A New Feature Set for Texture-Based Classification of Remotely Sensed Images in a Quantum Framework.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Texture feature extraction is vital for land-use and land-cover (LULC) classification of remote sensing images.
• Quantization of images to limited gray levels degrades conventional texture descriptors like Local Binary Patterns (LBP).
• Gray-Level Co-occurrence Matrix (GLCM) captures macro-texture but can be sensitive to noise.

Purpose of the Study:

• To introduce Singular Values of the Gray-Level Co-occurrence Matrix (SVGM) as a novel texture feature set.
• To evaluate SVGM's effectiveness in LULC classification under coarse quantization.
• To compare SVGM against traditional texture descriptors (LBP, CLBP, Haralick's GLCM) and assess its performance with various classifiers.

Main Methods:

• Investigated singular values derived from GLCM matrices.
• Compared SVGM with LBP, CLBP, and original GLCM features.
• Evaluated performance using Support Vector Machines (SVMs) with classical and quantum kernels, and neural networks (ANN, 1D-CNN).

Main Results:

• SVGM demonstrated superior class separability compared to LBP and its variants, especially under coarse quantization.
• SVGM effectively preserves dominant spatial structures while suppressing noise and redundancy.
• SVGM consistently improved classification performance across all tested models, with quantum kernel SVMs showing competitive results.

Conclusions:

• SVGM presents a robust and effective texture feature for LULC classification, particularly for quantized remote sensing data.
• The proposed SVGM method offers a valuable alternative to conventional texture descriptors, enhancing classification accuracy.
• The study highlights the potential of SVGM in conjunction with both classical and quantum machine learning models for improved image analysis.