Satellite Image Dehazing using UNet

Abstract

Satellite Image Dehazing using UNet This project presents a deep learning approach for the restoration of clear visual data fromhaze-affected satellite imagery. Atmospheric haze poses a significant challenge in satellite imageanalysis, obscuring ground features and diminishing the utility of remote sensing data for criticalapplications such as environmental monitoring, urban development, and disaster response. Ourmethodology employs a U-Net convolutional neural network architecture, specifically adaptedfor image-to-image translation tasks, to effectively remove haze and reconstruct high-qualityimages. The model was trained and validated on the comprehensive Haze1k dataset, whichprovides a diverse collection of hazy and corresponding clear satellite images. We detail theimplementation of a custom PyTorch Dataset and DataLoader for efficient data handling andpreprocessing, including image resizing and tensor conversion. The U-Net model, optimizedwith the Adam optimizer and trained using the Mean Squared Error (MSE) loss function,learns an end-to-end mapping to transform hazy inputs into clear outputs. The performanceof the proposed model is rigorously evaluated on a separate test set comprising images withvarying degrees of haze density (thin, moderate, and thick). Quantitative analysis using PeakSignal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) demonstrates the model’ssuperior performance in haze removal and image quality enhancement compared to existingmethods. Visual comparisons between hazy, dehazed, and ground truth images further highlightthe effectiveness of our approach. This project contributes a robust and efficient solution forsatellite image dehazing, paving the way for improved accuracy in subsequent image analysisand interpretation tasks.