Background: Optical satellite imagery is integral to various Earth observation initiatives, including environmental monitoring, urban planning, and disaster management. However, approximately 67% of the Earth’s surface is obscured by clouds and their shadows, which significantly hampers ground visibility and limits the effectiveness of this imagery. Traditional techniques for cloud removal typically depend on multi-temporal data, which is not always readily available, rendering these methods less reliable in regions characterized by persistent cloud cover.

Objectives: This thesis endeavors to develop a comprehensive synthetic data generation framework, integrated with a deep learning architecture, aimed at reconstructing cloud-free satellite images. The primary objectives include: (1) the creation of realistic synthetic overlays of clouds and shadows for training purposes, (2) the design of an efficient model that accurately performs image reconstruction, and (3) the evaluation of the model’s performance in comparison to established generative adversarial networks.

Methods: A novel Progressive Multi-scale Attention Autoencoder (PMAA) was designed to eliminate shadows and clouds from single-image inputs using attention refined processing (ARPD) at different scales. A dataset based on Perlin noise was developed for volumetric rendering to simulate varying atmospheric conditions. The model was trained and compared with CTGAN and STGAN baselines using different metrics of PSNR, SSIM, MSE, RMSE, MAE and Inference Time. Additional ablation experiments evaluated the contribution of architectural components to the design.

Results: PMAA significantly outperformed baseline models, achieving a PSNR of 35.08 dB and an SSIM of 0.9673, indicating superior reconstruction quality and structural fidelity. It demonstrated robustness across scenarios with varying cloud densities and shadows, while remaining computationally efficient (average inference time of 0.0341s per image). The attention mechanism and progressive refinement were confirmed as key contributors through ablation analysis.

Conclusions: This study confirms the effectiveness of synthetic training data and multi-scale attention in enhancing satellite image clarity under atmospheric occlusions. The PMAA framework sets a new benchmark for cloud and shadow removal, offering high-quality reconstructions with practical deployment potential in real world Earth observation applications.