Compression of multispectral images by spectral classification and transform coding.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Multispectral images contain rich spectral information crucial for land cover analysis.
• Conventional compression techniques often struggle to efficiently exploit the spatial and spectral redundancies in multispectral data.
• Existing methods may not optimally handle the varying statistical properties across different land cover types within a single image.

Purpose of the Study:

• To develop and evaluate a new compression technique for multispectral images.
• To improve rate distortion performance by leveraging image segmentation based on land cover homogeneity.
• To effectively exploit statistical redundancies within spectrally similar regions.

Main Methods:

• Image segmentation into homogeneous land cover regions.
• Vector quantization for pixel classification and minimum distortion encoding.
• Entropy encoding of the classification map as side information.
• Karhunen-Loeve transform (spectral domain) and Discrete Cosine Transform (spatial domain) on residual vectors grouped by class.

Main Results:

• The proposed technique demonstrates improved rate distortion performance compared to conventional transform coding.
• Achieved performance gains of 1 to 2 dB across various data rates.
• Validated on a six-band thematic mapper image, showing effectiveness for real-world remote sensing data.

Conclusions:

• Segmentation-based compression effectively exploits stationary statistics and linear dependencies within homogeneous land cover regions.
• The proposed method offers a significant improvement over traditional transform coding for multispectral imagery.
• This approach provides a more efficient way to compress multispectral data for applications like remote sensing and environmental monitoring.