Exploring Generative Adversarial Networks (GANs) for cloud detection and cloud-obstructed region reconstruction from satellite imagery.
Remote sensing stands as a technique capable of comprehensively monitoring the Earth’s surface, providing valuable insights across a wide array of domains. While satellite imagery plays a vital role in this context, the presence of cloud cover introduces a significant challenge, obstructing accurate analysis and interpretation. This project intends to leverage the capabilities of Generative Adversarial Networks (GANs) to effectively address cloud detection and reconstruct cloud-obstructed regions in satellite images. Existing research has explored two primary approaches in cloud detection: classical algorithm-based and machine learning-based methods. Classical algorithms rely on threshold-based techniques and exhibit limitations due to the variability of environmental conditions. On the other hand, machine learning methods offer adaptability but require a robust training dataset. Specific atmospheric conditions such as the presence of aerosols can impede accuracy by scattering light. However, the reliability of detection depends on the satellite sensor’s spatial and spectral resolution, with modern advanced sensors providing more accurate results. Furthermore, different cloud types vary in detectability, with thick clouds being more detectable than thin or low-lying ones. To strengthen the accuracy of cloud detection, researchers have recommended integrating diverse atmospheric models, accounting for physical parameters, and adopting hybrid methodologies that combine atmospheric data with artificial neural networks. GANs have gained prominence in various fields, including image generation and translation. They consist of two neural networks: the Generator, which aims to capture the true data distribution, and the Discriminator, a binary classifier distinguishing between genuine and generated samples. GANs are known for their ability to generate synthetic data that closely resembles real data. Therefore, their utilization in cloud-obstructed image reconstruction of satellite imagery is beneficial for enhancing the accuracy and reliability of remote sensing applications. During this research, reconstructed satellite images will be employed in an environmental monitoring task (such as post-disaster damage assessment, vegetation growth and deforestation, coastal erosion and shoreline changes, etc.), focusing on a specific geographical location (the location will be considered having variable weather conditions, availability of adequate data sources, and depending on the selected monitoring activity). The research will perform cloud detection and masking of the pre-processed data and enhance the images through data fusion. Quantitative analysis will be performed following feature selection and analysis. The project will visualize and report on the findings to support decision-making, and therefore, continuous monitoring can be carried out eventually. In summary, this project pioneers the integration of GANs in the domain of cloud detection and cloud-obstructed region reconstruction from satellite imagery. By leveraging the transformative potentials of GANs, this research aims to enhance the accuracy and utility of satellite data, contributing to improved Earth observation and remote sensing applications.