TRUE-LOSSLESS AUTOENCODER RESIDUAL COMPRESSION (TLARC) FOR HIGH-RESOLUTION SATELLITE IMAGERY

Abstract

The high rate of proliferation of earth observation missions has resulted in the production of very large quantities of high-resolution satellite imagery with on-board processing posing serious challenges because of storage capacity, bandwidth, and real time transmission expense. The classical lossless and near-lossless compression methods are based on predetermined transformations, including DCT, DWT, or predictive entropy coding, that are typically not capable of deriving the spatial-spectral correlation and structural variations in satellite data. To overcome such drawbacks, we introduce a True-Lossless Autoencoder Residual Compression (TLARC) model consisting of a convolutional autoencoder plus deterministic residual compression in order to provide a model with a perfect pixel-level reconstruction and high compression efficiency. The encoder is trained on a small data-driven latent representation that can be on different land-cover types, cloud coverage, and illumination levels. A quantization module encodes Latent activation values into values that can be stored as integers, and the values corresponding to the quantization differences, which are usually discarded in the standard autoencoder approaches, are encoded as tensors of the residual values. These tensors are reintroduced back after decoding to obtain perfect reversibility and actual lossless reconstruction. Largely-scale experiments have shown that TLARC has always been offering an ideal fidelity, lower storage and better compression. The proposed method has a compression ratio of 2.18x +- 0.08, bit rate of 10.98 bpp, and space savings of over 54% and the encoding and decoding time are 0.182 s and 0.157 s respectively. These results affirm the fact that TLARC is better than the conventional compression techniques.