Spectral-Spatial Attention Transformer with Dense Connection for Hyperspectral Image Classification

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise in hyperspectral image (HSI) classification.
• CNNs face limitations in HSI analysis due to redundant information and restricted receptive fields, hindering effective sequence feature extraction.
• HSIs possess inherent sequential characteristics that traditional CNN models struggle to fully exploit.

Purpose of the Study:

• To develop an advanced deep learning model for more efficient and accurate hyperspectral image classification.
• To address the limitations of CNNs in capturing both local and global spectral-spatial features in HSIs.
• To enhance the mining of sequential features within HSI data.

Main Methods:

• Proposed the SSA-Transformer, a hybrid model combining a modified CNN-based spectral-spatial attention mechanism with a self-attention-based Transformer.
• Integrated dense connections within the Transformer component to improve feature propagation and model depth.
• Employed a spectral-spatial attention mechanism to refine feature extraction before Transformer processing.

Main Results:

• The SSA-Transformer model demonstrated competitive classification accuracy across three benchmark HSI datasets: University of Pavia (PU), Salinas (SA), and Kennedy Space Center (KSC).
• The model effectively combined local features (via CNN) and global features (via Transformer) for improved HSI classification performance.
• Achieved superior results compared to existing CNN-based classification methods, highlighting the efficacy of the proposed architecture.

Conclusions:

• The SSA-Transformer offers an effective approach for hyperspectral image classification by leveraging both spectral and spatial information.
• The hybrid CNN-Transformer architecture successfully addresses the limitations of purely CNN-based models in handling HSI data.
• This model provides a robust framework for extracting complex spectral-spatial features, leading to enhanced classification accuracy.