Supplementary material to "Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral data"

Eya Cherif, Teja Kattenborn, Luke A. Brown, Michael Ewald, Katja Berger, Phuong D. Dao, Tobias Hank, Étienne Laliberté, Bing Lu, Hannes Feilhauer2025 · DOI

Introduction

Deep learning (DL) models have become indispensable tools for retrieving plant functional traits from hyperspectral remote sensing data. However, these models often lack a reliable measure of uncertainty, leading to overconfident predictions when encountering out-of-domain (OOD) data. For instance, a model trained on temperate forests might produce nonsensical values when applied to urban areas or water bodies. This document explores robust methods for uncertainty estimation to ensure that AI-driven plant trait mapping is both accurate and trustworthy.

Methods for Uncertainty Estimation

The study evaluates three primary techniques for quantifying uncertainty:

  1. Distance-based (Dis_UN): This novel approach calculates the normalized median cosine distance of a sample to its 50 nearest neighbors. This calculation is performed in both the raw feature space and the model's final "embedding space"—the internal representation just before the prediction layer.
  2. Ensemble Deep Learning (Ens_UN): Multiple models are trained independently, and the variance among their outputs serves as a proxy for uncertainty.
  3. Monte Carlo Dropout (MCDrop_UN): Dropout layers remain active during inference, generating a distribution of outputs for a single input to estimate predictive variance.

The researchers utilized t-distributed Stochastic Neighbor Embedding (t-SNE) to visualize how data points from different land covers cluster within these high-dimensional spaces.

Datasets and Traits

The research leverages a massive compilation of 50 distinct datasets (DS1–DS50), covering a wide array of vegetation types including crops, forests, grasslands, shrublands, and tundra. The data originates from various proximal, airborne, and spaceborne sensors.

Descriptive statistics for the six functional plant traits across the compiled datasets.

As shown in the table above, the models target six essential functional traits: Leaf Mass per Area (LMA), Nitrogen (N), Leaf Area Index (LAI), Chlorophyll (Chl), Equivalent Water Thickness (EWT), and Carotenoids (Car). The broad range of values across these traits highlights the complexity of the retrieval task.

Key Insights: The Power of Embedding Space

A critical finding of this work is that the model's internal embedding space provides a superior environment for OOD detection compared to the raw input space.

t-SNE visualization showing the separation of land cover types in the model's embedding space.

In the embedding space, training data and OOD data (such as urban surfaces or water) separate more distinctly. This separation allows the distance-based method (Dis_UN) to naturally assign high uncertainty to non-vegetated pixels, effectively creating an "automatic mask." This self-masking capability prevents the model from making unreliable predictions on surfaces it was never intended to analyze.

Comparative Results

The distance-based method significantly outperformed standard ensemble and dropout techniques in terms of calibration and separability:

  • Calibration Accuracy: Dis_UN achieved a much lower Expected Normalized Calibration Error (ENCE). For Nitrogen retrieval, Dis_UN recorded an ENCE of 0.441, compared to 1.821 for Ens_UN and 14.995 for MCDrop_UN.
  • Statistical Separability: Using Jeffries Matusita (JM) distances, Dis_UN demonstrated a superior ability to distinguish vegetated from non-vegetated pixels based on uncertainty values alone (JM distances of ~0.5-0.6 vs. <0.16 for other methods).

Conclusion and Future Directions

While the distance-based approach offers a robust solution for uncertainty quantification, it comes with a higher computational cost due to the nearest-neighbor calculations required for every pixel. Furthermore, the effectiveness of this metric is inherently tied to the diversity and representativeness of the training data. Future research will focus on optimizing these calculations for large-scale satellite imagery and integrating these uncertainty "error bars" into global ecological and climate models to improve their predictive reliability.

Summary updated

BibTeX
@misc{cherif2025supplementary,
  author = {Eya Cherif and Teja Kattenborn and Luke A. Brown and Michael Ewald and Katja Berger and Phuong D. Dao and Tobias Hank and Étienne Laliberté and Bing Lu and Hannes Feilhauer},
  title = {{Supplementary material to "Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral data"}},
  year = {2025},
  doi = {10.5194/egusphere-2025-1284-supplement},
}
Figure