SPanet: Intelligent Invertible Super-Panchromatic Neural Video Transcoder for Scene-Adaptive Satellite Surveillance

Abstract

Onboard video transmission in satellite surveillance missions is critically bottlenecked by the limited processing capability of radiation-hardened hardware and the scarcity of downlink bandwidth. To address these challenges, this paper proposes SPanet, a scene-adaptive, low-bitrate video compression framework designed with onboard deployment considerations. The proposed architecture utilizes a three-stage dimensionality reduction pipeline to maximize coding efficiency while reducing computational overhead. First, an analytically invertible grayframe generation (IGG) module employs a hardware-efficient wavelet transform to map colour frames into a super-panchromatic representation, reducing the encoder input dimensionality by approximately 66.6% while preserving chromatic cues. Second, this representation is compressed into a compact latent stream leveraging an intra-frame neural compression (INC) autoencoder, enhanced by instance-specific weight adaptation (ISWA), which models the network as a mission-specific digital lens, tailored to exploit geographic scene redundancy. Finally, an inter-frame residual temporal compression (IRTC) module ensures temporal consistency by alpha-dampened blending of transformed difference frames with reconstructed output of INC, suppressing motion-induced noise in low signal-to-noise ratio (SNR) environments. Experimental results on benchmark satellite video datasets demonstrate consistent improvements in rate–distortion performance and perceptual quality at extremely low bitrates. At 0.2 bits-per-pixel (bpp), the proposed framework achieves average gains of 1.17dB in PSNR and 1.01% in SSIM over state-of-the-art learned codecs. This integrated design establishes a compute-efficient and adaptive compression paradigm, enabling onboard edge processing and scene-adaptive optimization, while maintaining low computational complexity for next-generation satellite video systems.