A Low-Measurement-Cost-Based Multi-Strategy Hyperspectral Image Classification Scheme.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification is crucial but hindered by the high cost of pixel annotation.
• Reducing labeled data while maintaining accuracy is a key challenge in HSI classification.

Purpose of the Study:

• To develop a novel classifier that effectively addresses the issue of limited labeled data in HSI classification.
• To improve learning strategies for enhanced feature extraction and classification accuracy.

Main Methods:

• Introduced a multi-strategy triple network classifier (MSTNC) incorporating contrast learning for low sample dependence.
• Implemented an active learning strategy with a novel feature-mixed active learning (FMAL) method for valuable pixel selection.
• Utilized dual-threshold pseudo-active learning (DSPAL) to expand the training set without increasing labeling costs.

Main Results:

• The MSTNC demonstrated superior performance compared to state-of-the-art methods across various labeling ratios on benchmark HSI datasets.
• Achieved high overall accuracy under extreme small-sample conditions (e.g., 82.97% on IP, 87.94% on PU, 86.57% on WHU).

Conclusions:

• The proposed MSTNC effectively reduces the dependence on labeled data while significantly improving HSI classification accuracy.
• The combination of contrast learning, active learning, and pseudo-active learning offers a robust solution for HSI classification with limited annotations.