Performance and Transferability Assessment of Convolutional Neural Network (CNN) Based Building Detection Models for Emergency Response

Remote Sensing data from Earth Observation (EO) is used for a wide variety of applications. Over the last decade, in the event of a natural calamity, the importance of using geo referenced products from satellite and aerial imagery has been on the rise. They play a vital role in helping the first responders by providing valuable information in the form of hazard zone maps that help in relocation of people, in post disaster evaluation to get a better understanding of the impact on the disaster zone and in the rehabilitation and reconstruction of damaged property. In remote sensing-based emergency mapping, there are major limitations during the acquisition and processing of earth observation data. In most cases, satellite data can be acquired only from that set of EO satellites that are in orbit over the hazard zone during the time of the disaster. This can be compensated by deploying sensors on board airplanes and Unmanned Aerial Vehicles (UAVs) like drones for data acquisition. This gives rise to an archive of multi modal data that have different acquisition geometry, radiometry, acquisition conditions and Ground Sampling Distance. This forces the data processing and analysis team to be equipped with methods that can readily handle such versatile data. With the dominance of artificial intelligence in earth observation, this thesis focuses on developing a Convolutional Neural Network (CNN) model that provides a robust performance for detecting exposed buildings when subjected to optical data from different kinds of sensors and platforms. This thesis starts with an approach of training a region-based network to obtain a baseline model, which then is improved gradually by using advanced techniques like data augmentation and fine tuning. A comprehensive performance evaluation is carried out under consideration of different training-testing scenarios. Furthermore, the influence of tile-size on the detection performance is tested. The resultant model after improvements is tested on an independent validation dataset acquired during rapid mapping activation of the Centre for satellite-based crisis information (ZKI) during the floods in Germany, July 2021. Contrary to intuition, the model owning the implementation of augmentation technique on the xView global dataset, shows the best performance for transferability. Due to resource limitation, the pipeline has been trained with a small sliver of the available dataset. The model weights obtained by retraining on the entire dataset with much powerful machines will provide new benchmarks for transferability models in object detection. By combining the resultant exposure with hazard information, we can get a first insight into which areas are likely to be affected in the event of a catastrophe. The importance of this work is that it provides an up-to-date picture of the building stock compared to Open Street Map or cadastre data, at different phases of the disaster