Distance-Based Uncertainty Quantification for Deep Learning in Remote Sensing
📌 Overview
Developed novel distance-based uncertainty quantification method (Dis_UN) for deep learning models applied to satellite imagery analysis. Addressed critical challenge of reliable uncertainty estimation when models encounter out-of-domain data (unseen regions, species, biomes, or scene components like clouds and water). Achieved 36% higher uncertainty contrast than traditional variance-based approaches while being 2.6-7.7× faster at inference.
Type: PhD Research Project
Duration: May 2023 – February 2026
Institution: ScaDS.AI Leipzig, RSC4Earth
Status: Accepted for publication in Biogeosciences (2026)
Technical Highlights:
- Designed distance-based uncertainty estimation framework that quantifies prediction reliability by measuring dissimilarity in predictor space (spectral inputs) and embedding space (learned features)
- Implemented FAISS-based nearest neighbor search for efficient dissimilarity computation in high-dimensional feature spaces
- Developed 95-quantile regression model using residuals as proxy for worst-case prediction errors
- Applied method to predict 6 plant traits (leaf mass per area, chlorophylls, carotenoids, nitrogen, water thickness, leaf area index) from hyperspectral imagery
- Evaluated performance on challenging out-of-domain scenarios: urban surfaces, bare ground, water bodies, clouds, and mixed pixels at 30m resolution
Technologies: TensorFlow, Keras, FAISS, Python, Scikit-learn, Quantile Regression, Distance Metrics, KS Statistics
Key Results:
- Performance: 36% higher uncertainty contrast (KS distance: 0.648 vs. 0.475) for out-of-domain data
- Efficiency: 2.6-7.7× faster inference requiring only single forward pass
- Robustness: No normality assumptions required - handles asymmetric errors
- Interpretability: Uncertainty directly linked to training data dissimilarity
Impact:
- Publication: Accepted in Biogeosciences (2026)
- Conference Presentation: GFÖ 2023 (German Ecological Society)
Links: