Conference Agenda

The EO-AI4ResilientCities project integrates Earth Observation (EO), artificial intelligence (AI), geospatial modelling, and urban analytics to strengthen urban resilience under growing climate and socio-environmental pressures. By combining multi-source satellite observations, deep learning, generative AI, foundation models, and geospatial analysis, the project develops scalable methods for disaster damage assessment, urban climate monitoring, vulnerability analysis, and infrastructure risk evaluation, contributing to resilient and sustainable urban planning across Europe and China.
One major research component of the project focuses on urban 3D mapping using multi-source EO data and advanced deep learning models. To support large-scale and accurate building height estimation, the project developed T-SwinUNet that integrates Sentinel-1 SAR and Sentinel-2 multispectral time series for 10 m building height mapping and demonstrated strong generalization capability across multiple European cities (Yadav et al., 2025). Complementing this work, a Tri-Modal Height Estimation Network (TMHEN) was developed by integrating SDGSAT-1 nighttime light imagery with Sentinel-1 SAR and Sentinel-2 MSI data for building height estimation across major Chinese urban agglomerations. The inclusion of high-resolution nighttime light data substantially reduced height underestimation, further demonstrating the value of combining multi-source EO data for accurate urban 3D mapping (Z. Li et al., 2026).
Another component focuses on exploiting foundation models for climate hazard analysis in urban areas. Using the SMARTIES foundation model (Sumbul et al., 2025), an approach was developed to extract information related to urban hazards and support resilience to climate change. The Deep Learning methodology exploits the SMARTIES embeddings to characterize the temporal changes in heterogeneous sequences of data, which are all projected into a common latent space. Currently tested for floods, the approach can also be jointly exploited with 2D and 3D buildings information extracted from EO imagery, enabling more comprehensive urban risk characterization.
Urban thermal environment monitoring forms another major theme of the project. A generative AI framework was developed for producing daily 30 m land surface temperature (LST) products by combining daily VIIRS observations with high-resolution SDGSAT-1 thermal imagery. Building on diffusion-based super-resolution approaches, the framework expanded the training dataset with additional cloud-free SDGSAT-1 acquisitions, improving reconstruction accuracy across multiple cities. Ongoing work further compares diffusion models with conditional flow matching and parameter-efficient fine-tuning strategies to improve model efficiency and performance, particularly for cities with limited training data (Brune et al., 2026). Complementary studies investigated relationships between urban morphology and land surface temperature in major South Asian cities using SDGSAT-1 thermal observations. Results showed that building density and building volume are dominant warming drivers, while vegetation volume provides significant cooling effects, with substantial inter-city variability indicating the need for locally adapted heat mitigation strategies.
Another component of the project addresses disaster damage assessment and urban risk analysis under natural hazards. A disaster-adaptive deep learning framework was developed for building damage assessment using very-high-resolution satellite imagery. By incorporating disaster-type information into the learning process, the approach improves the generalization capability of deep learning models across previously unseen disaster events and demonstrates that data-centric strategies can provide more reliable and scalable solutions for operational disaster response, infrastructure recovery, and resilience planning (Hafner et al., 2025). In addition, long-term surface deformation and associated building damage risks were analyzed using multi-temporal InSAR observations combined with building footprint data, revealing substantial spatial variability in subsidence dynamics and infrastructure vulnerability driven by groundwater extraction, geological conditions, and compressible soil layers (Wang et al., 2026).
The project also advances urban resilience analysis through economic vulnerability assessment, infrastructure monitoring, and change detection. For economic vulnerability assessment, nighttime light observations were integrated with explainable AI and causal discovery algorithms to develop a Resource-based Economic Vulnerability Index (REVI) for analyzing factors influencing urban resilience. The results showed an overall decline in vulnerability between 2000 and 2023, while highlighting persistent inequalities in resource-dependent counties dominated by mining and industrial activities (C. Li et al., 2026). Additional work combined remote sensing imagery, street-view data, GIS information, simulation and deep learning to investigate wind damage assessment, earthquake casualty estimation considering human emergency behavior, and three-dimensional fire spread simulation in dense urban environments (Gu et al., 2025; Xu et al., 2025a; Tian et al., 2025; Xu et al., 2025b). In addition, a Cross-Resolution Change Detection Network (CRCD-Net) was developed to address spatial misalignment, heterogeneous resolutions, and semantic inconsistencies in multi-temporal EO imagery, demonstrating strong performance for complex urban applications. These studies enhance support for urban early warning systems and resilience-oriented disaster management.
Overall, the EO-AI4ResilientCities project demonstrates the strong potential of integrating multi-source Earth observation, AI, and urban analytics to support resilient and sustainable cities. The midterm results highlight significant advances in disaster assessment, urban heat monitoring, vulnerability analysis, and infrastructure risk evaluation, while providing practical EO-based methodologies and decision support tools for urban resilience, climate adaptation and disaster risk reduction.

In Dragon 6, urban security research focuses on leveraging Earth Observation (EO), GIS, and modelling techniques to address critical urban challenges associated with urban thermal risk and land subsidence. While the former concerns the amplification of heat exposure driven by urban form and function, the latter reflects long-term ground deformation processes linked to hydrogeological dynamics and human activities. Together, these themes highlight the need for integrated EO-based frameworks capable of capturing processes that increasingly affect the resilience and sustainability of urban systems under rapid urbanization and climate change.

(a) Urban Heat Risk

Urban climate change adaptation is addressed through an EO driven framework aimed at mitigating thermal risk associated with extreme heat and heatwaves. CMIP6-based, statistically downscaled projections of temperature and heatwave metrics are incorporated to capture future thermal risk dynamics and support scenario-based assessments. The approach builds on the linkage between the urban thermal environment and the 3Us, namely urban functions, urban form, and urban fabric, to better understand and quantify intra-urban heat dynamics. A multiparameter architecture is developed, combining advanced downscaling techniques and machine learning to extract spatial and temporal patterns of urban heat from multi-source EO data, including Sentinel-2, Sentinel-3, and SDGSAT-1. Key variables include Land Surface Temperature (LST), vegetation indices, and proximity to green spaces, along with derived indicators such as LST anomalies and compound heat-drought events. A classification scheme based on urban typologies enables the identification of thermally burdened areas.

The cooling potential of urban greenery is assessed through analysis of the Park Cool Island effect, considering park size, shape, and response to heatwaves and droughts. In parallel, targeted mitigation strategies are evaluated using advanced urban climate simulation models, including the microscale CFD model ENVI-met, to quantify the impact of urban form, materials, and vegetation on thermal conditions and ventilation, supporting optimized, climate-resilient urban planning.

(b) Urban land subsidence

Focusing on the complex evolution and formation mechanism of urban land subsidence in Beijing and the Beijing-Tianjin-Hebei region under the coupled effects of climate change and human activities, this study integrates multi-disciplinary theories of remote sensing, geophysical exploration, hydrogeology and artificial intelligence, and adopts advanced technologies including InSAR, GRACE, peridynamics and physics-informed neural networks (PINN), to clarify the differentiated evolution of urban groundwater storage and the response mechanism of land subsidence driven by the natural-artificial dual water cycle. A "prior-posterior-scenario" physically constrained multimodal data fusion and intelligent decomposition framework is established to realize the refined separation, reconstruction and verification of multi-source remote sensing signals of land subsidence and groundwater in the Beijing-Tianjin-Hebei Plain (Young Scientists’ Research Achievement). A GRACE downscaling algorithm coupled with InSAR groundwater inversion and weight optimization is proposed to generate 0.05° high-resolution groundwater storage data, and reveal the spatiotemporal heterogeneity and synergistic response between groundwater dynamics and surface deformation. Constrained by the dual water cycle framework, the PINN model is constructed to identify the coupling interaction between groundwater level variation and land subsidence. Furthermore, a water balance constrained liquid time-constant neural network model is developed to predict groundwater dynamics in karst aquifers; combined with in-situ groundwater observations and time series analysis, the impact of Yongding River ecological water replenishment on urban security is evaluated; a relative sea level rise inundation model is established to predict flooding characteristics under diverse land subsidence and CMIP6 climate change scenarios; a land subsidence-ground fissure evolution model based on peridynamics is coupled with the finite element method for efficient numerical simulation, which provides theoretical basis and technical support for regional land subsidence prevention, risk assessment and disaster early warning.