Flood Mapping Using Satellite Images

Abstract

The aim of this thesis is to study the subject of flood mapping utilizing images captured from satellites and machine learning algorithms. Since the rise of ESA's Copernicus program and the consecutive Sentinel satellites a vast number of freely available images are captured every day expanding the potential applications. With the launch of Sentinel 1 and Sentinel 2 sensing the planet Earth in an unprecedented frequency and spatial resolution, scientists and engineers can now develop tools in order to understand the processes of the Earth and make more informed decisions. Floods are one of the most devastating natural disaster affecting many people each year, causing a lot of deaths, infrastructure damages and loss of properties. In order to mitigate the effects of floods on people critical decision making is needed, which can be assisted by satellite images and machine learning methods. This study examines the performance of three different machine learning methods in identifying pixels of satellite images containing flooded areas. More speci cally, the three tested methods are based on deep learning architecture, transfer learning and traditional swallow learning pixel based semantic segmentation, consequently. In particular, the deep learning method is based on the UNET architecture, the transfer learning uses as backbone the VGG16 network and the traditional swallow machine learning approache utilizing a Randome Forest. Experiments involve training models either through strict or through weak supervision as well as multimodal feature spaces, combining sentinel 1 and sentinel 2. Finally, the machine learning techniques are compared in terms of performance with a technique based on a single image band histogram segmentation, called as baseline model.