Destripe Any Scale: Effective Stripe Removal Via Multiscale Decomposition Using the Luojia3-02 Stripe-Noise Dataset

Stripe noise is a significant challenge in remotely sensed satellite images, primarily originating from inconsistencies in push-broom scanning systems. This noise manifests as linear, repetitive patterns that degrade image quality and impede the accurate data analysis. Existing destriping methods, including traditional statistical and frequency-based techniques, optimization-based approaches, and deep-learning models, each exhibit limitations in effectively addressing both small-scale and large-scale stripe noise, particularly in complex and heterogeneous imaging scenarios. To overcome these challenges, we present Destripe Any Scale, a novel hybrid destriping framework that integrates multiscale decomposition with deep-learning techniques. We first introduce the Luojia3-02 stripe-noise dataset, constructed from Luojia3-02 (Wuhan-1) satellite imagery, including stripe-noise patterns across multiple spatial scales. Building upon this dataset, we propose a multilevel perceptual stripe-noise-removal network. This network employs an independent-weight learning strategy to separately capture and optimize large-scale and small-scale noise features, facilitating the precise extraction of multiscale noise features. To further enhance performance, we construct a multiscale image pyramid that decouples mixed noise into distinct resolution levels. Within this framework, stripe noise is progressively suppressed across hierarchical levels, while a column-guided correction mechanism ensures consistency across scales and preserves the fidelity of image details. Experimental evaluations using Luojia3-02 data demonstrate that, without relying on external calibration, the proposed network and processing framework effectively eliminate multiscale stripe noise caused by sensor-induced radiometric errors, significantly improving both radiometric quality and visual clarity. Compared with traditional methods, our approach exhibits superior robustness and adaptability in handling complex, large-swath scenarios, demonstrating the advantages of integrating deep-learning techniques with multiscale decomposition for stripe-noise removal.