🌍 Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral Data

📌 Description

With the advent of Hyperspectral satellites and Deep Learning, we have the chance to monitor biodiversity on a global scale as a contribution to global digital data streams.

However, there is still a need for methods that fully report the uncertainty of deep learning-derived products. In this regard, this study introduces Dis_UN, a novel distance-based method for estimating predictive uncertainty in deep learning models for plant trait retrievals.

Unlike traditional uncertainty estimation approaches, Dis_UN quantifies worst-case errors by:

• Using residuals as a proxy for uncertainty.
• Measuring dissimilarity in data manifolds.

🔹 Key Findings:

• Dis_UN effectively detects Out-Of-Domain (OOD) spectral variations, such as urban surfaces, clouds, and open water.
• Provides more reliable uncertainty estimates compared to conventional methods, which often underestimate uncertainty.
• With the increasing availability of hyperspectral data, robust uncertainty quantification is crucial for Earth Observation applications.
• This method contributes to uncertainty quantification in machine learning-derived maps, offering a practical tool for uncertainty assessment valuable for ecological and environmental analyses.

---

⚙️ Setup & Installation

🔹 Requirements

• Python 3.9.5
• TensorFlow 2.7.0
• Additional dependencies are listed in requirements.txt.

🔹 Installation Steps

1. Clone the repository
2. conda create -n python==3.9
3. conda activate
4. Install tensorflow (The trained model was tested for =2.7.0). May vary with your system.
5. cd into the cloned repo
6. pip install -r requirements.txt

📂 Dataset

Raw data available at:
👉 Data

• Spectra of the training data
• Use case scenes: EnMAP and NEON scenes

👉 Corresponding distances in : distances

• DisQuTraining_50neighFaiss_SpLastlayQuNorWithResiduals.csv
• DistQuInferenceEnMAP_50neighFaiss_SpLastlayQuNor.csv
• DistQuInferenceNEON_50neighFaiss_SpLastlayQuNor.csv

🚀 Usage

1️⃣ Distance Calculation

To generate the distances of a new scene to the training data, run:

python distInference.py \
  --r ./Data/TrainingSpectra_1522.csv \
  --modelpath ./DL_Model/ \
  --inferpath ./Data/CaseStudies/EnMAP/clip2_south.tif \
  --metapath ./Data/CaseStudies/EnMAP/EnmapBands.csv \
  --ne 50 \
  --lay -4 \
  --norVec True \
  --emb True \
  --sp True \
  --nor True \
  --gpu 0 \
  --sceneText clip2_southSpLastlayQuNor_Test

2️⃣ Apply the Model for Uncertainty Estimation

To infer uncertainty maps, run:

python distance_UN.py \
  --modelpath ./Un_models_95QuReg/ \
  --path_dist_ts ./distances/DistQuInferenceEnMAP_50neighFaiss_SpLastlayQuNor.csv \
  --sceneText clip2_southSpLastlayQuNor_Un