A smart compression approach core based on the CCSDS 123.0-B-2 standard for CHIME

Abstract

The Copernicus Hyperspectral Imaging Mission for Environment (CHIME) is introduced by ESA in the future Copernicus 2.0 program to provide routine hyperspectral observations for the monitorization of natural resources, including applications such as sustainable agricultural and biodiversity management, soil properties characterization, sustainable mining practices and environment preservation [1]. CHIME shall provide continuous spectral data in the range between 400 and 2500 nm at input data rates up to 2 Gbps. A mission requirement is to employ adaptive compression, detecting clouds in the scene and introducing in those pixels a higher level of losses. This is relevant because clouds, which it is estimated that cover more than 54% of the Earth’s land area and 68% of the oceans, make more than half of the acquired scenes unusable for scientific applications. The solution adopted for CHIME is to modify the CCSDS 123.0-B-2 Low-Complexity Lossless and Near-Lossless Multispectral and Hyperspectral Image Compression [2] to achieve this goal, while at the same time a low-complexity algorithm is provided, capable to compress in lossless and near-lossless modes. In the CHIME pre-development phase (A/B1), a proof of concept was developed and tested [3]. This solution, developed following an HLS workflow, do not achieve the goals in terms of throughput defined for the mission. Besides, different modifications to the standard were analysed at algorithmic level in [4], in order to remove clouds from the acquired scenes. From this study, it was concluded that the Different Absolute Error (DAE) strategy provides promising results for selective compression in terms of both compression ratio (a data reduction between 20% and 35% compared to the compliant CCSDS 123.0-B-2 algorithm for cloud coverage around 40%) and image quality after decompression. Cloud detection is also performed by a Support Vector Machine (SVM) approach that allows identifying each pixel of the image as ground or cloud. These 2 classes are used to drive the selective compression with DAE settings for cloud and ground pixel classes. The algorithm is divided in four steps: Spectral Band Selection for performing cloud detection; Top of Atmosphere conversion of the pixels from the selected bands; SVM processing; and Morphological Cloud Map filtering. This work presents a VHDL-made hardware solution, based on the CCSDS 123.0-B-2 standard and including cloud detection and the DAE approach, to efficiently compress hyperspectral images acquired by the CHIME instrument. Only the prediction features that maximize the throughput under BIL order are implemented. As for the HLS design, the block-adaptive alternative is used for entropy coding. The inclusion of the DAE approach implies that the IP core is able to compress clouds present in the scene with a higher level of losses than the rest of the image and thus improving the compression ratio. Architecture and behaviour of the proposed IP core are detailed, and preliminary results in terms of area footprint and throughput are presented.