Prototyping Crop Traits Retrieval Models for CHIME: Dimensionality Reduction Strategies Applied to PRISMA Data

Area of Science:

• Remote Sensing
• Spectroscopy
• Machine Learning

Background:

• Next-generation imaging spectrometers will generate vast amounts of hyperspectral data.
• Optimized models are crucial for routine vegetation trait retrieval from this data.
• Hybrid models combining radiative transfer and machine learning are promising but face spectral collinearity challenges.

Purpose of the Study:

• To evaluate spectral dimensionality reduction methods (PCA and BR) within a hybrid model workflow for retrieving six key vegetation traits.
• To compare the performance of Gaussian process regression (GPR) models incorporating PCA (GPR-20PCA) versus BR (GPR-20BR).
• To assess the mapping capabilities of these models using real-world hyperspectral data.

Main Methods:

• Simulated training datasets using the SCOPE model and optimized with active learning.
• Gaussian process regression (GPR) algorithms trained for specific vegetation traits (SLA, LAI, CWC, CCC, FAPAR, FVC).
• Two dimensionality reduction strategies: Principal Component Analysis (PCA) and Band Ranking (BR), each selecting 20 features.

Main Results:

• GPR-20PCA models demonstrated slightly better performance than GPR-20BR models across all six vegetation traits.
• Normalized Root Mean Squared Error (NRMSE) ranged from 13.9% (CWC) to 22.3% (CCC) for GPR-20PCA.
• Mapping consistency was highest for Fractional Vegetation Cover (FVC) and Leaf Area Index (LAI) (R² = 0.91, R² = 0.86).

Conclusions:

• The GPR-20PCA approach is recommended as the most efficient strategy for retrieving multiple crop traits from hyperspectral data streams.
• This workflow supports preparations for next-generation operational missions like the Copernicus Hyperspectral Imaging Mission for the Environment (CHIME).
• The study highlights the effectiveness of PCA in handling spectral collinearity for improved vegetation trait retrieval.