Conference Agenda

Roofing material classification is critical for urban sustainability, energy efficiency, public health and environmental protection, and regulatory compliance. Despite the need for scalable solutions, existing approaches are hindered by reliance on costly, specialised, multispectral/hyperspectral imagery, and spatial biases, and they overlook the potential of deep learning and multimodal data fusion. This paper addresses these research gaps by introducing RoofSense, a multimodal semantic segmentation dataset for roofing material classification in diverse urban contexts, leveraging~qty{8}{cm} aerial~acs{rgb} imagery and airborne lidar data. Representing eight diverse classes and spanning more than~qty{138}{ha} and 480 buildings across five Dutch cities, RoofSense is the largest publicly available dataset of its kind. By fusing spectral and geometric information at the pixel level and employing a novel weighting scheme to address class imbalance, RoofSense was used to achieve competitive classification and segmentation performance with a tuned, off-the-shelf model based on ResNet-18-D and DeepLabv3+. Lidar-derived features were added to improve performance in difficult classes and materials commonly used in pitched roofs, but results were ultimately sensitive to material and building context, clutter, and modality alignment. The implementation is publicly accessible at url{https://github.com/ANONYMOUS}.

This paper introduces BldgWeaver, a novel adaptive generative model for creating 3D building digital cousin (BDC) models using pre-trained Transformer architecture. Unlike traditional approaches that require complete 3D reconstruction with extensive visual data, BldgWeaver approximates building geometries using artificial intelligence-generated content to address data deficiencies in urban digital twin development. The proposed method employs a token-based approach to convert triangle mesh coordinates into discrete tokens for auto-regressive prediction, incorporating parallel conditional controls and an optimized footprint-masked training strategy. Experiments conducted on the PLATEAU dataset demonstrate our model’s capability to generate Level of Detail 2 (LoD2) building models with diverse roof structures, achieving an average 49% improvement in geometric proximity compared to basic LoD1 representations. The proposed model effectively addresses challenges in wide-range urban mapping by reducing data dependencies while maintaining satisfactory architectural fidelity.

Semantically rich 3D city models play a vital role in a variety of applications, such as urban planning. Enhancing these models with currently unavailable attributes, such as roof material types, can unlock new opportunities to tackle pressing challenges, including climate change mitigation and sustainable urban development. In this work, we present an end-to-end pipeline for the automatic detection of roof materials to semantically enrich 3D city models. To support this, a comprehensive training dataset was prepared by labeling roofs across Germany using OpenStreetMap (OSM) attributes and high-resolution orthophotos. Our detection results enabled the automatic augmentation of CityGML-based 3D models, filling in missing roof material information. This enrichment supports advanced applications, such as assessing roof suitability for blue-green infrastructure or simulating urban heat island mitigation strategies. We validated the feasibility of our approach with real-world data and applied the method to a district in the city of Bremen, Germany. The paper also includes a detailed discussion of the learning process quality, the integration, and the visualization of the enriched 3D city model. The code used in this study is available at: [github].

Recent advances in Gaussian Splatting (GS) have demonstrated its effectiveness in photo-realistic rendering and 3D reconstruction. Among these, 2D Gaussian Splatting (2DGS) is particularly suitable for surface reconstruction due to its flattened Gaussian representation and integrated normal regularization. However, its performance often degrades in large-scale and complex urban scenes with frequent occlusions, leading to incomplete building reconstructions. We propose GS4Buildings, a novel prior-guided Gaussian Splatting method leveraging the ubiquity of semantic 3D building models for robust and scalable building surface reconstruction. Instead of relying on traditional Structure-from-Motion (SfM) pipelines, GS4Buildings initializes Gaussians directly from low-level Level of Detail (LoD)2 semantic 3D building models. Moreover, we generate prior depth and normal maps from the planar building geometry and incorporate them into the optimization process, providing strong geometric guidance for surface consistency and structural accuracy. We also introduce an optional building-focused mode that limits reconstruction to building regions, achieving a 71.8% reduction in Gaussian primitives and enabling a more efficient and compact representation. Experiments on urban datasets demonstrate that GS4Buildings improves reconstruction completeness by 20.5% and geometric accuracy by 32.8%. These results highlight the potential of semantic building model integration to advance GS-based reconstruction toward real-world urban applications such as smart cities and digital twins. Our project is available: [anonymized for the submission].