ChronoEarth-Benchmark: Evaluation Suite

• ChronoEarth-Benchmark is a unified evaluation suite for spatiotemporal hyperspectral representation learning with clearly defined static, short-horizon, and long-horizon regimes.
• It leverages a deterministic UID-based gridding of 492K hyperspectral patches from NASA EO-1 Hyperion data spanning global observations from 2001 to 2017.
• The benchmark rigorously separates spatial-spectral encoding from temporal modeling, providing standardized protocols across six diverse open-source geospatial products.

ChronoEarth-Benchmark is a unified evaluation suite for spatiotemporal hyperspectral representation learning built on top of ChronoEarth-492K, a temporally calibrated archive derived from NASA’s EO-1 Hyperion mission. It was introduced to address the absence of a standardized, large-scale, temporally grounded benchmark for hyperspectral self-supervised learning and temporal modeling, and it organizes evaluation into static, short-horizon temporal, and long-horizon temporal regimes using six open-source geospatial products covering land cover, crop type, forest dynamics, and soil properties (Si et al., 15 May 2026).

1. Data foundation and temporal calibration

ChronoEarth-Benchmark is inseparable from ChronoEarth-492K, the underlying hyperspectral corpus from which its temporal structure is derived. ChronoEarth-492K contains 492,354 patches from 185,398 unique global locations over 2001 to 2017, with 56,491 locations having at least two observations and 28,786 locations having three or more timestamps. The source sensor is NASA EO-1 Hyperion Level-1T terrain-corrected scenes, described as the world’s longest continuous spaceborne hyperspectral archive to date, with more than 240 narrow spectral bands, spectral range roughly 400–2500 nm, and 30 m spatial resolution (Si et al., 15 May 2026).

Patch construction follows a deterministic geospatial indexing pipeline. Each Level-1T scene is kept in its native UTM projection, resampled to a uniform 30 m pixel size, divided into non-overlapping $128 \times 128$ patches, and aligned to a fixed global grid within its UTM zone. Each patch receives a deterministic UID of the form [\text{UTM\_zone} : \text{column} : \text{row}], and patches sharing the same UID across dates are treated as repeated observations of the same location. This UID-based gridding is the benchmark’s core mechanism for temporal sequence formation (Si et al., 15 May 2026).

The benchmark inherits the corpus preprocessing decisions of ChronoEarth-492K. Only Level-1T acquisitions with reported cloud coverage below 10% were retained. Hyperion’s original 242 bands were filtered to 155 spectrally consistent bands following Datt et al. (2003), with unstable and low-signal bands removed. The retained samples are released as $128 \times 128 \times 155$ GeoTIFF cubes, and the archive is organized hierarchically as dataset/<Region>/<UID>/<UID>_<YYYYDDD>.TIF, accompanied by dataset/metadata.parquet (Si et al., 15 May 2026).

A notable design choice is that spatial consistency is achieved through fixed-grid patching rather than learned co-registration. The paper states that Level-1T products are already orthorectified and radiometrically calibrated by USGS, so no additional atmospheric corrections and no additional geometric corrections were applied. This suggests that temporal alignment is benchmarked under realistic archive irregularities rather than under a heavily normalized remote-sensing preprocessing regime (Si et al., 15 May 2026).

2. Temporal regimes and benchmark logic

ChronoEarth-Benchmark is explicitly partitioned into three temporal regimes: Static, Short-Horizon temporal (SH), and Long-Horizon temporal (LH). This partition is the benchmark’s central methodological contribution, because it distinguishes spatial-spectral representation quality from temporal aggregation and from explicit future prediction (Si et al., 15 May 2026).

In the static regime, a sample pairs a label map or label vector with one spatially aligned hyperspectral observation from the same year. This evaluates representation learning without temporal aggregation. In the short-horizon regime, the model receives multiple temporally adjacent observations corresponding to the same label; SH is therefore defined by temporal aggregation under fully observed supervision for the target year, not by a fixed interval in days or months. In the long-horizon regime, the task becomes explicit future prediction: observations from 2001 to year $n$ are used to predict labels at year $n+1$ (Si et al., 15 May 2026).

Sequence formation is UID-centric. All patches with the same UID across timestamps are grouped into temporal sequences, then filtered with a minimum observation threshold of 2 for benchmark temporal evaluation. Sequences are normalized to a fixed length $T$ by sub-sampling or zero-padding, and evaluation uses deterministic sequence fixing for fair comparison. The paper also notes that the archive is sparse, non-uniform, irregularly sampled: most acquisition intervals are under one year, while over one-third of multi-temporal locations span more than two years. The benchmark is therefore designed for irregular temporal archives rather than dense revisit schedules (Si et al., 15 May 2026).

The temporal decomposition is reflected in the evaluation settings. SH experiments compare $T=1$ against $T \le 4$. LH experiments use historical context lengths $T \le 2$, $T \le 4$, and $T \le 8$. This structure makes the benchmark useful for separating three failure modes: insufficient static spatial-spectral encoding, inability to exploit nearby temporal context, and inability to convert historical observations into future-label prediction (Si et al., 15 May 2026).

3. Label products, task inventory, and alignment rules

ChronoEarth-Benchmark integrates six enumerated geospatial products, although one sentence in the paper refers to “seven publicly available geospatial products,” a reporting inconsistency explicitly noted in the source synthesis. The actual datasets listed in the tables and appendices are GFC, ISDASoil, CDL, CORINE, NLCD-S, and CLCD (Si et al., 15 May 2026).

The benchmark first constructs a dictionary of ChronoEarth UIDs independent of time, then reconstructs the corresponding UTM-zone grid for each external label layer and matches UIDs to extract overlapping labels. Temporal alignment rules vary by product: same-year alignment for annual products in static tasks, same-label-year multiple observations for SH, history through year $128 \times 128 \times 155$0 to predict year $128 \times 128 \times 155$1 for LH, nearest-epoch mapping for CORINE, geographic-overlap-only for ISDASoil, and between-time change labels for GFC (Si et al., 15 May 2026).

Product	Task type	Benchmark role
GFC	Temporal change detection	17,214 paired samples from 319 locations; retains patches with >10% changed pixels
ISDASoil	Static and SH multi-label classification	17,808 pairs from 5,502 locations; uses intersecting soil texture classes
CDL	Static segmentation, SH, LH	10,162 pairs from 2,864 locations; final taxonomy 15 agricultural classes + 1 background
CORINE	Static multi-label classification; spatial/temporal generalization	19,774 pairs from 9,706 locations; 44 classes mapped to 19 categories
NLCD-S	Static segmentation, SH, LH	55,476 pairs from 24,662 locations; 16 thematic classes
CLCD	Static segmentation, SH, LH	17,614 pairs from 9,752 locations; nine major land-cover types

For segmentation products, the benchmark applies entropy filtering to remove effectively homogeneous patches. The paper gives the normalized entropy as

$128 \times 128 \times 155$2

with threshold $128 \times 128 \times 155$3, and samples below this threshold are removed (Si et al., 15 May 2026).

This task construction yields a heterogeneous downstream suite: static segmentation, static multi-label classification, short-horizon temporal aggregation, long-horizon future prediction, spatial-temporal generalization, continental generalization, and forest change detection. A plausible implication is that ChronoEarth-Benchmark is not merely a temporal benchmark, but a staged test of whether hyperspectral representations remain useful under increasingly realistic temporal and distribution-shift constraints (Si et al., 15 May 2026).

4. Evaluation protocol and model families

ChronoEarth-Benchmark uses a distance-aware grouping strategy based on EO-1 orbital swaths to prevent spatial leakage. Patches from the same observation are grouped into spatial units, overlapping units are merged into connected components, and train/validation/test are assigned at the component level. The stated aim is to ensure no overlapping spatial regions across splits, no patch-level leakage, and geographically distinct test partitions. The benchmark therefore defines an explicitly spatial OOD setting rather than a random split (Si et al., 15 May 2026).

Evaluation metrics are task-specific: mIoU for segmentation, mAP for multi-label classification, and F1 in addition to mIoU for GFC. The paper does not provide formal mathematical definitions of these metrics, but it standardizes downstream evaluation by using a linear classification head for multi-label classification and UPerNet for segmentation. Comparability is further enforced by using all models at ViT-Base scale, deterministic temporal sequence fixing at evaluation time, common optimizer settings, and validation-based hyperparameter tuning (Si et al., 15 May 2026).

The static baselines are SpectralViT, HyperSigma, DOFA, LESSViT, SatMAE (adapted), and DINOv3 (adapted). For temporal evaluation, the paper compares three temporal adaptation strategies: Max Pooling, AttentionPool, and Temporal SSL. Beyond max pooling, the temporal module is evaluated only with SpectralViT (Si et al., 15 May 2026).

The second-stage temporal pretraining protocol uses the 28,786 locations with at least three timestamps. Given a sequence $128 \times 128 \times 155$4, the model predicts $128 \times 128 \times 155$5 from $128 \times 128 \times 155$6 using causal masking and an $128 \times 128 \times 155$7 reconstruction loss in pixel space. The encoder is frozen, and only the temporal module plus reconstruction decoder are trained for 50 epochs. The token tensor is defined as $128 \times 128 \times 155$8, then reshaped to $128 \times 128 \times 155$9 for temporal attention over the $n$0 dimension (Si et al., 15 May 2026).

Downstream optimization is standardized: AdamW, learning-rate sweep $n$1, batch size 32, 5% warmup, cosine annealing, and weight decay $n$2. This standardization is central to the benchmark’s function as a comparative evaluation suite rather than a collection of loosely comparable datasets (Si et al., 15 May 2026).

5. Empirical behavior and benchmark discriminativeness

The benchmark produces distinct rankings across static, short-horizon, and long-horizon regimes. On static tasks, LESSViT is the strongest model on the major segmentation datasets: 54.84 mIoU on CLCD, 23.91 on CDL, and 33.59 on NLCD-S, compared with 46.04, 19.38, and 32.47 for SpectralViT. On ISDASoil, SpectralViT reaches 57.70 mAP, while the supervised SpectralViT baseline is slightly higher at 57.79, one of the few cases where SSL does not clearly dominate (Si et al., 15 May 2026).

The benchmark’s generalization settings expose a more nuanced picture. On CORINE, performance degrades from ID to T-OOD to S-OOD to ST-OOD; the paper highlights that spatial shift causes the largest drop, making geographic variability the main generalization bottleneck. On GFC, SpectralViT outperforms the alternatives with 19.29 mIoU / 32.34 F1 in-domain and 16.13 / 27.78 OOD. The authors attribute this to sparse change supervision favoring stronger spatial aggregation (Si et al., 15 May 2026).

In the short-horizon setting, adding adjacent temporal observations usually helps. Representative results include LESSViTmax on CLCD, improving from 51.19 at $n$3 to 54.69 at $n$4, and on CDL, from 24.80 to 26.15. Among controlled SpectralViT variants, Temporal SSL > AttentionPool > Max Pooling on most segmentation tasks, while ISDASoil benefits less from temporal SSL. The paper links this to the reconstruction objective being less aligned with classification via the global token (Si et al., 15 May 2026).

The long-horizon regime is the benchmark’s most distinctive evaluative layer. On CLCD, SpectralViTtemporal is best at all tested context lengths, reaching 43.61 / 51.26 / 55.61 for $n$5, $n$6, and $n$7. On NLCD-S, the same model reaches 37.34 / 39.69 / 40.76, substantially ahead of the alternatives. The main exception is CDL, where crop-type forecasting is unstable: some models degrade with longer context, and the paper suggests that older observations can become stale or harmful for rapidly changing agricultural targets (Si et al., 15 May 2026).

These results make ChronoEarth-Benchmark discriminative in a specific technical sense. Strong static hyperspectral encoders are not automatically strong temporal predictors; nearby temporal aggregation is easier than future prediction; and temporal self-supervision is most beneficial when the downstream target evolves slowly enough that historical context remains informative. This suggests that the benchmark succeeds in isolating long-horizon temporal modeling as a distinct capability rather than as a trivial extension of static spectral-spatial representation learning (Si et al., 15 May 2026).

6. Position within Earth benchmark research and known limitations

ChronoEarth-Benchmark occupies a distinct position within the recent Earth-science benchmark landscape. Unlike MSEarth, which targets graduate-level multimodal scientific comprehension from figures and refined captions, ChronoEarth-Benchmark is a hyperspectral representation-learning benchmark with explicit static/SH/LH temporal protocols rather than a figure-centered MLLM reasoning benchmark (2505.20740). Unlike OmniEarth-Bench, which evaluates multimodal reasoning across six Earth spheres and cross-sphere interactions, ChronoEarth-Benchmark does not define perception, general reasoning, scientific-knowledge reasoning, or CoT tiers; its core abstraction is temporally calibrated remote-sensing sequences linked to downstream geospatial products (Wang et al., 29 May 2025). Unlike EarthSE, which assesses LLM-based Earth scientific exploration through QA and multi-turn dialogue, ChronoEarth-Benchmark is not a literature-grounded language benchmark and does not evaluate open-ended methodology induction or limitation analysis (Xu et al., 22 May 2025). It is closer in spirit to ChaosBench in its emphasis on long-horizon temporal difficulty, but the two benchmarks operate on different substrates: ChaosBench evaluates global subseasonal-to-seasonal forecasting on gridded geophysical fields, whereas ChronoEarth-Benchmark evaluates irregularly sampled hyperspectral patch sequences under fine-tuning protocols (Nathaniel et al., 2024). A further distinction can be made from throughput-oriented Earth-system benchmarking such as global fully coupled kilometer-scale ICON runs, where the benchmark objective is exascale simulation throughput rather than spatiotemporal hyperspectral representation quality (Klocke et al., 3 Nov 2025).

The benchmark’s limitations are explicitly discussed. Temporal modeling is implemented as an extension on top of pretrained spatial backbones, not as a native end-to-end spatiotemporal architecture. The observations are sparse and irregular, Hyperion’s narrow swath produces partial coverage and nodata handling requirements, and regional sample counts are uneven. Label construction also introduces approximation and noise: CORINE uses nearest-epoch mapping, ISDASoil assumes soil texture is stable during 2001–2017, class imbalance is substantial in several products, and crop forecasting is sensitive to stale temporal context (Si et al., 15 May 2026).

Two further caveats shape its interpretation. First, the benchmark paper contains a small but consequential reporting inconsistency between six and seven source products; the enumerated products are six. Second, while reproducibility is emphasized through release of the processing pipeline, deterministic UID indexing, temporal sequence construction, and split generation, some exact split counts and implementation specifics remain in the appendices. The assembled ChronoEarth dataset and benchmark are to be released under an open academic license permitting non-commercial research use, which situates the benchmark as a research infrastructure artifact as much as a leaderboard instrument (Si et al., 15 May 2026).