasting, i.e. predicting label states at time t+n using a sequence of historical observations.
- Generalization Settings: Controlled OOD splits for spatial, temporal, and continental shift are defined, enabling rigorous evaluation under realistic domain transfer scenarios.
Notably, split design employs a distance-aware grouping aligned with orbital swaths to minimize label leakage and enforce strong geographic OOD settings.
Baseline Models and Temporal Extensions
ChronoEarth-Benchmark evaluates multiple state-of-the-art (SOTA) HSI foundation models including SpectralViT, HyperSigma, DOFA, LESSViT, SatMAE, and vision foundation models like DINOv3, all adapted for high-bandwidth input. For temporal tasks, baseline backbones are extended via three approaches:
- Max Pooling: Nonparametric aggregation over frame-level features.
- AttentionPool: Trainable attention module aggregating temporal context post-encoding.
- Temporal SSL: Two-stage pipeline, leveraging frozen spatial backbone with a trainable temporal head, trained through self-supervised next-frame prediction using masked, causal attention.
This approach isolates temporal modeling effects while holding spatial encoding constant.
Empirical Results
Self-supervised pretraining on ChronoEarth consistently outperforms supervised training across tasks. LESSViT demonstrates the highest static segmentation accuracy, reflecting benefits from explicit spatial-spectral interaction modeling. Generalization tasks on CORINE (Europe) and GFC (global forest change) show controlled, progressive accuracy drop as models traverse from in-distribution to spatial-, temporal-, and joint OOD splits. Notably, transfer to continental OOD settings (train on Europe/NA/East Asia, test on Africa/Latin America/Oceania/SW Asia) reveals significant performance gaps, particularly for models requiring dense annotation.
Cross-Satellite Transfer
Pretraining on ChronoEarth achieves competitive downstream performance on EnMAP-based datasets (e.g., SpectralEarth), despite differing sensor characteristics, demonstrating robust cross-sensor generalization powered by long-term, diverse observation sequences.
Temporal Modeling Insights
- Short-Horizon Tasks: Aggregating temporally adjacent frames yields consistent performance improvements; LESSViT shows the highest gains using naive max pooling, indicating that powerful static backbones can exploit temporal context even with simple methods.
- Long-Horizon Tasks: Increasing context length benefits land cover and land use segmentation (CLCD, NLCD-S), with the temporal SSL variant of SpectralViT outperforming pooling and attention baselines. However, in crop classification (CDL), performance gains plateau or degrade with long input horizons, suggesting that outdated historical context may introduce label noise or confounding due to rapid land use changes.
A strong, consistent result is that temporally pretrained models (via self-supervised objectives) provide additive benefits over both static pretraining and supervised temporal attention.
Practical and Theoretical Implications
ChronoEarth-492K and its benchmark establish a new standard for temporally aware HSI modeling by enabling end-to-end evaluation of spatial, spectral, and temporal generalization. The ability to evaluate both short-term aggregation and long-horizon forecasting under realistic, split-based OOD scenarios is particularly useful for remote sensing tasks where geographic and inter-annual domain shift dominates downstream utility.
The results indicate that current generation HSI foundation models benefit from both massive scale static pretraining and explicit temporal adaptation, but there is substantial headroom, especially in OOD robustness and long-term forecasting. The empirical evidence suggests that simple pooling is insufficient for fine-grained temporal dependencies and that the quality of temporal adaptation is highly contingent on the baseline backbone's capacity.
Limitations and Future Directions
A primary limitation is that temporal modeling is implemented as a separate stage atop static spatial backbones, rather than performed in an end-to-end, native spatiotemporal manner. Additionally, temporal observation sparsity and irregularity, inherent to orbital HSI archives, introduce challenges for deriving uniformly consistent temporal features. Future research should target:
- Architectures integrating spatial, spectral, and temporal dynamics natively
- Self-supervised objectives that can disentangle long-term trend from sensor noise and environmental variability
- Learning paradigms robust to variable and sparse observation intervals
Addressing these avenues is vital for operational deployment in climate monitoring, agricultural analysis, and forecasting frameworks.
Conclusion
ChronoEarth-492K provides the first large-scale, globally diverse, and temporally calibrated HSI dataset supporting rigorous spatiotemporal SSL, alongside a unified and realistic benchmark suite. Empirical analyses establish that large-scale pretraining confers consistent benefits in downstream performance and generalization, while explicit temporal adaptation via self-supervised learning yields further gains, particularly for long-horizon predictive tasks. ChronoEarth-492K is positioned to serve as a foundational resource for the development, evaluation, and deployment of temporally grounded hyperspectral foundation models in Earth observation applications (2605.15666).