🌿 PlantTraitNet

An Uncertainty-Aware Multimodal Framework for Global-Scale Plant Trait Inference from Citizen Science Data

License: CC-BY-4.0
Paper: https://arxiv.org/abs/2511.06943, accepted in AAAI

Introduction

PlantTraitNet is a curated multimodal dataset for training and evaluating global plant trait inference models using citizen science images paired with trait distributions from TRY.
Each sample contains an RGB observation, species information, geolocation, and quantified trait ranges supporting uncertainty-aware modeling.

Dataset Splits

Split	Description	Purpose
train	Refined dataset used for training	Model fitting
val	Validation subset of refined data	Tuning + early stopping
reference	Reference set used to identify turning points in uncertainty
benchmark	Benchmark set (benchmark + val) for global evaluation vs sPlotOpen	Cross-dataset benchmarking
traitmap	Global .tif layers (TRY community-weighted means + sPlotOpen)	Global mapping + inference
train_filtered	Raw dataset used for training (train + train_filtered)

Column Schema

Column	Type	Meaning
Image	path	Filepath to RGB image
Latitude	float	WGS84 latitude
Longitude	float	WGS84 longitude
Species	string	Identified plant species
<trait>_lowerbound	float	Q1 (25th percentile) trait value from TRY
<trait>_upperbound	float	Q3 (75th percentile) trait value
<trait>_median	float	Median trait estimate
<trait>_sd	float	Standard deviation of trait distribution

<trait> is replaced by each trait included in the dataset.

Citation

Coming soon: Citation of published verison

@misc{sharma2025planttraitnetuncertaintyawaremultimodalframework,
      title={PlantTraitNet: An Uncertainty-Aware Multimodal Framework for Global-Scale Plant Trait Inference from Citizen Science Data}, 
      author={Ayushi Sharma and Johanna Trost and Daniel Lusk and Johannes Dollinger and Julian Schrader and Christian Rossi and Javier Lopatin and Etienne Laliberté and Simon Haberstroh and Jana Eichel and Daniel Mederer and Jose Miguel Cerda-Paredes and Shyam S. Phartyal and Lisa-Maricia Schwarz and Anja Linstädter and Maria Conceição Caldeira and Teja Kattenborn},
      year={2025},
      eprint={2511.06943},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.06943}, 
}