Research map / Label-efficient and robust learning / Self-supervised and contrastive
GreenHyperSpectra: A multi-source hyperspectral dataset for global vegetation trait prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GreenHyperSpectra, a multi-source hyperspectral pretraining dataset (~140k canopy spectra), and benchmarks semi- and self-supervised models (SR-GAN, RTM-AE, MAE) against a supervised EfficientNet-based CNN for multi-output plant trait regression.
Problem
Plant trait prediction from hyperspectral data suffers from label scarcity and domain shifts across sensors and ecosystems, and existing benchmarks are single-sensor and lack semi/self-supervised multi-output evaluation.
Approach
Pretrain on a large unlabeled, cross-sensor, cross-ecosystem spectral dataset and adapt semi- and self-supervised methods (GAN, RTM-physics autoencoder, masked autoencoder) for simultaneous multi-trait regression, evaluated in-distribution, half-range, and out-of-distribution settings.
Key results
- MAE fine-tuned (MAE_FR_FT) reached average R^2 0.641 vs 0.587 for the supervised baseline on full-range spectra (about 9% relative gain).
- In OOD cross-dataset evaluation, MAE_FR_FT achieved average R^2 0.311 vs 0.243 for the supervised baseline.
- On half-range (VNIR) spectra, semi/self-supervised methods gained roughly 100-200% in R^2 over the supervised baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pretrain on GreenHyperSpectra and fine-tune MAE instead of training the supervised baseline | Full-range trait prediction · Full-range trait prediction | average R^2: 0.587 (± 0.010) -> 0.641 (± 0.020) | Helped |
| Use full fine-tuning rather than linear probing (ablation) | MAE pretrained on full-range spectra · Full-range trait prediction | average R^2: 0.466 (± 0.008) -> 0.641 (± 0.020) | Helped |
| Pretrain on full-range spectra and fine-tune for half-range inputs rather than train the supervised baseline on half-range data | Half-range trait prediction · Half-range trait prediction | average R^2: 0.163 { (± 0.084)} -> 0.566 { (± 0.015)} | Helped |
| Pretrain on full-range rather than half-range spectra before fine-tuning with half-range inputs (ablation) | MAE · Half-range trait prediction | average R^2: 0.465 { (± 0.040)} -> 0.566 { (± 0.015)} | Helped |
| Pretrain on GreenHyperSpectra and fine-tune MAE rather than train the supervised baseline | OOD trait prediction · OOD cross-dataset evaluation | average R^2: 0.24 vs. 0.31 | Helped |
| Increase the quantity of unlabeled pretraining data (ablation) | Semi- and self-supervised models · Trait prediction | R^2 | No clear effect |
Surprising findings
- Varying the size of the unlabeled pretraining set did not substantially change performance.
- MAE linear probing alone underperformed the supervised baseline (average R^2 0.466 vs 0.587 full-range).
Limitations
- Run-to-run variance remains high and OOD evaluation uses a single training run due to computational constraints.
- Labeled data are geographically and ecologically limited, with biases toward certain vegetation types.
- Only default VNIR+SWIR and VNIR sensor configurations were explored.
Experimental setup
Datasets: GreenHyperSpectra aggregated annotated plant trait dataset (7,900 spectra, 50 experiments)
Benchmarks: GreenHyperSpectra trait-annotated dataset (7,900 spectra, 7 traits)
Metrics: R^2 nRMSE
Baselines: Supervised CNN (EfficientNet-B0-based) SR-GAN RTM-AE MAE linear probing
Builds on: masked autoencoder SR-GAN PROSAIL-PRO EfficientNet-B0
Closest papers by meaning
- Multi-modal Learning for Geospatial Vegetation Forecasting (CVPR 2024 · similarity 0.85)
- GeoPlant: Spatial Plant Species Prediction Dataset (NeurIPS 2024 · similarity 0.84)
- FoMo: Multi-Modal, Multi-Scale and Multi-Task Remote Sensing Foundation Models for Forest Monitoring (AAAI 2025 · similarity 0.83)
- Hybrid Phenology Modeling for Predicting Temperature Effects on Tree Dormancy (AAAI 2025 · similarity 0.82)
- LESSViT: Robust Hyperspectral Representation Learning under Spectral Configuration Shift (NeurIPS 2026 · similarity 0.81)
- DivShift: Exploring Domain-Specific Distribution Shifts in Large-Scale, Volunteer-Collected Biodiversity Datasets (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.