Advanced Global Prototypical Segmentation Framework for Few-Shot Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning methods show promise for Hyperspectral Image (HSI) classification.
• Existing HSI classification methods struggle with capturing global information due to patch-based inputs.
• Limited labeled samples hinder the performance of current HSI classification techniques.

Purpose of the Study:

• To propose an Advanced Global Prototypical Segmentation (AGPS) framework to address challenges in HSI classification.
• To enhance the capture of global information and improve the utilization of limited labeled samples in HSI classification.

Main Methods:

• A patch-free feature extractor segmentation network (SegNet) based on a fully convolutional network (FCN) processes entire HSIs.
• A Fusion of Lateral Connection (FLC) structure integrates low-level and high-level features.
• An Atrous Spatial Pyramid Pooling-Position Attention (ASPP-PA) module captures multi-scale spatial information.
• An advanced global prototypical representation learning strategy with supervised contrastive learning (CL) optimizes the network using three constraints.

Main Results:

• The proposed AGPS framework effectively captures global information from HSIs.
• The method demonstrates superior performance in utilizing limited labeled samples.
• Experimental results on three public datasets show that the AGPS framework outperforms state-of-the-art methods.

Conclusions:

• The AGPS framework offers a significant advancement in HSI classification.
• The integration of global information capture and effective sample utilization leads to improved classification accuracy.
• The proposed approach provides a robust solution for HSI classification tasks with limited labeled data.