The adequate acquisition and analysis of a scene are of great interest for photogrammetry, remote sensing, computer vision, computer graphics and robotics. In this regard, 3D data have become a standard data source for numerous applications. The capability to interpret such 3D data, in turn, allows dealing with a variety of high-level tasks such as scene modelling, autonomous exploration of unknown scenes or autonomous navigation without collisions. To obtain such a capability, many applications rely on a semantic interpretation of the acquired data. In this course, we will first provide an overview of different concepts to obtain 3D data either directly by measurements or indirectly from acquired imagery. Subsequently, we focus on the semantic interpretation of 3D data. Thereby, we will present the classic approach relying on the use of handcrafted features and the state-of-the-art end-to-end learning approach relying on deep learning techniques.

The course introduces recent classification schemes with the goal to produce and update 2D topographic databases. The inclusion of the spatial descriptors such as geometry and shape are important to characterize topographic objects in orthoimages. This course will present some of the challenges in mapping from high- resolution orthoimages and highlights the necessity of characterizing the orthoimages in the spatial domain. The solution to the above challenges will be provided by the efficient tool called morphological attribute profiles. They are multi-scale attributes and are constructed by a hierarchical representation of the images, thus enabling object-based image analysis. These characterizations are classified using well-established machine learning methods and different data sources (either raw or derived features). Different approaches to assess the thematic and geometric accuracy of map data will be discussed, and lastly the cartographic enhancement of the classification maps at different levels of quality will be presented. Solutions to the tasks are given by means of detailed course material including open source programs.

Deep Learning has led to significant breakthroughs in various fields including natural language processing and computer vision. Remote sensing also benefits from such methodological advances and deep networks currently achieve state-of-the-art results in many automatic tasks, such as object detection, semantic segmentation (e.g. for land cover mapping), change detection, etc. The goal of this course is to introduce deep learning, review the main architectures relevant for cartography, photogrammetry and other EuroSDR related fields, as well as to train the participants with available software and codes. The audience will be offered to know about and practice with recent architectures proposed by the lecturers.

It is complementary to the course “Topographic Maps through Description and Classification of Remotely Sensed Imagery and Cartographic Enhancement” that focuses on the traditional approach to automated classification (i.e. feature extraction and supervised classification) while deep learning brings a paradigm change by learning both the features and the classifier, at the possible cost of higher labeled datasets and higher computational resources.

This introductory course aims to provide a comprehensive overview on the state-of- the art in Open SDI and its key components, and introduce participants to the underlying principles of Open SDI, and let them experience hands-on what it means to establish and maintain an Open SDI. A number of topics will be tackled: spatial data infrastructures as well open data principles, key standards, architectures, (network) services, relevant EU-regulations and policies, governance strategies, and key institutions. At the end of the course, participants are: informed about Open SDI strategies around the world, aware of the main strengths, weaknesses, opportunities and threats of Open SDI, familiar with the latest technological developments, capable to facilitate the opening of open data using latest developed tools, and able to evaluate Open SDIs.