ChronoEarth-492K: Spatiotemporal Hyperspectral Dataset

• ChronoEarth-492K is a large-scale, temporally deep hyperspectral Earth observation dataset designed for spatiotemporal modeling using 17 years of global data.
• It provides 492,354 radiometrically harmonized patches and multi-temporal sequences that support static, short-horizon, and long-horizon geospatial tasks.
• The associated benchmark establishes a standardized evaluation protocol for segmentation, classification, and change detection, fostering sensor transferability and spatial consistency.

ChronoEarth-492K is a large-scale, globally distributed, temporally deep hyperspectral Earth observation dataset built from NASA’s EO-1 Hyperion mission and paired with ChronoEarth-Benchmark, a standardized evaluation suite for static, short-horizon, and long-horizon spatiotemporal modeling. It comprises 492,354 radiometrically harmonized patches across 185,398 global locations over 17 years, with 28,786 sites containing multi-temporal sequences with $\geq 3$ observations, and was introduced as the first large-scale, temporally calibrated hyperspectral SSL dataset built upon EO-1 Hyperion, the world’s longest continuous hyperspectral archive up to date (2001–2017) (Si et al., 15 May 2026).

1. Definition and nomenclature

ChronoEarth-492K denotes a hyperspectral self-supervised learning dataset explicitly designed for spatiotemporal modeling rather than only static hyperspectral representation learning. Its defining properties are scale, temporal depth, and deterministic spatial indexing: 492,354 hyperspectral patches, 185,398 unique spatial locations, and 17 years of observations from May 2001 to March 2017 (Si et al., 15 May 2026). The associated ChronoEarth-Benchmark adds a unified evaluation suite spanning static prediction, short-horizon temporal aggregation, and long-horizon forecasting.

The dataset is built on EO-1 Hyperion Level-1T products. Each sample is a 128 $\times$ 128 patch at 30 m ground sampling distance with 155 harmonized spectral bands derived from 242 original Hyperion bands. The temporal dimension is central rather than incidental: 56,491 locations have at least 2 observations, and 28,786 locations have multi-temporal sequences with at least 3 timestamps, enabling both short-term dynamics and long-horizon change analysis (Si et al., 15 May 2026).

A recurrent source of confusion arises from unrelated uses of similar wording. In "AgriGPT-Omni" (Yang et al., 11 Dec 2025), the string "ChronoEarth-492K" does not appear by name; the only 492K-scale component there is an agricultural synthetic speech corpus, and its association with the term is an interpretive convenience rather than an established dataset name. In the CHRONOS detector paper (Inoue et al., 7 Apr 2026), “ChronoEarth-492K” is discussed only as an extrapolative label “interpreted in the spirit of CHRONOS,” not as an actual gravitational-wave instrument. In established usage, ChronoEarth-492K refers to the EO-1 Hyperion-based hyperspectral dataset and benchmark introduced in 2026 (Si et al., 15 May 2026).

2. Data source, preprocessing, and spatial-temporal indexing

ChronoEarth-492K is constructed from NASA’s EO-1 Hyperion mission, using Level-1T products acquired between May 1, 2001 and March 13, 2017. Hyperion provides 242 original spectral bands with continuous coverage over approximately 400–2500 nm at 30 m GSD and a swath width of about 7.7 km. The Level-1T products are terrain-corrected and radiometrically calibrated by USGS (Si et al., 15 May 2026).

Scene selection is governed by both quality and spatial diversity constraints. Only scenes with cloud coverage below 10%, as reported in metadata, are retained. To reduce ocean coverage and ensure geographic diversity, the dataset is organized into nine broad regions: Africa, North America, Southwest Asia, Latin America, East Asia, Oceania, Europe, Arctic, and Southeast Asia.

Spectral harmonization begins with the 242-band Hyperion products and applies well-established band filtering to remove unstable or low-signal bands, especially around strong water absorption and noisy ranges such as approximately 1400 nm and 1900 nm, following Datt et al. (2003). The result is a consistent 155-band representation. Because Level-1T products are already radiometrically calibrated and orthorectified, no additional atmospheric or geometric correction is applied; “radiometrically harmonized” refers primarily to filtering unstable bands and aligning all samples to a consistent wavelength grid across scenes (Si et al., 15 May 2026).

Spatial consistency is enforced through a fixed global gridding procedure within each UTM zone. Each scene is resampled to 30 m pixel size in its native UTM projection, and then divided into non-overlapping 128 $\times$ 128 patches aligned to a zone-specific grid anchored at projected coordinate $(0,0)$. Every patch receives a deterministic unique location identifier,

$\text{UID} = \text{[UTM\_zone : column : row]}.$

Patches with the same UID but different timestamps form a spatiotemporal sequence for a fixed geographic location. Up to 10% nodata pixels are allowed per patch in order to balance coverage and quality. This indexing scheme is what makes precise multi-year co-registration possible at scale (Si et al., 15 May 2026).

3. Dataset composition and temporal depth

The global statistics of ChronoEarth-492K are summarized below.

Component	Value	Description
Total patches	492,354	128 $\times$ 128 $\times$ 155
Unique locations	185,398	Global UIDs
Locations with $\geq 2$ observations	56,491	Temporal sites
Locations with $\geq 3$ observations	28,786	Multi-temporal sequences
Temporal span	2001–2017	17 years
Spatial resolution	30 m	Hyperion nominal GSD
Spectral bands	155	Harmonized from 242

Regional coverage is also explicitly reported: Africa has 26,835 locations and 105,597 patches; North America 40,583 locations and 96,259 patches; Southwest Asia 34,750 and 96,034; Latin America 22,874 and 58,565; East Asia 22,958 and 45,081; Oceania 13,693 and 42,319; Europe 12,777 and 25,430; Arctic 5,692 and 14,363; and Southeast Asia 5,236 and 8,706 (Si et al., 15 May 2026). This distribution reflects both EO-1 tasking priorities and the narrow-strip acquisition geometry of Hyperion.

The temporal structure is irregular and physically realistic rather than synthetically regularized at collection time. Most acquisition intervals between consecutive Hyperion passes at a given UID are below 1 year, which supports short-term temporal analysis. At the same time, more than one-third of multi-temporal locations have a span greater than 2 years between first and last observation, which supports long-term change analysis. This combination of short revisit intervals and multi-year sequences is one of the dataset’s distinctive properties (Si et al., 15 May 2026).

The released data are stored as GeoTIFF hyperspectral cubes with per-patch geospatial metadata in the headers. The directory structure is $(0,0)$0 where YYYYDDD encodes year and day-of-year. A global metadata table, $(0,0)$1 lists paths, UIDs, timestamps, and other metadata, enabling sequence construction by querying all samples sharing the same UID (Si et al., 15 May 2026).

4. ChronoEarth-Benchmark and the evaluation protocol

ChronoEarth-Benchmark aligns ChronoEarth-492K with external geospatial products to define downstream tasks in segmentation, multi-label classification, and change detection. The paper states that the benchmark is built from six open-source geospatial products covering land cover, crop type, forest dynamics, and soil properties, although one passage notes “seven publicly available geospatial products”; the detailed task definitions revolve around six core products (Si et al., 15 May 2026).

Product	Geography	Benchmark use
CDL	United States	Crop type segmentation
CLCD	China	Land cover segmentation
NLCD-S	United States	Land cover segmentation
ISDASoil	Africa	Multi-label soil texture classification
CORINE	Europe	Multi-label land cover classification
GFC	Global	Forest loss change detection

Location-label alignment is performed by reconstructing the same UTM-based grid for each label layer, matching labels to UIDs, and extracting corresponding label patches. Each label patch may align with multiple Hyperion observations in the same year or with none. All valid location-label pairs are retained, which permits later construction of temporal tasks (Si et al., 15 May 2026).

To remove low-information label patches, the benchmark applies entropy-based filtering. For a label map $Y \in \mathbb{N}^{H \times W}$ with at most $\times$0 classes, empirical class frequencies are defined by

$\times$1

The normalized entropy is

$\times$2

Patches are retained when $\times$3, with default $\times$4 (Si et al., 15 May 2026).

The benchmark taxonomy has three principal regimes. Static tasks use a single hyperspectral patch $\times$5 to predict a label $\times$6. Short-horizon tasks aggregate temporally adjacent observations for the same label epoch, learning

$\times$7

with sequences normalized to fixed length $\times$8 by deterministic subsampling or zero-padding and a minimum observation threshold of 2. Long-horizon tasks forecast future labels using observations up to a previous year, conceptually

$\times$9

with $\times$0 year for CDL, CLCD, and NLCD-S (Si et al., 15 May 2026).

A central feature of ChronoEarth-Benchmark is its leak-proof evaluation protocol. Spatial leakage is prevented by a distance-aware grouping strategy based on orbital swaths: patches from the same Hyperion observation are grouped into spatial units, overlapping units are merged into connected components, and each connected component is assigned to train, validation, or test. For temporal tasks, causal direction is enforced: short-horizon sequences are defined within label epochs, long-horizon tasks use only observations before the target label year, and temporal SSL pretraining uses causal masking (Si et al., 15 May 2026).

The benchmark also defines structured out-of-distribution settings. On CORINE, the authors construct ID, spatial OOD, temporal OOD, and joint spatial-temporal OOD splits by separating 2001–2012 from 2013–2017 and then applying distance-aware split generation within each temporal group. On GFC, cross-continental generalization is tested by training on Europe, North America, and East Asia and evaluating on Africa, Latin America, Oceania, and Southwest Asia (Si et al., 15 May 2026).

5. Self-supervised learning, baselines, and empirical results

ChronoEarth-492K is intended as a pretraining corpus for hyperspectral foundation models. The full 492K patches support static SSL, while the 28,786 sites with at least 3 timestamps support temporal SSL. The paper explicitly examines whether spatiotemporal pretraining improves initialization relative to supervised training from scratch, strengthens generalization to new regions and even new sensors, and yields gains from temporal context beyond static per-frame embeddings (Si et al., 15 May 2026).

The baseline suite includes hyperspectral and generalist foundation models. Static baselines are SpectralViT, LESSViT, HyperSigma, DOFA, SatMAE, and DINOv3. Temporal baselines are formed by attaching a temporal module on top of a frozen static backbone, mostly SpectralViT, using one of three strategies: max pooling over independently encoded frames, a supervised temporal attention module (“AttentionPool”), or a temporal module initialized by second-stage temporal SSL (Si et al., 15 May 2026).

Static pretraining on ChronoEarth-492K uses 200 epochs with AdamW, learning rate $\times$1, weight decay $\times$2, cosine scheduling, and a 10-epoch warmup. SpectralViT is pretrained with a masked autoencoder objective using 75% spatial patch masking; LESSViT uses Hyper-MAE with a 75% spectral mask (Si et al., 15 May 2026). Temporal SSL is then performed only on locations with at least 3 timestamps by next-frame prediction. A frozen static backbone produces token features, temporal self-attention is applied along time with temporal positional embeddings and causal masking, and a reconstruction decoder predicts the final frame. Training lasts 50 epochs with the same optimizer and schedule family, and only the temporal module plus decoder are updated (Si et al., 15 May 2026).

Downstream fine-tuning attaches task-specific heads: UPerNet for segmentation and linear heads for multi-label classification. Supervised baselines from random initialization are also included to quantify the benefits of SSL. Fine-tuning uses AdamW, a learning-rate search grid of $\times$3, batch size 32, linear warmup over the first 5% of steps, cosine decay, and weight decay $\times$4 (Si et al., 15 May 2026).

Several empirical conclusions structure the benchmark. On static tasks, SSL significantly improves over supervised training from scratch, and LESSViT generally achieves the best static performance, especially in segmentation, owing to explicit spatial-spectral modeling (Si et al., 15 May 2026). On CORINE, performance drops steadily from ID to T-OOD to S-OOD to ST-OOD, indicating that spatial shift is the main driver of degradation. On GFC, all models show substantial drops from ID to OOD continents, although SpectralViT can outperform LESSViT in sparse forest-loss settings, plausibly because strong spatial aggregation is advantageous for such signals.

Cross-satellite transfer is a notable result. SpectralViT pretrained on ChronoEarth and then fine-tuned on EnMAP-based benchmarks such as BDFORET, BNETD, EuroCrops, CORINE, and CDL achieves performance on par with or better than models trained entirely on SpectralEarth, which the paper interprets as evidence of sensor transferability from Hyperion-based pretraining (Si et al., 15 May 2026).

Temporal context yields further gains. In short-horizon tasks, moving from $\times$5 to $\times$6 generally improves performance, often by 2–5 mIoU points in segmentation settings. LESSViT with max pooling is already strong, indicating that a powerful static backbone can extract useful temporal information even under naïve aggregation. Among SpectralViT variants, temporal SSL initialization outperforms both supervised attention and max pooling on most segmentation tasks, confirming that sequence-aware pretraining improves temporal representation use (Si et al., 15 May 2026). In long-horizon tasks, increasing context from $\times$7 to $\times$8 to $\times$9 improves CLCD and NLCD-S performance, while CDL is less monotonic, suggesting that older history may introduce noise for annual crop classification.

6. Scientific role, limitations, and access

ChronoEarth-492K addresses a specific gap in hyperspectral learning: prior large datasets such as HySpecNet-11k, MSST, HyperGlobal-450K, and SpectralEarth are large but predominantly static, with little or no long-term temporal depth and at most about two years of temporal coverage, often used only for augmentation rather than temporal modeling (Si et al., 15 May 2026). ChronoEarth instead provides 17 years of calibrated, globally distributed hyperspectral time series and tens of thousands of sites with repeated observations, thereby supporting systematic research on long-horizon spatiotemporal hyperspectral models.

The benchmark’s immediate use cases include long-term ecosystem and land-cover monitoring, agricultural monitoring and crop modeling, forest disturbance studies, soil property mapping, hyperspectral foundation model development, temporal SSL, and robustness under spatial and temporal distribution shift (Si et al., 15 May 2026). The CDL tasks support crop type classification and crop rotation analysis; GFC enables forest-loss change detection and cross-continental generalization; ISDASoil connects hyperspectral response to soil texture classes; and the OOD settings on CORINE and GFC provide a controlled testbed for domain adaptation and domain generalization.

The limitations identified by the authors are methodological and data-centric. Temporal modeling is two-stage rather than end-to-end, because temporal modules are trained on top of frozen spatial backbones. Temporal sampling is sparse and irregular, which complicates learning consistent sequence representations. Label products such as ISDASoil and CORINE have coarser resolution or noisy labels relative to Hyperion, and coverage remains biased by EO-1 tasking priorities (Si et al., 15 May 2026). The future directions proposed in the paper are correspondingly direct: native spatiotemporal hyperspectral foundation models that jointly model spatial, spectral, and temporal structure; SSL objectives tailored to irregular time series; and expansion to additional sensors such as PRISMA, EnMAP, and Gaofen-5.

ChronoEarth-492K and ChronoEarth-Benchmark are distributed through the project page at https://uiuctml.github.io/ChronoEarth492K/. The release includes GeoTIFF hyperspectral cubes, benchmark label metadata stored as parquet files, label GeoTIFFs for segmentation tasks, and code for patch extraction, UID indexing, temporal sequence construction, split generation, entropy filtering, pretraining, and evaluation (Si et al., 15 May 2026). The underlying EO-1 Hyperion data and many label products are public domain or CC-BY, while ChronoEarth-492K and ChronoEarth-Benchmark are released under an open academic license for non-commercial research use. The paper recommends using the provided metadata tables, official splits, the native 30 m GSD, causal masking for temporal tasks, and appropriate per-band normalization to maintain comparability and avoid spectral bias (Si et al., 15 May 2026).