Reconstruction-Contrast Coupling Learning for Open-Set Semi-Supervised Hyperspectral Image Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Existing semi-supervised hyperspectral image (HSI) classification methods often assume a closed-set, where all data belong to known categories.
• This closed-set assumption is unrealistic in remote sensing due to the frequent presence of unknown categories in unlabeled data.
• This limitation hinders the practical application of HSI classification in real-world scenarios.

Purpose of the Study:

• To propose a novel open-set semi-supervised learning method for HSI classification that addresses the challenge of unknown categories.
• To leverage the synergy between masked feature reconstruction and contrastive learning for improved feature representation.
• To enhance the encoder's ability to capture both local spectral-spatial details and global semantic information.

Main Methods:

• Introduced Reconstruction-Contrast Coupling Learning (ReCo2L) for open-set semi-supervised HSI classification.
• Employed masked feature reconstruction with an adaptive masking strategy to improve local detail sensitivity.
• Utilized contrastive learning to enhance global feature discriminative ability and introduced a pixel-prototype deviation loss for better category separation.

Main Results:

• ReCo2L demonstrated superior classification performance on three benchmark datasets, effectively classifying both known and unknown categories.
• The proposed method significantly outperformed 10 state-of-the-art HSI classification techniques.
• The combination of reconstruction and contrastive learning proved effective in enhancing feature extraction for HSI data.

Conclusions:

• The ReCo2L method offers a robust solution for open-set semi-supervised HSI classification, overcoming the limitations of closed-set assumptions.
• The approach effectively enhances spectral-spatial feature representation by integrating reconstruction and contrastive learning objectives.
• The findings highlight the potential of ReCo2L for practical remote sensing applications requiring accurate classification of diverse land cover types.