Conference Agenda

The growing availability of open-access geographic data and high-resolution satellite imagery, coupled with advances in machine learning, is transforming the field of urban morphology. These developments enable more scalable, quantitative, and data-informed approaches to analyzing urban form. However, conventional methods often fall short when applied to complex, large-scale, and heterogeneous spatial contexts. Addressing this gap, the present study proposes a hybrid workflow that integrates Earth Observation (EO) imagery, image segmentation with AI, and morphometric analysis. By leveraging machine learning techniques, the method enables the detection, classification, measurement, and comparison of urban block patterns across diverse morphological settings, offering a replicable framework for large-scale urban form analysis.

The proposed approach adapts a U-Net convolutional neural network for urban block segmentation, using PlanetScope 3-meter resolution imagery as input and OpenStreetMap (OSM) data to generate training masks. Initially trained on the city of Barcelona, the model is tested across five additional cities, Turin, Paris, Copenhagen, New York, and Beijing, each representing distinct morphological patterns of urban blocks. This cross-contextual application allows for evaluating the model’s generalizability and adaptability. Following segmentation, the morphological characteristics of the detected urban blocks are analyzed using the Momepy library, enabling typological classification and quantitative comparison with manually curated reference data. To assess the robustness of the workflow, a four-step validation strategy is employed. First, form-based multi-scalar validation checks spatial consistency across macro, meso, and micro levels. Second, numerical–morphometric validation compares geometric properties of AI- and OSM-derived blocks using a range of morphometric indicators. Third, cluster comparison validation analyzes the structural similarity between datasets by clustering their extracted numerical features. Lastly, cross-case validation evaluates the workflow’s performance across diverse urban contexts, highlighting its robustness and limitations in varying spatial settings.

The results show that the adapted U-Net model performs with high spatial accuracy in detecting urban blocks in cities characterized by regular structures and delivers acceptable outcomes in more complex or irregularly developed urban environments. Subsequent morphometric analysis confirms that the AI-generated blocks successfully reproduce key typological features. Clustering based on geometric attributes further indicates a strong alignment between the AI-derived outputs and those obtained from reference datasets. Cross-case findings highlight the method’s ability to generalize across varied morphological settings while supporting scalable and transferable analysis, especially valuable for comparative urban studies in data-scarce regions.

By unifying image segmentation and morphological analysis into a single, replicable workflow, this study helps bridge the gap between geospatial AI and urban morphology. It advances the capacity to examine urban form objectively and systematically across diverse global contexts.

This paper describes the National Building Database (NBD), a digital twin of the building stock of Great Britain (England, Wales and Scotland). All buildings are covered, both domestic and non-domestic. The NBD has just been delivered to the Department of Energy Security and Net Zero (DESNZ) of the British Government. The paper explains how the Database has been constructed; lists the purposes for which it will be used; and reveals new findings about the stock uncovered by the work.

The NBD has been produced using the 3DStock modelling method developed over the last twenty years at the Energy Institute, University College London. It is constructed on a digital map base. The three-dimensional envelopes of buildings are modelled from LiDAR (laser) measurements made by overflying aircraft. Floor areas come from taxation records or are inferred from the 3D envelopes. Other data are incorporated on materials of construction, building age, heating systems and energy efficiency measures installed. These come from commercial databases and Energy Performance Certificates. Conservation areas and buildings listed as of special historical interest are recorded. Actual annual electricity and gas consumption from meters are matched to premises and buildings.

The Department of Energy plans to use the NBD for research on energy use and carbon emissions, and to inform government strategy and policy in three areas: improvement of energy efficiency and reduction of costs; examination of practical approaches to low carbon retrofitting; and investigation of the potential for integrating low carbon and renewable energy technologies. The Department will make data available in anonymised form for other users. Because all buildings are located geographically on their sites, the NBD lends itself to analysis of district heating systems, shared heat pump loops, and the location of ground source heat pumps as well as building-based measures.

Production of the Database has allowed mapping of the spatial distribution of building uses, building forms, and densities of development. The paper will include a sample of maps. One major finding is the very great extent of mixed-use buildings, in which different non-domestic uses are found (e.g. offices above shops), or domestic combined with non-domestic (e.g. flats above shops). These have tended to escape attention in research and policy but present special problems for retrofit.

Urban land use mapping is a key application in city planning, infrastructure construction, and efficient management of resources. The process, traditionally, has been based on precise field surveys and remote sensing, which, as useful as they have proven, sometimes require huge capital outlay, time, and effort. Satellite imagery has made some of the impediments easier to solve because it has allowed large-scale land cover and indirect anthropogenic feature mapping. Nonetheless, data from satellites mainly record physical characteristics and do not have the level of detail necessary to portray dynamic socioeconomic activities and human uses of urban places.

In response to this void, the utilization of Geospatial Big Data (GBD) offers a revolutionary prospect for increased urban land use mapping. Social media data with geotags, especially, holds unparalleled promise for extracting real-time, spatially-delineated understanding of human activity patterns, thus offering an enhanced spatiotemporal concept of land use processes. Several recent research works have put forward frameworks for utilizing such data using temporal and content analysis, showcasing its ability to augment traditional land use databases.

Drawing upon these developments, our study investigates the incorporation of geotagged social media metadata to enhance the accuracy and interpretability of urban land use maps. We particularly use an exhaustive list of land use categories and subcategories developed by the Atal Mission for Rejuvenation and Urban Transformation (AMRUT), prepared by the Government of India. In order to further improve the coverage and richness of our dataset, we also include the corresponding synonyms and substitute phrases for these land use categories.

We have utilized natural language processing (NLP) models to examine the text within social media posts and prepare a dataset. This data corpus was used to identify keywords and activity patterns that are characteristic of certain land use types. Our initial analyses identify significant spatial clusters that reflect different land use categories. Additionally, through combining topic modeling with Geographic Information System (GIS) methodologies, we effectively group these activity patterns, providing insight into the spatial distribution and temporal dynamics of urban land use.

Our results show that data from social media, when integrated with machine learning and GIS models, greatly improves the level of detail and contextual depth of urban land use maps. This method offers planners and policymakers a cost-saving, real-time, and participatory tool to observe and control urban spaces better.

Urban areas are increasingly affected by the impacts of climate change, including rising temperatures, more frequent extreme weather events, and altered precipitation patterns. Addressing these challenges requires precise, up-to-date land cover data to support effective spatial planning and climate adaptation strategies. This study presents a novel deep learning approach for fine-grained urban land cover classification, utilizing high-resolution aerial imagery, and provides planners with actionable insights into surface composition and vegetation structure.

Our method classifies urban surfaces into eleven semantically and environmentally relevant categories defined in a participative process: three vegetation height classes (low, medium, high), three levels of surface sealing (fully sealed, partially sealed, unsealed), three levels of rooftops (standard, green, with photovoltaic (PV)), open space PV, and water bodies. The classification leverages a convolutional neural network (CNN) based on the U-Net architecture, trained using 20 cm resolution RGBi (red, green, blue, infrared) imagery and normalized digital surface models. This combination enables highly detailed semantic segmentation of urban environments.

To address class imbalance in training data, common in urban datasets, we apply targeted data augmentation techniques, ensuring adequate representation of minority classes. In addition, we introduce a multi-scale training approach, where separate CNN models are trained at resolutions of 20 cm, 50 cm, and 100 cm. This enables the model to leverage the specific advantages of each resolution: higher spatial detail for small features, such as rooftops, and improved consistency for larger elements, like open vegetation or agricultural fields.

Each resolution is validated independently and is then employed as a decision filter in a late-fusion strategy that combines the multi-scale outputs into a single high-accuracy land cover map. The final segmentation achieves an overall accuracy of over 90% in diverse test areas within the Frankfurt on the Main metropolitan region (Hesse), demonstrating both robustness and generalizability.

This work showcases the potential of AI-powered geospatial analysis to advance climate-sensitive urban planning. The resulting high-resolution land cover maps support a range of applications, enabling public authorities to monitor the implementation and effectiveness of climate adaptation strategies. By capturing detailed heterogeneity in urban surface materials and vegetation structure, the approach allows planners to identify spatial vulnerabilities and target interventions, such as cooling corridors, greening strategies, or reflective surfaces, with unprecedented precision.

Ultimately, our research bridges the gap between cutting-edge remote sensing technologies and practical planning needs. It provides municipalities and regional authorities with a scalable, automated tool to support evidence-based, adaptive, and resilient spatial development strategies in the face of accelerating climate change.