Random Forest (RF) Wrappers for Waveband Selection and Classification of Hyperspectral Data

Area of Science:

• Plant Pathology
• Remote Sensing
• Forestry

Background:

• Fusarium circinatum infection causes asymptomatic stress in Pinus radiata and Pinus patula seedlings.
• Early detection of plant stress is crucial for effective disease management in forest nurseries.

Purpose of the Study:

• To model and detect asymptomatic stress in Pinus seedlings infected with Fusarium circinatum using hyperspectral data.
• To evaluate the effectiveness of different feature selection algorithms in improving classification accuracy.

Main Methods:

• Hyperspectral data were collected using a field spectroradiometer.
• The random forest algorithm was employed for data analysis.
• Waveband selection was performed using Boruta, Recursive Feature Elimination (RFE), and AUC-RF algorithms.

Main Results:

• The Boruta feature selection algorithm yielded the highest classification accuracy for Pinus radiata (17.00% training error, 17.00% test error, 0.91 AUC).
• Classification accuracy was lower for Pinus patula seedlings (24.00% training error, 38.00% test error, 0.65 AUC).
• A hybrid waveband selection method improved accuracy for Pinus patula but not for Pinus radiata.

Conclusions:

• Feature selection methods, particularly Boruta, significantly enhance the accuracy of hyperspectral analysis for detecting plant stress.
• Hyperspectral imaging offers a promising framework for early stress detection in forest nursery environments.
• Further research may be needed to optimize detection for different pine species.