Quality Enhancement for Remote Sensing Imagery with Deep Learning

Abstract

Machine learning (ML) techniques have been integral to environmental remote sensing, particularly as the availability of "big data" from Earth observations has grown. The evolving capabilities of ML offer significant opportunities for improving environmental monitoring of the Earth through remote sensing imagery, which provides crucial data for various applications, including environmental monitoring, urban planning, and disaster management. Remote sensing data, acquired through active or passive imaging, is often compromised by noise and artefacts such as speckle in synthetic aperture radar (SAR) images, cloud cover in optical images, and resolution issues in multispectral images. This dissertation addresses these challenges by introducing innovative deep learning methods to enhance the quality of remote sensing data. Three main contributions are presented: 1. SAR Despeckling: A Residual-in-Residual Dense Generative Adversarial Network is developed to reduce speckle noise in SAR images while preserving essential geographical features. This method surpasses traditional filters by leveraging deep learning to better adapt to the complex statistical nature of SAR imagery. 2. Thin Cloud Removal: A novel deep learning architecture is proposed for removing thin clouds from optical images. This method employs multi-scale feature extraction and attention mechanisms, allowing for the selective processing of clouds and the accurate restoration of obscured surfaces. This approach significantly improves upon traditional methods, which often result in loss of critical information. 3. Multispectral Pansharpening: The dissertation advances pansharpening techniques by proposing a Heterogeneous Two-stream Network (HTSNet) with Hierarchical Feature Prefusion. This method effectively combines high-resolution panchromatic images with multispectral data, resulting in superior pansharpened imagery that supports more accurate decision-making in various applications. These contributions represent significant improvements over existing state-of-the-art techniques, providing automated, robust solutions for enhancing the quality of remote sensing imagery. The deep learning methods introduced not only improve the clarity and usability of the data but also ensure adaptability and accuracy, offering substantial benefits for a wide range of environmental and industrial applications.

Date of Award	2024
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Changjing Shang (Supervisor) & Qiang Shen (Supervisor)