DISASTER DAMAGE DETECTION FROM OPTICAL EARTH OBSERVATION DATA

Abstract

Natural disasters inflict severe damage on landscapes and communities, necessitating rapid and accurate assessment for effective disaster response and recovery. Traditional methods of damage assessment, primarily based on manual analysis of satellite imagery, are time consuming and prone to human error. This thesis explores the potential of deep generative networks as an automated solution for disaster damage detection from satellite images, aiming to enhance the speed and accuracy of the assessments. Deep learning, specifically deep generative networks such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) have shown promise in image processing and anomaly detection. This research focuses on leveraging these models to identify and quantify damage in areas affected by disasters, using high resolution satellite imagery. The model is designed to process satellite images, identify damage patterns and quantify the extent of the destruction. The performance of this model is evaluated based on accuracy, efficiency and robustness across various types of disasters including earthquakes, floods, hurricanes and wildfires. The findings indicate that deep generative networks can effectively identify subtle and complex damage patterns that are often overlooked in manual analyses. The thesis concludes with a discussion of the implications of these results for disaster management, highlighting the potential of this approach to transform disaster response strategies by providing quicker and more reliable damage assessments.. this research contributes to the field by showcasing how deep generative networks can be applied to enhance disaster damage detection from satellite images, offering a path forward for more efficient and effective disaster response and recovery efforts.