A Refined Rotation Forest-Based Ensemble Classifier for Lithological Mapping with ZY1-02D Hyperspectral Remote...

Area of Science:

• Geoscience
• Remote Sensing
• Machine Learning

Background:

• Hyperspectral remote sensing excels at lithological classification in exposed bedrock areas.
• Ensemble learning methods like Rotation Forest (ROF) show promise but face high training costs with high-dimensional hyperspectral data.
• Optimizing these methods is crucial for efficient and accurate geological mapping.

Purpose of the Study:

• To propose an optimized ROF-LightGBM ensemble algorithm for hyperspectral lithological classification.
• To integrate Minimum Noise Fraction (MNF) for improved rotation matrix construction within the ensemble.
• To evaluate the model's accuracy and computational efficiency against other machine learning techniques.

Main Methods:

• Developed an optimized ROF-LightGBM algorithm incorporating MNF for rotation matrix construction.
• Conducted experimental validation using ZY1-02D hyperspectral data for lithological mapping.
• Performed parameter optimization (L × T, bootstrap) and comparative analysis with Random Forest, LightGBM, and traditional ROF.

Main Results:

• The optimized ROF-LightGBM (PCA, L × T = 4 × 25) achieved 74.28% accuracy, surpassing RF and LightGBM.
• This model demonstrated significant efficiency, with a training time of 4.86 s, ~37 times faster than traditional ROF.
• The ROF-LightGBM (MNF) model further boosted accuracy to 82.17% with a kappa coefficient of 0.81.

Conclusions:

• The proposed optimized ROF-LightGBM ensemble algorithm significantly enhances lithological classification accuracy and efficiency.
• Integrating MNF for rotation matrix construction provides superior performance compared to PCA-based methods.
• This approach offers a powerful and efficient solution for hyperspectral geological mapping.