Multi-Temporal Hyperspectral Classification of Grassland Using Transformer Network.

Area of Science:

• Remote Sensing
• Ecology
• Computer Science

Background:

• Grassland monitoring traditionally relies on field surveys, but remote-sensing methods offer greater scalability.
• Existing remote-sensing techniques struggle to achieve desired accuracy in grassland classification.
• Multi-temporal hyperspectral data holds significant potential for distinguishing grassland species and growth stages.

Purpose of the Study:

• To explore the application of transformer networks for analyzing multi-temporal hyperspectral data in grassland classification.
• To introduce and evaluate a novel Multi-Temporal Hyperspectral Classification of Grassland using Transformer network (MHCgT).
• To enhance the accuracy of high-powered grassland detection and species identification.

Main Methods:

• Collected 16,800 multi-temporal hyperspectral data points from grassland samples across different growth stages (400-1000 nm).
• Developed the MHCgT network with a hierarchical architecture and multi-head self-attention mechanism for feature extraction.
• Conducted ablation studies and comparative experiments against state-of-the-art methods (CNN, LSTM-RNN, SVM, RF, DT).

Main Results:

• The proposed MHCgT framework achieved a high accuracy of 98.51% in multi-temporal hyperspectral grassland identification.
• MHCgT outperformed existing methods by 6.42-26.23%.
• Average species classification accuracy exceeded 95%, with mature stages (August) being easier to identify than growth stages (June).

Conclusions:

• The MHCgT framework demonstrates significant potential for precise multi-temporal hyperspectral species identification.
• This approach offers valuable applications for sustainable grassland management and species diversity assessment.
• Transformer networks are a powerful tool for extracting features from complex hyperspectral time-series data.