This PhD course is part of the thematic path "Technologies, Techniques and Methodologies for Sustainable Development" of the PhD programme in Urban and Regional Development. GeoAI (Geospatial Artificial Intelligence) represents a transformative intersection of geospatial technology and artificial intelligence. This emerging field harnesses the power of AI algorithms, particularly machine learning and deep learning, to analyze and interpret vast amounts of geographic data. The importance of GeoAI lies in its ability to uncover patterns and insights that were previously unattainable, enabling more informed decision-making. By automating the analysis of complex geospatial data, GeoAI not only enhances the accuracy and efficiency of geographic information systems but also drives innovation in the ways we understand and interact with our world, ultimately contributing to more sustainable and resilient communities. The course will be structured with both theoretical and practical lectures, covering the following topics: • Machine and deep learning algorithms • Satellite or drone images processing with AI algorithms or neural networks • Semantic segmentation and point-cloud related tasks • Photogrammetric AI and NeRFs The proposed course will help students exploit both 2D and 3D data through AI algorithms, unlocking the potential of the AI techniques and analysing geographic/geospatial pattern, with at the same time a critical thinking on the obtained results. Expected outcomes: • knowledge of 2D and 3D geospatial data, from acquisition methods (mobile mapping systems, laser scanners, satellites), structure and management; • basic knowledge of machine learning methods and deep learning fundamentals, from training methods and optimization techniques, to specialized architectures for 2D and 3D data; • knowledge of machine learning techniques and methods for processing 2D and 3D geospatial data, with application for segmentation, classification and/or localization activities using QGIS, ArcGIS Pro, Google Earth Engine or Python. From the point of view of learning ability, students, at the end of the course, will be able to apply the knowledge acquired to further application cases, possibly using other software or programming languages, taking advantage of the design phases described during the course.

Basic knowledge of geospatial data and GIS systems

Course organization: The course includes approximately 6 hours of lectures and 9 hours of practical exercises. The topic of the exercise will be divided into 2D and 3D data applications. Course topics: • Geospatial data (3h of theory): - 2D data: digital images: structure, content, spectral bands and indexes; - 3D data: point clouds, geometric structure, acquisition methods, local geometric characteristics; - software and plugins for processing geospatial data with machine or deep learning algorithms. • Machine learning fundamentals (3h of theory): - Deep Neural network (DNN) concepts: perceptron, loss function, optimization - DNNs architecture, backpropagation, stochastic gradient descent - DNNs training: data preprocessing, weight initialization, hyperparameter selection - Results visualization - Convolutional architectures for classification, detection and segmentation (on 2D or 3D data) - GeoGPT • Geospatial AI applications (9h of exercises). The topics of the exercises will be decided together with the students at the beginning of the course and can range from: - Satellite image classification for land use purposes - Image Segmentation with SAM: Segment Anything Model - Point cloud semantic segmentation with user-defined Level of Detail - Automatic urban mapping from point clouds: 2D (footprint) and 3D (volume) data extraction - Time Series Analysis and Forecasting - Map creation and aggregation from textual data

On site

Written report presentation

P.D.2-2 - April

The course will mainly use open software. Students must have a personal computer to perform the practical exercises.