Florida International University
Lawrence Egharevba is a Ph.D. candidate at the Knight Foundation School of Computing and Information Sciences at Florida International University. Under the guidance of Dr. Naphtali Rishe, his research focuses on satellite imagery, cloud detection and extraction, image processing, and machine learning. A graduate of Florida International University, he earned his Master’s in Telecommunications and Networking in Spring 2021 and began his Ph.D. studies in Fall 2021. He has been an NSF Graduate Research Fellow for two consecutive years and received a Department of Homeland Security grant. In his recently published paper for IPSI Transactions, he explores the application of Deep Learning Techniques in Satellite Imagery for Cloud Extraction. Among his research interests are ecological research and Geographic Information Systems. His passion is applying Machine Learning for Atmospheric Science Analysis.
Remote sensing is the interpretation and inversion of radiometric measurements of electromagnetic radiation from a distance. Electromagnetic radiation can be classified according to certain basic parameters, such as wavelength, amplitude, direction of propagation, and polarization. It is possible to illustrate the relative proportions of reflected and emitted radiation as a function of wavelength, the emissivity of the surfaces observed, and the solar illumination of the area under observation using the Planck radiation distribution function. Atmospheric molecules such as clouds absorb wavelength-specific radiation. Acquired signals from remote sensing are significantly distorted by the fact that the region cloud cover is not adequately captured. Considering the amount of cloud cover the Earth experiences every day, the use of remote sensing for land surface studies is greatly compromised by cloud contamination. As a result, remote sensing images (data) are less useful, and imagery analysis is more challenging. Clouds impede the practical application of remote sensing images. Thus, removing clouds for effective analysis and interpretation of remote sensing data is a significant challenge.
The main objective of the proposed dissertation is to analyze, explore, and implement deep learning models and architectures for recovering missing information. The novel model learns to capture the statistical distribution of training data, which allows the synthetization of novel samples from the learned distribution. The proposed novel techniques can be used to perform image retrieval and classification tasks using this information. In particular, the proposed research aims to retrieve cloudy imagery from satellite imagery by learning the statistical distribution of cloud-free images from the training data. The learned representation of the novel model can be used to synthesize cloud-free satellite imagery. This research work will implement a novel generative adversarial networks model that can: (1) translate a cloudy image into a cloud-free image for effective analysis and interpretations of satellite imagery; and (2) remove cloud-contaminated regions from satellite imagery.
Our proposed model is expected to perform at the highest level of removing thick cloud regions from satellite imagery. This proposed cloud removal method will greatly improve Earth’s surface observation and monitoring from space, which is essential for managing climate change, natural resources, and mitigating natural disasters. Our cloud removal technique will also improve the usefulness of satellite imagery in a variety of fields by providing more accurate and clearer data for analysis, research, and decision-making.