Feature Consistency-Based Prototype Network for Open-Set Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral Image (HSI) classification has advanced significantly.
• Existing methods often rely on a closed-set assumption, limiting their effectiveness in real-world scenarios with unknown classes.
• Open-world scenes require models capable of handling both known and previously unseen classes.

Purpose of the Study:

• To develop an effective method for open-set Hyperspectral Image (HSI) classification.
• To address the limitations of closed-set assumptions in HSI classification.
• To accurately identify both known and unknown classes in HSI data.

Main Methods:

• A Feature Consistency-based Prototype Network (FCPN) was proposed.
• A three-layer convolutional network with a contrastive clustering module was used for discriminative feature extraction.
• A scalable prototype set was constructed from extracted features.
• A prototype-guided open-set module (POSM) was developed to classify known and unknown samples.

Main Results:

• The proposed FCPN demonstrated superior performance in open-set HSI classification.
• The method effectively identified unknown samples, overcoming closed-set limitations.
• Experimental results showed remarkable classification accuracy compared to state-of-the-art techniques.

Conclusions:

• The FCPN offers a robust solution for open-set HSI classification challenges.
• The integration of feature consistency and prototype learning enhances classification accuracy.
• This approach paves the way for more reliable HSI analysis in diverse, real-world applications.