LEO-VE: Low-Light Earth-Observation Video Enhancement with Linear-Memory Spatio-Temporal Modeling

<jats:p>Nighttime and low-light Earth observation offers unique value for urban mapping, volcanic monitoring, and disaster response, yet captured video suffers from low signal-to-noise ratio, underexposure, and motion blur. Existing low-light enhancement methods restore frames largely independently and neglect temporal coherence, trustworthiness, and edge deployability. We propose LEO-VE (Low-light Earth-observation Video Enhancement), a dual-branch spatio-temporal network that combines physics-informed static degradation modeling with a dynamic branch using local–non-local and linear-memory long-range attention to enforce temporal consistency without quadratic memory cost. A perceptual adversarial loss suppresses artifacts, a watermark and manipulation-localization head makes outputs forensically traceable, and a risk-adaptive KV compression enables edge inference. On the LEO-Syn benchmark, LEO-VE achieves 25.41 dB PSNR, 0.903 SSIM, and 0.118 LPIPS at 69 ms per frame, surpassing LVE-S2D (23.67 dB) and all baselines. Enhanced video yields downstream gains of 66.9 mAP on nighttime change detection and 63.8 IoU on thermal-glow segmentation. After edge compression, memory drops from 9.8 GB to 3.1 GB and latency to 41 ms/frame while retaining 25.18 dB PSNR, demonstrating practical on-board deployment.</jats:p>