Avatarr05/Dist_Uncertainty

Dist_Uncertainty – Hyperspectral Case Studies for Plant Trait Uncertainty Assessment

[image]

Dataset Description

This dataset contains hyperspectral remote sensing imagery and associated land cover labels used as out-of-domain (OOD) test cases for evaluating uncertainty estimation methods in deep learning-based plant trait retrievals. It accompanies the paper by Cherif et al. (2025, Biogeosciences) and supports the evaluation of a distance-based uncertainty method (Dis_UN) against traditional approaches (deep ensembles and Monte Carlo dropout).

The dataset includes two real-world hyperspectral scenes from different sensors and geographic locations:

• —EnMAP scene over south Leipzig, Germany
• —NEON AOP scene over the Liro site, Wisconsin, USA

Each scene comes with spectral band metadata and pixel-level land cover labels, enabling the identification of OOD components such as urban surfaces, bare ground, water bodies, and clouds — elements not represented in the training data.

[image]

The TrainingSpectra_1522.csv file serves as the reference distribution for the distance-based uncertainty method (Dis_UN): at inference time, the dissimilarity between any new unseen spectrum and this training set is computed in the predictor and embedding space to quantify how far the new data lies from the known training distribution.

Dataset Structure

Dist_Uncertainty/
├── TrainingSpectra_1522.csv              # 1522 labeled training spectra
└── CaseStudies/
    ├── EnMAP/
    │   ├── clip2_south.tif               # EnMAP hyperspectral image (south Leipzig): Contains modified EnMAP data © DLR [2024].
    │   ├── EnmapBands.csv                # EnMAP band wavelengths
    │   ├── southLeipzig2_mask_lc.csv     # Land cover metatada (all classes)
    │   └── southLeipzig2_mask_lc_veg.csv # Land cover metatada (vegetation only)
    └── NEON/
        ├── clip_Liro_3.tif               # NEON hyperspectral image (Liro, tile 3): Contains modified NEON data
        ├── NeonBands.csv                 # NEON band wavelengths
        └── Liro3_mask_lc.csv             # Land cover metatada

Note: The .tif GeoTIFF images are available as raw file downloads. The HF Dataset Viewer loads the CSV files only.

Citation

If you use this dataset, please cite the associated paper:

Cherif, E., Kattenborn, T., Brown, L. A., Ewald, M., Berger, K., Dao, P. D., Hank, T. B., Laliberté, E., Lu, B., and Feilhauer, H.: Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral data, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2025-1284, 2025.

@article{cherif2025uncertainty,
  author    = {Cherif, Eya and Kattenborn, Teja and Brown, Luke A. and Ewald, Michael
               and Berger, Katja and Dao, Phuong D. and Hank, Tobias B.
               and Laliberté, Etienne and Lu, Bing and Feilhauer, Hannes},
  title     = {Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals
               from Hyperspectral data},
  journal   = {EGUsphere},
  year      = {2025},
  note      = {preprint},
  doi       = {10.5194/egusphere-2025-1284}
}

Acknowledgements

The EnMAP hyperspectral data were provided by the German Aerospace Center (DLR) through the EnMAP Science Service System (https://planning.enmap.org/).

The NEON hyperspectral data were provided by the National Ecological Observatory Network (NEON, https://data.neonscience.org).