GreenHyperSpectra: Global Hyperspectral Pretraining

• GreenHyperSpectra is a comprehensive hyperspectral resource aggregating over 140,000 real-world canopy spectra across sensors and ecosystems.
• The dataset leverages advanced pretraining methods, including masked autoencoders, to improve multi-trait regression accuracy and label efficiency.
• It establishes a robust benchmark for evaluating models under both in-distribution and out-of-distribution settings, enhancing ecological trait prediction.

GreenHyperSpectra is a large, multi-source hyperspectral pretraining resource introduced by Cherif, Ouaknine, Brown, Dao, Kovach, Lu, Mederer, Feilhauer, Kattenborn, and Rolnick for global prediction of plant functional traits under label scarcity and domain shift. It aggregates real-world at-surface canopy reflectance spectra across sensors, platforms, geographies, and ecosystems, and couples this corpus to a standardized benchmarking protocol for in-distribution and out-of-distribution evaluation. In the reported experiments, semi- and self-supervised pretraining on GreenHyperSpectra, especially with masked autoencoders, improves multi-trait regression accuracy and label efficiency relative to a strong supervised baseline (Cherif et al., 9 Jul 2025).

1. Problem setting and scientific scope

GreenHyperSpectra is motivated by a core difficulty in ecological remote sensing: plant functional traits are essential variables for biodiversity monitoring and climate-change science, yet conventional field sampling cannot feasibly cover trait variation at ecologically meaningful spatial scales. The target variables include leaf mass per area, chlorophyll, nitrogen or protein, carbon-based constituents, water content, carotenoids, anthocyanins, and leaf area index, all of which influence productivity, carbon cycling, stress responses, and ecosystem functioning (Cherif et al., 9 Jul 2025).

The dataset is designed around three obstacles that recur in hyperspectral trait prediction. The first is label scarcity: trait labels require field sampling and laboratory analysis, and are therefore expensive, slow, and inconsistent across studies and protocols. The second is distribution shift: models trained in limited regions or with limited sensors often fail across changes in spectral resolution, calibration, illumination, atmosphere, sun-sensor geometry, and ecosystem composition. The third is high-dimensional spectral covariate shift induced by heterogeneous preprocessing pipelines, spectral ranges, and noise characteristics. GreenHyperSpectra addresses these conditions by assembling a large unlabeled corpus for representation learning and by testing models under both in-distribution and out-of-distribution regimes.

This orientation places GreenHyperSpectra within a broader shift in hyperspectral machine learning from narrowly supervised predictors toward pretraining-based representation learning. A plausible implication is that the dataset is intended not merely as a benchmark, but as an infrastructure layer for semi-supervised and self-supervised spectral learning in ecology.

2. Dataset composition and harmonization

GreenHyperSpectra contains over 140,000 canopy surface reflectance spectra collected from proximal, airborne, and spaceborne platforms between 1992 and 2024. The source domains span forests, croplands, grasslands and savanna, shrublands, chaparral and steppe, tundra, coastal and wetland environments, urban vegetation, Mediterranean systems, and aquatic vegetation, with geographic coverage across North America, Europe, Africa, and beyond (Cherif et al., 9 Jul 2025).

The platform composition is summarized below.

Platform	Examples	Resolution and count
Proximal	ASD FieldSpec, SVC HR-1024i	\<1 m; 1–4 nm; 5,620 spectra
Airborne	AVIRIS-Classic, AVIRIS-NG, NEON AOP, Specim AisaFenix, HySpex	1–20 m; 3–7 nm; 96,699 spectra
Spaceborne	Hyperion, PRISMA, EMIT, EnMAP	30–60 m; 6–12 nm; 36,059 spectra

To harmonize these sources, all spectra are resampled to a uniform 1 nm grid across 400–2500 nm, producing 2101 bands per spectrum. Strong atmospheric absorption windows are removed at 1351–1430 nm, 1801–2050 nm, and 2451–2500 nm, and the remaining spectra are smoothed with a Savitzky–Golay filter using a 65 nm window. The resulting full-range input has 1721 bands, while a half-range subset restricted to 400–900 nm has 500 bands (Cherif et al., 9 Jul 2025).

A defining feature of this preprocessing is that the dataset uses real-world canopy reflectance rather than simulated spectra. The authors explicitly position this as a way to mitigate the simulation–reality domain gap that often limits downstream performance when purely synthetic radiative transfer outputs are used for pretraining. They also note that GreenHyperSpectra is explicitly vegetation-centric, multi-platform, and curated to reduce non-vegetated redundancy, in contrast to broader hyperspectral corpora such as HySpecNet-11k, SpectralEarth, and HyperSigma.

3. Trait benchmark and evaluation protocol

GreenHyperSpectra separates unlabeled pretraining from downstream supervised evaluation. The trait benchmark uses a distinct labeled dataset of 7,900 canopy spectra with co-located field or laboratory measurements for seven measured traits and one derived trait: Cab, Car, Anth, Cw, Cp, Cm, LAI, and $\mathrm{cbc} = C_m - C_p$ (Cherif et al., 9 Jul 2025).

The labeled data are drawn from 50 campaigns spanning forests, croplands, tundra, and other ecosystems. The reported trait statistics include, for example, Cab mean $39.12\ \mu g/cm^2$ with $n=2593$, Cm mean $0.0101\ g/cm^2$ with $n=4062$, LAI mean $3.49\ m^2/m^2$ with $n=1656$, Cp mean $0.0009\ g/cm^2$ with $n=3031$, and cbc mean $0.0104\ g/cm^2$ with $39.12\ \mu g/cm^2$0. A Box–Cox transformation is applied to targets to stabilize multi-output regression and better capture inter-trait correlations.

Two evaluation regimes are defined. In-distribution evaluation uses a fixed 80/20 hold-out split of the labeled dataset. Out-of-distribution evaluation uses a leave-five-datasets-out protocol across the 50 labeled campaigns: five datasets are held out for testing, and the remaining datasets are split 80/20 for training and validation. To reduce imbalance and $39.12\ \mu g/cm^2$1 instability, out-of-distribution reporting macro-averages predictions over five random subsamples per held-out dataset, each containing at most 30 samples.

The benchmark is reported in both full-range and half-range settings. The full-range setting uses the 1721-band 400–2450 nm representation. The half-range setting uses the 500-band 400–900 nm subset and also tests masked autoencoders pretrained on full-range data and then fine-tuned or applied to half-range inputs.

The evaluation metrics are the coefficient of determination, root mean square error, and normalized RMSE: $39.12\ \mu g/cm^2$2

$39.12\ \mu g/cm^2$3

$39.12\ \mu g/cm^2$4

This evaluation design makes domain shift a first-class benchmark criterion rather than a secondary robustness test. That choice aligns GreenHyperSpectra with later hyperspectral foundation-model work that likewise emphasizes sensor variability and distributional heterogeneity, although in different application domains such as PACE cloud retrieval (Tushar et al., 22 Apr 2026).

4. Learning methods and objective functions

The benchmark compares a supervised baseline with three representation-learning approaches: semi-supervised regression GAN, physics-informed autoencoding with a radiative transfer decoder, and masked autoencoding over 1D spectra (Cherif et al., 9 Jul 2025).

The supervised reference model is a 1D EfficientNet-B0 adapted for spectral inputs and equipped with a multi-output regression head. It serves as the state-of-the-art supervised baseline for sparse and heterogeneous trait data.

The semi-supervised regression GAN, or SR-GAN, is a 1D convolutional GAN in which the generator synthesizes reflectance spectra and the discriminator both regresses traits and learns a discriminative feature space $39.12\ \mu g/cm^2$5. The losses combine labeled supervision, unlabeled feature matching, fake-sample contrast, generator matching, and a gradient penalty: $39.12\ \mu g/cm^2$6

$39.12\ \mu g/cm^2$7

$39.12\ \mu g/cm^2$8

$39.12\ \mu g/cm^2$9

$n=2593$0

The discriminator loss is

$n=2593$1

The reported hyperparameters are batch size 128, 300 epochs, Adam with $n=2593$2, $n=2593$3, weight decay $n=2593$4, cosine embedding distance, Huber regression loss, and gradient penalty weight 10.0.

The physics-informed RTM-AE uses a small MLP encoder to predict traits, and a fixed PROSAIL-PRO radiative transfer decoder to reconstruct canopy reflectance. A learnable correction block compensates for the simulation–observation gap. Its reconstruction objective is

$n=2593$5

with supervised trait loss

$n=2593$6

The decoder includes PROSPECT-PRO leaf variables such as Cab, Car, Anth, Cw, Cp, CBC, structural parameters, and brown pigment, combined with 4SAIL canopy variables including LAI, LIDF, soil fraction, hotspot, and fixed geometry.

The best-performing method is the 1D masked autoencoder. It masks 75% of spectral tokens, reconstructs the full spectrum, and then supports either linear probing or full fine-tuning with a multi-output regression head. Its pretext loss combines cosine similarity and MSE: $n=2593$7 Representative hyperparameters are patch size 20, mask ratio 0.75, AdamW with learning rate $n=2593$8, weight decay $n=2593$9, batch size 128, and 500 pretraining epochs. In the reported ablations, patch size 20 and cosine-loss weight $0.0101\ g/cm^2$0 give the best validation performance.

Methodologically, GreenHyperSpectra therefore benchmarks three distinct forms of prior: adversarial semi-supervision, radiative-transfer structure, and masked spectral reconstruction. The empirical results favor the third, but the second remains an important interpretable baseline.

5. Empirical performance and generalization

The central empirical result is that masked reconstruction pretraining on GreenHyperSpectra yields the best average performance across full-range, half-range, and out-of-distribution evaluation, particularly when the pretrained model is fully fine-tuned. The table below reports average $0.0101\ g/cm^2$1 and nRMSE for the best masked-autoencoder configuration, denoted MAE_FR_FT, against the supervised baseline (Cherif et al., 9 Jul 2025).

Scenario	MAE_FR_FT (R^2 / nRMSE)	Supervised (R^2 / nRMSE)
Full-range ID	0.641 ± 0.020 / 12.777 ± 0.363	0.587 ± 0.010 / 13.697 ± 0.175
Half-range ID	0.627 ± 0.022 / 13.032 ± 0.368	0.163 ± 0.084 / 19.387 ± 0.982
OOD cross-dataset	0.311 ± 0.076 / 18.575 ± 2.584	0.243 ± 0.050 / 19.482 ± 2.278

In the full-range in-distribution setting, MAE_FR_FT reaches an average $0.0101\ g/cm^2$2 of $0.0101\ g/cm^2$3, compared with $0.0101\ g/cm^2$4 for the supervised baseline, and lowers average nRMSE from $0.0101\ g/cm^2$5 to $0.0101\ g/cm^2$6. Trait-wise $0.0101\ g/cm^2$7 for MAE_FR_FT are reported as 0.716 for Cm, 0.634 for Cw, 0.615 for LAI, 0.676 for Cp, 0.727 for cbc, 0.649 for Car, 0.598 for Anth, and 0.515 for Cab. The corresponding trait-wise nRMSE values are 9.856, 12.861, 16.285, 9.916, 9.833, 11.975, 14.117, and 17.374.

The half-range result is more striking. When only 400–900 nm information is retained, MAE_FR_FT still attains average $0.0101\ g/cm^2$8, whereas the supervised half-range baseline falls to $0.0101\ g/cm^2$9. Trait-wise half-range $n=4062$0 for MAE_FR_FT are 0.696 for Cm, 0.671 for Cp, 0.703 for cbc, 0.677 for Car, 0.567 for Cab, 0.592 for Cw, 0.556 for LAI, and 0.554 for Anth. The authors explicitly note that all semi- and self-supervised methods substantially outperform the supervised baseline in half-range settings.

In out-of-distribution evaluation, MAE_FR_FT again leads, with average $n=4062$1 against $n=4062$2 for the supervised model. Trait-wise, it improves Cm from 0.446 to 0.575, LAI from 0.074 to 0.229, Cw from 0.193 to 0.280, Cp from 0.183 to 0.275, cbc from 0.449 to 0.582, and Anth from 0.055 to 0.112, while Cab declines from 0.362 to 0.271 and Car declines slightly from 0.181 to 0.165. The authors also report that linear probing underperforms in OOD settings, indicating that fine-tuning is critical for adapting pretrained features to shifted spectral distributions.

Among the non-MAE methods, RTM-AE is the strongest runner-up in full-range settings and is second-best overall in half-range settings, while SR-GAN remains competitive on some traits but is generally weaker than MAE and RTM-AE. The reported label-efficiency study further shows that performance gains are strongest when only 20–40% of labeled data are available, and that increasing unlabeled pretraining size beyond a stratified subset yields diminishing returns. This suggests that coverage of sensor and ecosystem diversity matters more than raw sample count once a sufficiently broad unlabeled pool is available.

6. Significance, availability, and limitations

GreenHyperSpectra is significant because it recasts plant trait prediction as a cross-domain representation-learning problem rather than a narrowly supervised regression task. Its empirical results indicate that masked spectral pretraining can learn spectral shape and amplitude features that transfer across sensors, across ecosystems, and even across reduced spectral coverage. The authors therefore present it as a methodological framework as well as a dataset, with code and data released through the HyspectraSSL repository (Cherif et al., 9 Jul 2025).

The resource is also notable for its emphasis on real-world spectra. In adjacent hyperspectral research, large pretraining corpora and masked modeling have been used for reconstruction from multispectral inputs in greenhouse-gas monitoring (Avilés et al., 23 Apr 2025), but GreenHyperSpectra is explicitly organized around vegetation trait prediction and cross-ecosystem transfer. Its multi-platform design similarly distinguishes it from single-sensor corpora and from synthetic radiative-transfer datasets.

The limitations are also explicit. Although unlabeled coverage is broad, labeled trait data remain geographically and ecologically biased, and some traits such as Anth have fewer samples, increasing variance in out-of-distribution settings. Residual cross-sensor differences persist despite resampling and smoothing, including BRDF effects, atmospheric residuals, and calibration differences. The benchmark focuses on full-range VNIR+SWIR and VNIR-only settings; extension to other ranges or fused modalities such as LiDAR, SAR, or multi-angular measurements remains open. The authors also note run-to-run variance, indicating that stable representation learning across heterogeneous ecological distributions remains difficult.

Future directions proposed in the paper include multi-domain pretraining with spatio-temporal context, physics-informed self-supervision, active learning over underrepresented trait ranges, uncertainty quantification for trait maps, and scaling to regional and global canopy-level mapping. In that sense, GreenHyperSpectra functions both as a benchmark for current methods and as a substrate for future hyperspectral foundation models in ecology.