A Survey of Active Learning for Quantifying Vegetation Traits from Terrestrial Earth Observation Data

Area of Science:

• Earth Observation
• Machine Learning
• Vegetation Science

Background:

• Satellite data streams are rapidly increasing, offering opportunities for global vegetation monitoring.
• Hybrid retrieval approaches combine radiative transfer models (RTMs) with machine learning (ML) for vegetation variable estimation.
• Intelligent sampling using active learning (AL) can optimize ML model training.

Purpose of the Study:

• To review AL heuristics for Earth observation regression problems.
• To demonstrate the benefits of AL for estimating vegetation biophysical variables.
• To encourage the adoption of AL in operational vegetation monitoring workflows.

Main Methods:

• Systematic literature survey of AL heuristics in Earth observation.
• Case study using PROSAIL-PRO RTM simulations and experimental data.
• Gaussian Process Regression (GPR) for AL-based data subset optimization.

Main Results:

• AL-optimized datasets improved retrieval accuracy compared to random sampling.
• Euclidean distance-based (EBD) diversity was the most efficient AL technique.
• AL improved estimation of leaf carotenoid and water content, outperforming full datasets.

Conclusions:

• AL-based sub-sampling effectively selects informative samples, optimizing regression accuracy and reducing computational load.
• AL facilitates faster processing and smaller model sizes for kernel-based ML algorithms like GPR.
• AL is recommended for hybrid retrieval workflows in operational vegetation monitoring with satellite imaging spectroscopy.