Use of customizing kernel sparse representation for hyperspectral image classification.

Area of Science:

• Remote Sensing
• Machine Learning
• Hyperspectral Imaging

Background:

• Sparse Representation-based Classification (SRC) is effective for hyperspectral data.
• Kernel Sparse Representation-based Classification (KSRC) extends SRC non-linearly.
• KSRC lacks integration of prior domain knowledge, like spectral feature importance.

Purpose of the Study:

• To propose Customizing Kernel Sparse Representation-based Classification (CKSRC).
• To incorporate kth nearest neighbor density as a weighting scheme into kernels.
• To enhance classification accuracy and robustness in hyperspectral imagery.

Main Methods:

• Developed CKSRC by integrating kth nearest neighbor density into kernel functions.
• Applied CKSRC to two publicly available hyperspectral datasets.
• Compared CKSRC performance against other classification algorithms.

Main Results:

• CKSRC demonstrated increased overall classification accuracy.
• The method showed robust classification results across different training sample sets.
• CKSRC outperformed existing classification algorithms on the tested datasets.

Conclusions:

• CKSRC offers a significant advancement over traditional KSRC.
• The integration of domain knowledge via density weighting improves hyperspectral classification.
• CKSRC provides a more accurate and reliable approach for hyperspectral image analysis.