SYNTHETIC APERTURE RADAR (SAR) GUIDED RECONSTRUCTION AND CLOUD REMOVAL IN OPTICAL SATELLITE IMAGES

Abstract

Cloud cover affects approximately 60–67% of the Earth’s surface at any given moment, creating systematic gaps in optical satellite imagery that disrupt land monitoring, agricultural assessment, and disaster response workflows. Existing deep learning approaches to cloud removal share a fundamental deficiency: they process every pixel through the generative network indiscriminately, modifying cloud-free observations that require no reconstruction and occasionally hallucinating surface features that do not exist. This thesis proposes CMSRNet (Cloud-Aware Multi-Scale SAR-Guided Reconstruction Network), a lightweight deterministic architecture that addresses these deficiencies using co-registered Sentinel-1 SAR data as a structural prior for Sentinel-2 optical reconstruction. The architecture consists of three principal components. First, a dual-encoder design processes optical and SAR inputs through separate convolutional pathways before fusion, preventing the lower-information SAR signal from diluting optical feature representations. Second, MultiScaleSARFusion blocks integrate the two modalities at three encoder scales using channel-wise gating and spatial attention, allowing SAR features to guide optical reconstruction at both coarse structural and fine textural levels. Third, a CloudAwareRefinement module produces an explicit cloud importance map as a direct network output and applies a hard-replacement constraint: pixels identified as cloud-free are copied exactly from the input to the output, providing a strict mathematical guarantee that valid surface observations are never modified. CMSRNet is trained on the ROIs1158 spring subset of the SEN12MS-CR benchmark using a cloud-aware loss function that concentrates reconstruction penalties on cloud-affected regions. The network contains 3.26 million parameters and runs in a single forward pass, avoiding the iterative sampling cost of diffusion-based alternatives such as DiffCR. Ablation experiments validate the individual contribution of each architectural component. The results demonstrate that a deterministic, mask-aware architecture can achieve reasonable reconstruction quality within the operational constraints of large-scale satellite data processing.