Intelligent Sampling for Vegetation Nitrogen Mapping Based on Hybrid Machine Learning Algorithms.

Area of Science:

• Remote sensing
• Agricultural science
• Machine learning

Background:

• Satellite imaging spectroscopy missions provide spatiotemporal data for vegetation property mapping.
• Precise and rapid derivation of crop-specific information, like nitrogen content, is crucial for agricultural monitoring.
• Efficient machine learning models require intelligently sampled training databases for fast processing.

Purpose of the Study:

• To implement active learning (AL) heuristics for optimizing training databases for vegetation nitrogen estimation.
• To minimize training data size while maintaining model performance for mapping aboveground nitrogen content.
• To utilize variational heteroscedastic Gaussian processes regression (VHGPR) for accurate nitrogen estimation and uncertainty quantification.

Main Methods:

• Active learning (AL) heuristics with kernel ridge regression (KRR) were used to optimize a training database.
• Uncertainty and diversity criteria were applied to a lookup table (LUT) of simulated hyperspectral reflectance and nitrogen content.
• Euclidian distance-based diversity (EBD) was identified as the best-performing AL criterion, reducing the LUT by 81%.

Main Results:

• The optimized, reduced LUT enabled efficient training of VHGPR for aboveground nitrogen content estimation.
• Validation against in situ data yielded excellent results: RMSE of 1.84 g/m² and R² of 0.92.
• Mapping nitrogen content in an agricultural region provided reliable estimates with meaningful uncertainty assessments.

Conclusions:

• Active learning significantly reduces the data requirements for training machine learning models in remote sensing.
• VHGPR coupled with AL offers a competitive and robust approach for estimating vegetation nitrogen content from satellite data.
• This hybrid workflow shows promise for operational nitrogen monitoring using future satellite imaging spectroscopy missions.