ChronoEarth-492K: A Large Scale and Long Horizon Spatiotemporal Hyperspectral Earth Observation Dataset and Benchmark
Haozhe Si, Yuxuan Wan, Yuqing Wang, Minh Do, Han Zhao · 2026
Read ChronoEarth-492K: A Large Scale and Long Horizon Spatiotemporal Hyperspectral Earth Observation Dataset and Benchmark by Haozhe Si, Yuxuan Wan, Yuqing Wang, Minh Do, Han Zhao (2026) 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-2605-15666.txt.
Introduction
Haozhe Si, Yuxuan Wan, Yuqing Wang, Minh Do, and Han Zhao build ChronoEarth-492K, a hyperspectral data set for long-range Earth study. Past sets stay shallow in time, which blocks models of change over years. Their set rests on NASA's EO-1 Hyperion archive, the longest unbroken hyperspectral record, from 2001 to 2017. It holds 492,354 tuned patches at 185,398 sites over 17 years. A linked benchmark spans static, short-range, and long-range tasks from six open geospatial products.
Catalog record
| CREATOR | Haozhe Si, Yuxuan Wan, Yuqing Wang, Minh Do, Han Zhao |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61725 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.15666
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ChronoEarth-492K: A Large Scale and Long Horizon Spatiotemporal Hyperspectral Earth Observation Dataset and Benchmark Haozhe Si, Yuxuan Wan, Yuqing Wang, Minh Do, Han Zhao Abstract Hyperspectral imaging (HSI) provides dense spectral information for the Earth’s surface, enabling material-level understanding of land cover and ecosystem dynamics. Despite recent progress in hyperspectral self-supervised learning (SSL), existing datasets remain temporally shallow, limiting the development of long-horizon spatiotemporal modeling. To address this gap, we introduce ChronoEarth-492K, the first large-scale, temporally calibrated hyperspectral SSL dataset built upon NASA’s EO-1 Hyperion mission, the world’s longest continuous hyperspectral archive up to date (2001–2017). ChronoEarth-492K comprises 492,354 radiometrically harmonized patches across 185,398 global locations over 17 years, with 28,786 sites containing multi-temporal sequences ( ≥ \geq 3 observations) that enable both short- and long-horizon temporal analysis. Building on this foundation, we establish the ChronoEarth-Benchmark, a unified evaluation suite spanning static, short-horizon, and long-horizon temporal tasks, constructed from six open-source geospatial products covering land cover, crop type, forest dynamics, and soil properties. We further introduce a standardized evaluation protocol and report extensive baseline results across state-of-the-art hyperspectral foundation models. Together, ChronoEarth and benchmark provide the first large-scale, temporally grounded platform for systematic spatiotemporal hyperspectral representation learning. The code and latest version of the dataset are available through the project page: https://uiuctml.github.io/ChronoEarth492K/ . 1 Introduction Earth observation (EO) is essential for understanding global environmental dynamics, monitoring human activity impacts, and guiding sustainable development. Among various remote sensing modalities, hyperspectral imaging (HSI) acquires hundreds of contiguous spectral bands for each pixel, providing rich spectral–spatial information that enables material-level characterization of the Earth’s surface ( 18 ) . This fine spectral granularity enables precise discrimination of vegetation species, soil composition, and agricultural conditions, exceeding the descriptive capacity of multispectral sensors that typically contain fewer than twenty bands. Over the past two decades, spaceborne HSI missions ( 25 ; 21 ; 12 ; 20 ) have established orbital hyperspectral observation at tens-of-meters resolution with global coverage. As these archives expand in volume and temporal continuity, they increasingly support analysis beyond spatial–spectral representation toward temporal modeling, enabling the study of deforestation, land-use transitions, and long-term environmental change. Despite the rich spectral capacity of HSI, existing large-scale datasets are predominantly designed for static scene understanding rather than spatiotemporal modeling. Recent works ( 10 ; 26 ; 35 ; 3 ) have leveraged unlabeled HSI, primarily from EnMAP ( 12 ) , to develop self-supervised geospatial foundation models. While these efforts successfully capture global spatial variability, the limited temporal span of current HSI corpora restricts their ability to model long-term surface dynamics and, more importantly, prevents systematic evaluation of temporal generalization and forecasting capabilities. Consequently, spatiotemporal hyperspectral foundation modeling remains underexplored. Figure 1 : Global Distribution of the ChronoEarth dataset. The map illustrates the world-wide coverage of hyperspectral observations included in our dataset, highlighting its extensive spatial diversity across major continental regions. To address this limitation, we introduce ChronoEarth-492K, a large-scale, temporally calibrated hyperspectral dataset built upon NASA’s EO-1 Hyperion mission (2001–2017) ( 25 ) , the longest continuous spaceborne hyperspectral archive to date. ChronoEarth-492K comprises 492,354 radiometrically harmonized patches across 185,398 global locations ( Figure 1 ), with 28,786 locations containing multi-temporal sequences ( \geq 3 frames). This temporal depth enables both short-horizon aggregation and long-horizon prediction, providing a foundation for systematic spatiotemporal representation learning. Building on this dataset, we introduce ChronoEarth-Benchmark, a unified evaluation suite for systematic assessment of spatial representation learning and temporal modeling. The benchmark integrates six open-source geospatial products, including land-cover ( 8 ; 38 ) , crop ( 9 ; 2 ) , forest ( 13 ) , and soil datasets ( 16 ) . These datasets are carefully aligned with ChronoEarth observations. We define static, short-horizon, and long-horizon temporal tasks, enabling systematic evaluation of both spatial representation learning and temporal modeling. In addition, we design spatial-temporal and continental generalization settings, assessing the model robustness under domain shifts. Together, these components establish the first standardized platform for developing and evaluating temporally aware hyperspectral foundation models. This design enables controlled analysis of temporal aggregation, forecasting, and generalization under realistic distribution shifts. We evaluate a range of state-of-the-art hyperspectral foundation models ( 3 ; 37 ; 6 ; 33 ; 28 ) on both static and temporal tasks. For temporal evaluation, we extend static architectures with multiple temporal adaptation strategies, including learnable temporal modeling modules and non-parametric aggregation methods (e.g., max pooling), enabling consistent comparison across short-horizon and long-horizon settings. Extensive experiments demonstrate that ChronoEarth-Benchmark effectively differentiates model capabilities under varying temporal configurations and distribution shifts, demonstrating that the benchmark differentiates model capabilities under varying temporal configurations and distribution shifts. In summary, our contributions are threefold: • We present ChronoEarth-492K, the first large-scale hyperspectral SSL dataset offering 17 years of temporally calibrated global observations; • We introduce ChronoEarth-Benchmark, a unified evaluation suite comprising static, short-horizon, and long-horizon temporal tasks, enabling systematic assessment of spatiotemporal hyperspectral representation learning; • We provide extensive empirical evaluations of state-of-the-art hyperspectral foundation models, establishing strong reference baselines and demonstrating the effectiveness of temporal pretraining and modeling under diverse settings. 2 Related Works Geospatial Datasets. Large-scale geospatial representation learning has been primarily driven by multispectral imagery (MSI), which captures reflected radiance in a limited set of discrete spectral bands (typically 10–13) from platforms such as Sentinel-2 and Landsat. A number of self-supervised learning (SSL) datasets and benchmarks ( 19 ; 28 ) have been developed for MSI, covering both pretraining corpora ( 34 ; 23 ; 29 ) and standardized downstream tasks across classification ( 29 ; 15 ; 40 ) and segmentation ( 8 ; 2 ; 17 ; 39 ) . Despite their success, MSI datasets typically provide only 10–13 spectral bands, limiting spectral granularity and material-level characterization. Hyperspectral datasets provide substantially denser spectral sampling ( > 100 contiguous bands) and have recently been curated to support self-supervised representation learning ( 10 ; 26 ; 35 ; 33 ; 31 ; 5 ) . These datasets are primarily designed for static spatial representation learning, without sustained temporal coverage. SpectralEarth ( 3 ) is the only recent effort incorporating temporal observations, providing approximately two years of EnMAP ( 12 ) data and leveraging temporal views as augmentation for contrastive pretraining. However, its design does not explicitly support long-horizon temporal modeling or systematic spatiotemporal evaluation. Figure 2 : Global Patching and Spatial Alignment for ChronoEarth. (a) Raw EO-1 Hyperion observation. (b) Global patching using a UTM-zone–specific gridding system, where each patch is assigned a unique spatial identifier based on its zone and grid coordinates. (c) Spatial alignment of patches sharing the same identifier across multiple observations. (d) Formation of spatiotemporal patch sequences by aggregating aligned samples from different timestamps. 3 ChronoEarth-492K Data Source. ChronoEarth-492K is constructed from NASA’s EO-1 Hyperion mission. From all Level-1T terrain-corrected scenes available within this temporal span, we curate a globally distributed hyperspectral corpus designed specifically for large-scale spatiotemporal and hyperspectral self-supervised learning. To ensure spatial and radiometric consistency, we retained only acquisitions with reported cloud coverage below 10%. To mitigate excessive oceanic coverage and balance geographic diversity, we manually defined nine continental-scale regions (Africa, Arctic, Europe, East Asia, Latin America, Oceania, North America, Southeast Asia, Southwest Asia), spanning the major inhabited landmasses, from