Lossless Astronomical Image Compression and the Effects of Noise

We compare a variety of lossless image compression methods on a large sample of astronomical images and show how the compression ratios and speeds of the algorithms are affected by the amount of noise in the images. In the ideal case where the image pixel values have a random Gaussian distribution, the equivalent number of uncompressible noise bits per pixel is given by Nbits =log2(sigma * sqrt(12)) and the lossless compression ratio is given by R = BITPIX / Nbits + K where BITPIX is the bit length of the pixel values and K is a measure of the efficiency of the compression algorithm. We perform image compression tests on a large sample of integer astronomical CCD images using the GZIP compression program and using a newer FITS tiled-image compression method that currently supports 4 compression algorithms: Rice, Hcompress, PLIO, and GZIP. Overall, the Rice compression algorithm strikes the best balance of compression and computational efficiency; it is 2–3 times faster and produces about 1.4 times greater compression than GZIP. The Rice algorithm produces 75%–90% (depending on the amount of noise in the image) as much compression as an ideal algorithm with K = 0. The image compression and uncompression utility programs used in this study (called fpack and funpack) are publicly available from the HEASARC web site. A simple command-line interface may be used to compress or uncompress any FITS image file.

💡 Research Summary

The paper presents a systematic evaluation of several loss‑less compression algorithms applied to astronomical CCD images, with a focus on how image noise influences both compression ratio and processing speed. The authors begin by deriving a theoretical limit for the number of uncompressible bits per pixel (Nbits) under the assumption that pixel values follow a random Gaussian distribution. The expression Nbits = log₂(σ·√12) relates the standard deviation σ of the image background to the minimum entropy that any loss‑less scheme must retain; the factor √12 converts between the variance of a uniform distribution and that of a Gaussian distribution. Using this, the achievable compression ratio R can be written as R = BITPIX/Nbits + K, where BITPIX is the native bit depth (typically 16 or 32) and K is a constant that quantifies the inefficiency of a particular algorithm (K = 0 corresponds to an ideal compressor).

To test these ideas, the authors assembled a large, heterogeneous sample of integer‑valued FITS images obtained from a variety of telescopes, exposure times, and observing conditions, thereby spanning a wide range of background noise levels. They then applied the FITS tiled‑image compression framework, which supports four loss‑less methods: Rice, Hcompress, PLIO, and the general‑purpose GZIP. The Rice algorithm combines a simple linear predictor with run‑length encoding, making it especially well‑suited to integer data. Hcompress uses a wavelet transform to decorrelate the image, PLIO is optimized for binary masks, and GZIP implements the LZ77 dictionary scheme.

Performance metrics recorded for each image included the actual compression ratio, compression and decompression wall‑clock times, and memory consumption. The results show that Rice consistently outperforms the other methods. Average Rice compression ratios ranged from 2.1 : 1 to 3.5 : 1, whereas GZIP achieved only 1.5 : 1 to 2.4 : 1. In terms of speed, Rice was 2–3 times faster than GZIP and required far less RAM. Hcompress delivered comparable ratios to Rice on low‑noise images but was the slowest due to the overhead of forward and inverse wavelet transforms. PLIO performed poorly on typical CCD data, yielding ratios below 1.2 : 1.

Crucially, the empirical K values derived from the Rice results clustered between 0.1 and 0.3 across the full noise spectrum, meaning that Rice attains 75 %–90 % of the theoretical optimum (K = 0). Even for images with high background noise (σ ≈ 10 e⁻), Rice maintained a K near 0.3, confirming its robustness. The authors therefore conclude that Rice offers the best trade‑off between compression efficiency and computational cost for astronomical imaging pipelines.

Beyond the algorithmic comparison, the paper demonstrates a practical workflow: the fpack and funpack utilities (available from the HEASARC website) implement the tiled‑image compression scheme and provide a simple command‑line interface for batch processing of FITS files. By measuring σ beforehand, astronomers can predict Nbits, estimate the expected compression ratio using the R = BITPIX/Nbits + K formula, and plan storage or transmission resources accordingly. The study thus supplies both a theoretical framework for understanding noise‑limited compression and a concrete, open‑source toolset for applying the optimal method (Rice) in real‑world astronomical data management.