SSATNet: Spectral-spatial attention transformer for hyperspectral corn image classification.

Area of Science:

• Agricultural Science
• Computer Vision
• Data Science

Background:

• Hyperspectral imaging offers rich spectral and spatial data crucial for intelligent agriculture.
• Accurate classification of corn seed varieties is vital for agricultural applications.
• Existing methods struggle with large hyperspectral data volumes and complex features, limiting classification accuracy.

Purpose of the Study:

• To develop an advanced hyperspectral corn image classification method.
• To improve feature extraction and utilization for enhanced classification accuracy.
• To address the limitations of current methods in analyzing complex hyperspectral corn data.

Main Methods:

• Proposed a spectral-spatial attention transformer network (SSATNet).
• Utilized 3D and 2D convolutions for local feature extraction (spatial, spectral, textural).
• Incorporated spectral-spatial morphological structures and a transformer encoder with cross-attention for global feature refinement.

Main Results:

• SSATNet demonstrated superior performance on hyperspectral corn image classification.
• The model effectively extracts and refines spectral and spatial features.
• Achieved higher classification accuracy compared to state-of-the-art methods.

Conclusions:

• SSATNet is effective for hyperspectral corn image classification.
• The proposed network addresses challenges in feature extraction and utilization.
• This technology holds promise for advancing intelligent agriculture through accurate crop varietal identification.