Multi-Trait Prediction from Hyperspectral Satellite Imagery
📌 Overview
Developed multi-task deep learning models to simultaneously predict 20 vegetation parameters from hyperspectral satellite data across diverse ecosystems. Built end-to-end ML pipeline from heterogeneous data integration through model deployment, achieving state-of-the-art performance while being 20× more computationally efficient than existing approaches.
Type: PhD Research Project
Duration: June 2021 – May 2023
Institution: ScaDS.AI Leipzig, RSC4Earth
Status: Published in Remote Sensing of Environment (2023)
Technical Highlights:
- Compiled and harmonized heterogeneous dataset from multiple sensors and ecosystem types (7,000+ labeled samples across crops, forests, tundra, grasslands, shrublands)
- Designed custom 1D-EfficientNet architecture optimized for hyperspectral spectral data (100-400 wavelength bands)
- Implemented multi-task learning framework to predict 20 traits simultaneously (leaf area index, chlorophyll, nitrogen, carbon, water content, etc.)
- Developed custom loss functions to handle sparse labels and missing trait values
- Achieved 20-40% improvement in prediction accuracy (nRMSE) compared to state-of-the-art PLSR methods
- Outperformed single-trait CNN approaches by 0.03-19.6% while requiring only 1 model instead of 20
Technologies: TensorFlow, Keras, Scikit-learn, Python (NumPy, Pandas), HPC Systems, HuggingFace, Multi-Task Learning
Recognition & Impact:
- Publication: Remote Sensing of Environment (Impact Factor ~13)
- Award: EARSeL Young Scientist Award Finalist (Top 5)
- Conference Presentations: LPS 2022 (ESA), EARSeL 2022, EGU 2023, GFÖ 2023
- Open Source: Code and models publicly available for research community
- Applications: Models ready for deployment on satellite missions (EnMAP, PRISMA) for large-scale vegetation monitoring
Key Results:
Multi-trait CNN models simultaneously predicted 20 vegetation properties from hyperspectral reflectance. The approach significantly outperformed both single-trait CNNs (nRMSE improvement: 0.027–19.61%) and state-of-the-art PLSR models (nRMSE improvement: 1.94–40.07%) across diverse vegetation types and sensor configurations. The multi-task approach proved both computationally more efficient (single model vs. 20 separate models) and more accurate by leveraging trait co-variation.
Links: