Semisupervised Cross-Scale Graph Prototypical Network for Hyperspectral Image Classification | JoVE Visualize

Area of Science:

Remote Sensing
Computer Vision
Machine Learning

Background:

Hyperspectral image classification (HSIC) faces challenges due to limited labeled data, leading to model overfitting and performance issues in supervised methods.
Graph Convolutional Networks (GCNs) offer a promising semisupervised approach by propagating information between nodes transductively.
Existing GCN methods may not fully capture the multiscale nature of land cover appearances in remotely sensed scenes.

Purpose of the Study:

To propose a Cross-Scale Graph Prototypical Network (X-GPN) for high-quality semisupervised HSIC.
To address the limitations of data scarcity and improve classification performance in HSIC.
To leverage multiscale spectral-spatial features for more robust land cover classification.

Main Methods:

Constructing adjacency matrices using multiscale neighborhoods to capture diverse spatial contexts.
Implementing a multibranch framework with 1-D convolutions to extract both spectral and spatial features across different scales.
Developing a Self-Branch Attentional Addition (SBAA) module to adaptively weigh features from different branches.
Introducing an innovative prototypical layer with Distance-based Cross-Entropy (DCE) loss and Temporal Entropy-based Regularizer (TER) for enhanced feature discrimination and representativeness.

Main Results:

The proposed X-GPN significantly outperforms classic and state-of-the-art methods in HSIC.
Experimental results validate the effectiveness of the multiscale approach and the SBAA module in improving classification accuracy.
The novel prototypical layer enhances the discriminative power and representativeness of learned features.

Conclusions:

The X-GPN provides a superior semisupervised approach for hyperspectral image classification.
The method effectively utilizes multiscale information and attention mechanisms for enhanced feature learning.
The innovative prototypical layer contributes to improved discrimination and representativeness, advancing the field of HSIC.