If lossless compression is required, the 2 most efficient methods that were evaluated are the minimum variance and minimum entropy neighbourhood-pixel-based predictors. These approach a 2:1 compression ratio at 4.85 bits per pixel.
In the case of progressive encoding, pyramid encoding does an adequate job. It does not yield great compression, but is nevertheless more compact than the raw data at 6.76 bits per pixel.
If lossy compression is permissable, the optimal compression method found (by rate) is DCT compression after the mean is subtracted from the image. The coefficients should then be uniformly quantized, the DC coefficients DPCM coded, and the zig-zag reordered coefficients RLE encoded.
If, as is normally the case, the visual appearance of the images is more important than the SNR, then non-uniform quantization according to the CCIT-601 recommendations should be performed in place of uniform quantization.