Presentation at the ESA Φ-week 2018
During ESA Φ-week 2018, our team presented the outcomes of our research in the field of Earth Observation, with a particular focus on utilizing Deep Learning techniques for remote sensing image segmentation.
Remote sensing image segmentation holds significant importance in the realm of Earth Observation, prompting numerous methods to be explored and proposed. Our study specifically concentrates on Deep Learning network architectures for semantic image segmentation. While many existing solutions overlook image preprocessing and primarily emphasize network topologies and hyperparameters, our work introduces a comprehensive suite of tools tailored to Earth Observation data. These tools are designed to facilitate seamless experimentation and integration of various Deep Learning models, preprocessing techniques, and model ensemble methods.
To tackle large-scale image segmentation tasks, such as building footprint detection and road extraction, we leverage state-of-the-art machine learning models and incorporate them into our Earth Observation toolset. The motivation behind our research was driven by the Urban 3D and SpaceNet RoadDetector challenges, where the need for a reusable and extensible machine learning framework became evident. The Urban 3D challenge sought an algorithm capable of reliably and automatically labeling building footprints based on orthorectified color satellite imagery and 3D height data, while the SpaceNet RoadDetector challenge focused on automated methods for extracting routable road networks from high-resolution satellite imagery.
Throughout our study, we extended and evaluated prominent deep learning models, such as U-Net, Segnet, Xception, and DeepLabv3+, for Earth Observation tasks. We integrated powerful Machine Learning and Computer Vision tools, including TensorFlow, Keras, OpenCV, and SciKit-Learn, with modern Earth Observation tools like RasterIO (for raster and elevation model handling), Fiona, and Shapely (for vector data handling). In addition to assessing our tools and models using data from the aforementioned competitions, we also conducted evaluations against reference datasets like the ISPRS 2D Semantic Labeling - Vaihingen dataset.