Transfer Learning-Based Hyperspectral Image Classification Using Residual Dense Connection Networks.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image classification faces challenges due to limited labeled samples and uneven distribution, hindering effective feature extraction.
• Traditional few-shot learning methods struggle with complex hyperspectral data, leading to suboptimal classification performance.
• High costs of sample annotation further complicate the extraction of discriminative features from limited data.

Purpose of the Study:

• To develop an effective hyperspectral image classification approach for scenarios with insufficient labeled samples.
• To integrate advanced spectral-spatial feature extraction with meta-transfer learning to enhance classification accuracy.
• To address the limitations of existing methods in deriving deep, discriminative features from complex hyperspectral signals.

Main Methods:

• Proposed an integration of spectral-spatial feature extraction with meta-transfer learning for hyperspectral signal classification.
• Employed dense connection blocks and 3D convolutional residual connections to improve feature extraction.
• Utilized a model pre-trained on a large source domain dataset and transferred to a target domain with minimal samples.

Main Results:

• The proposed method significantly surpasses existing classification algorithms and small-sample techniques in accuracy.
• Demonstrated superior performance in extracting deep, discriminative features from limited hyperspectral data.
• Achieved high accuracy on diverse hyperspectral datasets (IP, UP, and Salinas) under label constraints.

Conclusions:

• The combined spectral-spatial feature extraction and meta-transfer learning approach effectively addresses hyperspectral image classification with limited labeled data.
• The method enhances spatial and spectral information retrieval, leading to improved classification performance.
• This technique is applicable to high-dimensional signal classification tasks where labeled samples are scarce.