Unsupervised Meta Learning With Multiview Constraints for Hyperspectral Image Small Sample set Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning models for hyperspectral image (HSI) classification require substantial labeled data, hindering accuracy.
• Existing meta-learning methods reduce sample dependence but still require extensive labeled source datasets for meta-training.
• This process is time-consuming and labor-intensive, necessitating a more efficient approach.

Purpose of the Study:

• To propose a novel unsupervised meta-learning method for HSI classification that minimizes the need for labeled samples.
• To overcome the limitations of existing supervised meta-learning techniques in terms of data requirements.
• To enhance the classification accuracy of deep learning models in small sample set scenarios.

Main Methods:

• An unsupervised meta-learning framework utilizing unlabeled hyperspectral images.
• Generation of multiple spatial-spectral multiview features from unlabeled samples to construct meta-learning tasks.
• A residual relation network for meta-training and classification using a voting strategy.

Main Results:

• The proposed unsupervised method effectively utilizes unlabeled HSIs, significantly reducing the requirement for labeled data.
• Experimental results on 8 public HSI datasets demonstrate competitive or superior performance compared to supervised meta-learning and other advanced models.
• The method shows effectiveness in both cross-domain and in-domain classification scenarios with small sample sets.

Conclusions:

• The developed unsupervised meta-learning approach offers a viable solution for HSI classification with limited labeled data.
• This method significantly reduces the practical challenges associated with data acquisition and labeling in HSI analysis.
• The approach holds promise for improving the efficiency and accuracy of hyperspectral image classification in resource-constrained settings.