An end-to-end deep learning image compression method for satellite images based on entropy model

Area of Science:

• Remote Sensing
• Image Processing
• Artificial Intelligence

Background:

• Satellite image data is crucial for environmental monitoring and mapping.
• Existing compression methods struggle with high-resolution, multi-spectral satellite data, leading to inefficiency and quality loss.
• The rapid increase in satellite data volume presents transmission and storage challenges.

Purpose of the Study:

• To develop an end-to-end deep learning framework for efficient compression of VIIRS satellite imagery.
• To improve compression efficiency and reconstruction image quality compared to existing methods.
• To address the challenges of transmitting and storing large volumes of satellite data.

Main Methods:

• An end-to-end deep learning framework incorporating an analysis transform encoder, synthesis transform decoder, hybrid quantizer, and probability model for entropy coding.
• Utilized a Gaussian-mixture entropy model with a cumulative distribution function (CDF) for likelihood computation.
• Integrated a checkerboard context structure and a VIIRS-specific block-processing pipeline, including a block compression strategy for large images.

Main Results:

• Achieved a compression rate of 0.51 bpp, 38.39 dB PSNR, and 0.973 SSIM on NASA VIIRS datasets.
• Outperformed ELIC and Transformer-CNN baselines by reducing bpp by 13.6% and 10.5%, and improving PSNR by 0.78 dB and 0.61 dB, respectively.
• Compressed data to 1.5%-4% of original size (approx. 30:1 compression ratio), with block compression enabling reconstruction of large images.

Conclusions:

• The proposed deep learning framework offers superior performance for VIIRS satellite image compression.
• The method effectively balances compression efficiency with high reconstruction quality.
• The block compression strategy ensures scalability for processing high-resolution satellite imagery.