Efficient Lossy Satellite Image Compression Using Hybrid Autoencoder Model

The challenge of handling the vast amount of data generated by onboard satellite instruments arises due to limited bandwidth and memory capacity, which must be managed cau-tiously. Image compression is used as a tool to minimize the data size for easier storage and transmission between ground stations. An efficient compression method maintains the quality of the source representation in the reconstruction stage. Conventional lossy image compression techniques have been used for the last few decades. However, recent learning-based approaches have gained substantial interest in the artificial intelligence field of research that achieves highly promising image reconstruction results under adequate storage, and low bandwidth. In this paper, a hybrid approach was proposed, a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) autoencoder to improve compression efficiency. The proposed model was applied to the RGB benchmark satellite images dataset (EuroSat dataset). The experiment outcomes indicate that the proposed solution surpasses the Joint Photographic Experts Group (JPEG) family in regards to reconstruction evaluation criteria Peak Signal Noise Ratio (PSNR) and Structural Similarity Index (SSIM), with an improvement up to 34.7% and 6.36% respectively. Additionally, it outperforms the most recent deep learning approach, by 3.63% in terms of PSNR and 0.43% in terms of SSIM.