Robust Self-Ensembling Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning shows promise for hyperspectral image (HSI) classification.
• Training deep learning models for HSI classification typically requires extensive labeled data, which is costly and time-consuming to acquire.
• Developing methods that perform well with limited labeled data is crucial for practical HSI analysis.

Purpose of the Study:

• To propose a novel deep learning approach for HSI classification that addresses the challenge of limited labeled data.
• To introduce a robust self-ensembling network (RSEN) that leverages both labeled and unlabeled HSI data.
• To enhance the utilization of unlabeled HSI data for improved classification accuracy.

Main Methods:

• A robust self-ensembling network (RSEN) comprising a base network and an ensemble network was developed.
• A self-ensembling mechanism was implemented by constraining both supervised and unsupervised losses, enabling mutual learning between the networks.
• A novel consistency filter was introduced to bolster the robustness of the self-ensembling learning process.

Main Results:

• The proposed RSEN method demonstrated competitive performance against state-of-the-art techniques on three benchmark HSI datasets.
• The self-ensembling approach effectively utilized unlabeled data to improve classification accuracy in low-data regimes.
• The consistency filter contributed to the stability and effectiveness of the self-ensembling learning.

Conclusions:

• The RSEN offers a promising solution for hyperspectral image classification, particularly in scenarios with scarce labeled data.
• This work represents the first application of self-ensembling techniques to HSI classification, providing a new paradigm for leveraging unlabeled data.
• The proposed method enhances the robustness and accuracy of HSI classification models through innovative self-ensembling and data utilization strategies.