Flood Detection in Satellite Images using Deep Learning and Quantum Computing

Abstract

Floods represent an important natural hazard that have historically impacted, and continue to impact, human lives and the economic infrastructure worldwide. It is therefore essential to detect those events for mitigating their impact and protecting populations. Quantum computing is an emerging field of computer science that attempts to solve computational problems using the laws of quantum mechanics. Deep learning, a branch of artificial intelligence, proved to be successful for image data analysis and pattern recognition. This thesis explores the intersection of deep learning and quantum computing by building a hybrid quantum-classical deep learning model architecture to detect floods in satellite images. The research consisted in preparing a dataset of flood and non-flood satellite images, and using it to train eight hybrid models with eight classical models, all based on pre-trained convolutional neural networks. The models were evaluated and compared in terms of their performance. Results showed that only three hybrid models achieved higher (or comparable) performance than their classical counterparts, while the remaining classical models outperformed their hybrid versions; however, the differences of performance were generally small, suggesting that hybrid quantum-classical models are already performing effectively. While quantum computing has a great potential for classical deep learning tasks, these findings suggest that further research is still needed to fully realize a quantum advantage.