Onboard deep lossless and near-lossless predictive coding of hyperspectral images with line-based attention

TL;DR

The paper tackles onboard hyperspectral image compression under tight computational and memory constraints by moving away from traditional autoencoders to a predictive coding framework. It introduces LineRWKV, a line-based predictive model that recursively processes image lines using RWKV-inspired line and spectral predictors, achieving linear memory and parallel training while maintaining Transformer-like context. Empirical results on HySpecNet-11k and transfer tests on PRISMA show that LineRWKV outperforms the CCSDS-123.0-B-2 standard in lossless and near-lossless regimes, with substantial gains that scale with model size, and feasible throughput on a 7 W embedded platform. The approach demonstrates the viability of deep-learning-based onboard predictive coding for high-rate, high-resolution hyperspectral data, with potential extensions to further throughput optimization and broader mission applicability.

Abstract

Deep learning methods have traditionally been difficult to apply to compression of hyperspectral images onboard of spacecrafts, due to the large computational complexity needed to achieve adequate representational power, as well as the lack of suitable datasets for training and testing. In this paper, we depart from the traditional autoencoder approach and we design a predictive neural network, called LineRWKV, that works recursively line-by-line to limit memory consumption. In order to achieve that, we adopt a novel hybrid attentive-recursive operation that combines the representational advantages of Transformers with the linear complexity and recursive implementation of recurrent neural networks. The compression algorithm performs prediction of each pixel using LineRWKV, followed by entropy coding of the residual. Experiments on the HySpecNet-11k dataset and PRISMA images show that LineRWKV is the first deep-learning method to outperform CCSDS-123.0-B-2 at lossless and near-lossless compression. Promising throughput results are also evaluated on a 7W embedded system.

Figures (6)

• Figure 1: Overview of LineRWKV for $z>1$ and $y>0$. Encoder, decoder, line and spectral predictors are neural networks trained end-to-end to minimize the prediction error for all pixels. Each square represents a pixel in the corresponding spatial-spectral position, as it is processed by the neural modules. Rounded prediction errors are then entropy-coded with a standard code.
• Figure 2: Line and spectral predictors are made of the repetition of line/band mixers and channel mixers according the depicted architecture. $z^{-1}$ denotes unit delay in the sequence dimension (lines or bands). LN denotes LayerNorm and LIN denotes linear projection of features.
• Figure 3: Rate-PSNR performance comparison on HySpecNet-11k hard test set.
• Figure 4: Left to right: band 187 of a test image (colormap $[0,4000]$, positive-mapped prediction residuals (colormap $[0,100]$) for CCSDS-123.0-B-2 and LineRWKV-XS.
• Figure 5: Positive-mapped prediction residual for CCSDS-123.0-B-2 and LineRWKV-XS. Median, over the entire HySpecNet-11k hard test set, of spatial medians.