Onboard hyperspectral image compression with deep line-based predictove architectures

AI-driven onboard compression of hyperspectral images remains a challenging problem due to the need to balance the high computational requirements of neural networks and their representational capability. So far, autoencoder-based approaches have dominated the literature and offer interesting performance at very low bitrates. However, they do not scale well into the high-quality, high-bitrate regimes, requiring large amounts of compute and memory and are significantly outperformed by classic and simpler approaches such as CCSDS-123. In this paper, we present a novel approach towards AI-driven onboard compression, consisting of two fundamental ingredients: i) neural predictive coding; ii) line-based architectures. With the former, we depart from the autoencoder approach to develop a neural network that predicts the value of a pixel based on a causal spatial-spectral context of past pixels. As in classic predictive coding, this allows to easily control maximum distortion and is particularly suited to the high-quality regime. While in-loop quantization is possible, the use of a prequantizer followed by lossless predictive coding offers the higher throughput at minimal rate-distortion suboptimality. The other ingredient of our approach seeks to limit computational requirements and, in particular, memory usage by developing a neural network architecture that can process one line with all its spectral bands at a time. Several mechanisms for sequence processing can be used to achieve this goal such as recurrent neural networks. However, we show that a recent hybrid recurrent-attentive operation can overcome the limitations of recurrent neural networks and offer Transformer-like performance with no need to store the whole line sequence in memory. This approach greatly limits memory requirements and potentially allows for continuous operation with pushbroom sensors. Experiments show that the proposed approach outperforms CCSDS-123 in lossless and near-lossless compression.