Semantic-Guided Transformer Network for Crop Classification in Hyperspectral Images.

Area of Science:

• Agricultural remote sensing
• Computer vision
• Machine learning

Background:

• Hyperspectral remote sensing images offer rich spectral information for crop monitoring.
• Existing methods struggle with complex backgrounds and varying crop scales, reducing classification accuracy.
• Spectral similarity and scale variations hinder effective feature extraction in hyperspectral crop analysis.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and robust hyperspectral crop classification.
• To address limitations in current methods regarding background interference and scale variations.
• To enhance the extraction of semantic and spatial-spectral features for improved crop recognition.

Main Methods:

• A novel semantic-guided transformer network (SGTN) was proposed.
• A multi-scale spatial-spectral information extraction (MSIE) module was designed to handle scale variations.
• A semantic-guided attention (SGA) module was developed to reduce background interference and focus on crop semantics.
• A two-stage feature extraction structure was employed for optimized feature learning.

Main Results:

• The SGTN achieved high overall accuracies on benchmark datasets: 98.24% (Indian Pines), 98.34% (Pavia University), and 97.89% (Salinas).
• The model demonstrated superior classification accuracy and generalization performance compared to existing methods.
• The MSIE and SGA modules effectively improved feature extraction and reduced background noise.

Conclusions:

• The SGTN effectively overcomes limitations of traditional deep learning methods for hyperspectral crop classification.
• The proposed model offers enhanced accuracy and robustness in challenging remote sensing scenarios.
• The SGTN shows potential for future applications in precision agriculture, including disease detection and yield prediction.