Automatic optical satellite image orthorectification and registration with advanced deep learning methods

Earth observation has become an important source of geospatial data. Due to the large amount of new and archived satellite data of high- and very high-resolution (HR and VHR) the automation of its processing is becoming imperative, as only automatic processing can provide large amounts of end products. To ensure reliable interpretability and maintain the high quality of the derived data, satellite data must be geometrically corrected prior to their use. The most reliable and accurate methods for geometric corrections are orthorectification and image registration. The accuracy and automation of the geometric correction has a critical influence on all further processes. For VHR satellite data, existing automatic pre-processing methods do not yet achieve technical and operational maturity required for accurate positioning of processed data in reference coordinate systems. Many automatic methods achieve good relative co-registration accuracy among overlapping images of the same type. However, the processing required to achieve good absolute accuracy is usually semi-automatic and time-consuming. The proposed project aims to contribute to this challenge by developing a robust, accurate and automated process for geometric corrections of various HR and VHR optical sensors using state-of-the-art methods. All implemented processes will be interconnected, automated and integrated into the stable and automatic STORM processing chain developed by the research team over the past years. The developed procedure will include all processing steps required for either orthorectification or image registration. The implementation will be generic, automated and enhanced with modern deep learning techniques. The research will be conducted in five interconnected work packages (WP1 to WP5), and a separate work package (WP6) dealing with project management tasks and dissemination of results. In WP1, we will establish access to source satellite and ancillary data and prepare them for the development and testing of our methods. WP2 will deal with the automatic extraction of ground control points (GCPs) through a road matching procedure. Deep learning techniques will be used to match the roads detected on a satellite image with roads from the reference road network. In WP3, we will develop different geometric models using the extracted GCPs to generate orthoimages, applying and comparing different orthorectification methods and using different digital elevation models. WP4 is dedicated to the development of an image registration procedure using deep learning. This procedure will serve as a complement to the orthorectification procedure when for example no road network is available. In WP5, we will integrate the procedures developed in the previous WPs into a common workflow; the accuracy of the results obtained will be evaluated and the whole procedure validated. The described workflow offers several innovations over previous research. The most important novelties are: • fully automated real-time retrieval and preparation of ancillary data from various databases, • the integration of deep learning methods in the GCP extraction phase, • implementation of image registration that can complement orthorectification, and • automatic generation of very accurate spatial data from different sensors using a generalized approach. If the proposed project is successfully realized, it will contribute many innovations to the satellite image processing and remote sensing in general. Moreover, it has the potential to contribute novel approaches in other scientific fields specialized in image processing. The prototype implementation of the developed procedures will provide a firm basis for transfer of the mastered innovations to current operational systems for obtaining or updating spatial databases already in use by public or private users. The proposed project is also harmonized with the needs and requirements of the European and international space programmes.