Conference Agenda

This project aims to advance the application of artificial intelligence (AI) and Earth observation (EO) technologies to address critical urban and environmental challenges. As cities face increasing pressures from rapid urbanization and climate change, the need for high-resolution, timely, and actionable geospatial intelligence has become critical. This project brings together complementary expertise in AI, remote sensing, urban analytics, and environmental resilience to develop novel methods for understanding and managing complex urban-ecological systems.

To tackle the challenge of producing fine-scale thermal data, a conditional diffusion probabilistic model (DDPM) was developed to generate daily 30-meter resolution Land Surface Temperature (LST) maps. By fusing high-temporal-resolution VIIRS data with high-spatial-resolution SDGSat-1 imagery, and leveraging a U-Net-based transformer-enhanced architecture, this model significantly outperformed traditional interpolation methods. The resulting LST products provide a scalable solution for urban heat island (UHI) monitoring and climate adaptation planning in data-sparse environments (Brune et al., 2025).

Urban heat risk was further examined in Karachi using SDGSat-1 thermal infrared imagery in combination with local climate Zone (LCZ) classifications and socioeconomic vulnerability indicators. The study applied Crichton’s Risk Triangle framework to produce a Heat Health Risk Index (HHRI), revealing that the highest risk was concentrated in LCZs with high population density and informal settlements. The results support tailored urban planning responses in heat-vulnerable areas and contribute directly to SDGs 11, 13, and 10 (Baqa et al., 2025).

A deep learning-based approach was also developed to monitor 3D structural changes in urban areas using synthetic aperture radar (SAR) data. A CycleGAN-based image translation framework was applied to convert Sentinel-1 and COSMO-SkyMed SAR images into a common representation, enabling high-frequency monitoring of building construction and demolition. Applied over Shanghai, the model significantly improved spatial consistency and signal alignment across sensors. This approach was further extended to support 2D and 3D urban structure extraction in Shenzhen, providing a strong foundation for SAR-based urban monitoring and resilience applications (Memar et al., 2025).

Complementing this, a novel generative AI model was introduced for building damage assessment, focusing on the transferability of high-resolution satellite imagery models to lower-resolution datasets. The diffusion-based model is trained on the xView2 Wildfire Building Damage Benchmark, a dataset specifically designed for wildfire-induced building damage detection. The model is further evaluated on real-world wildfire incidents in Lahaina, Hawaii, and Athens, Greece, demonstrating its effectiveness in damage localization across varying spatial resolutions (Shibli et al., 2025).

To support urban infrastructure monitoring, the Poly-BRBLE framework was proposed for individual building localization and extraction. Combining boundary refinement, deformable convolution feature fusion, and cross-layer regularization, the model achieved state-of-the-art accuracy on multiple building datasets including SH, WHU, and Inria. It is particularly suited to complex urban environments and enables scalable footprint extraction for resilience assessment (Tan, 2025).

For urban land use monitoring, developed to detect functional zone changes. By fusing satellite remote sensing and street view imagery using attention mechanisms, the model provided high-precision classification of commercial, residential, and public service zones. Tested in Nanjing, MSTFM outperformed traditional CNN-based approaches in spatial detail and temporal coherence (Yang et al., 2025).

In arid urban regions, ecological sensitivity to anthropogenic activity was analyzed using a combination of Morphological Spatial Pattern Analysis (MSPA), InVEST habitat quality modeling, and omnidirectional connectivity assessment. Structural equation modeling and Geodetector analyses revealed that cultivated and bare lands are the primary drivers of ecological fragmentation, with habitat quality playing a crucial mediating role. Even modest improvements in habitat conditions yielded measurable gains in ecological connectivity (Mu et al., 2025).

To evaluate long-term ecological vulnerability, a Dryland Ecological Vulnerability Index (DEVI) was developed for the Hohhot-Baotou-Ordos-Yulin urban agglomeration. Incorporating six remote sensing-based indicators, DEVI mapped 40 years of ecological change. Using XGBoost and SHAP for model interpretation, the study identified high-risk areas, thresholds of vulnerability, and zones where ecological restoration has been effective (Li et al., under review).

A comprehensive multi-hazard risk framework, termed “risk source–risk exposure–mitigation force,” was implemented for urban disaster simulation. This framework integrated satellite, street view, and video data with deep learning and physical models (e.g., CFD) to assess vulnerability to wind, seismic, and fire hazards. Applications included city-scale wind damage analysis through multi-view image fusion and a spatiotemporal casualty assessment method for earthquake-induced debris and fire (Gu et al., 2025; Xu et al., 2025).

By combining diverse datasets, advanced deep learning techniques, and validated case studies across multiple regions, this project delivers innovative, scalable solutions for urban heat mitigation, land use management, ecological conservation, and disaster resilience. The methodologies developed offer practical tools for urban planners, ecologists, and risk managers committed to building sustainable and adaptive cities.

Brune et al. (2025). Generating Daily High-Resolution Urban Land Surface Temperature Maps using a Conditional Diffusion Model with SDGSat-1 and VIIRS Data. (Abstract).

Baqa et al. (2025). Investigating Heat-Related Health Risks Related to Local Climate Zones Using SDGSAT-1 High-Resolution Thermal Infrared Imagery in an Arid Megacity. International Journal of Applied Earth Observation and Geoinformation, 136, 104334.

Memar et al. (2025). Building Height Estimation from COSMO-SkyMed Imagery through Deep Learning Methods. Proceedings of JURSE 2025.

Shibli et al. (2025). Very High- to High-Resolution Imagery Transferability for Building Damage Detection Using Generative AI. Proceedings of JURSE 2025.

Tan (2025). Poly-BRBLE: A Boundary Refinement-Based Individual Building Localization and Extraction Model Combined with Regularization. IEEE Transactions on Geoscience and Remote Sensing, Early Access.

Yang et al. (Under Review). Urban Functional Zone Change Detection via Multi-Source Temporal Fusion of Street View and Remote Sensing Imagery. IEEE Transactions on Geoscience and Remote Sensing.

Mu et al. (2025). Quantifying the Anthropogenic Sensitivity of Ecological Patterns in Arid Urban Agglomeration. Applied Geography, 178, 103595.

Li et al. (Under Review). Dynamic Monitoring of Fine-Grained Ecological Vulnerability in Dryland Urban Agglomerations Integrating Remote Sensing Index and Explainable Machine Learning. GIScience & Remote Sensing.

Gu et al. (2025). Multi‐View Street View Image Fusion for City‐Scale Assessment of Wind Damage to Building Clusters. Computer‐Aided Civil and Infrastructure Engineering, 40(2), 198–214.

Xu et al. (2025). A Spatiotemporal Casualty Assessment Method Caused by Earthquake Falling Debris Considering Human Emergency Behaviors. International Journal of Disaster Risk Reduction, 105206.

In Dragon 6, urban security research primarily focuses on utilizing earth observation to address two distinct themes: urban thermal risk, and geological disasters related to land subsidence.

(a) Urban Heat Risk

The complex interplay between heat dynamics and urban form and urban fabric was examined in view of recognizing relationships that need to be considered as far as the scientific basis of urban heat risk is concerned. Results demonstrate the importance of both urban form and urban fabric to the state of the urban thermal environment and consequently to urban heat risk. Furthermore, a multiparameter architecture, combining downscaling methods and machine learning, has been shaped to unravel the spatial and temporal dimension of EO data, quantify the urban heat dynamics also with respect to urbanization trends and finally develop a classification scheme of urban heat risk on the basis of urban typologies. The architecture addresses the exploitation of such parameters as Land Surface Temperature (LST), vegetation, proportion of ground surface with building cover, proportion of impervious surface (paved, rock), proportion of pervious surface (bare soil, vegetation, water), distance from green area/park, etc. as well as derived indices such as LST Anomaly, NDVI Anomaly, combined pervious surface/NDVI anomaly, compound heat and drought events, etc. Special attention was given to the assessment of the impact of urban greenery on urban heat dynamics.

(b) Urban land subsidence:

Addressing the complex evolution and mechanism of land subsidence in Beijing-Tianjin-Hebei urban area under the combined impact of climate change and human activities, a comprehensive interdisciplinary approach integrating remote sensing, geophysical exploration, hydrogeology, and artificial intelligence has been employed. Advanced technologies such as InSAR, GRACE, peridynamics, and physical information neural networks (PINNs) have been integrated to analyze the differential evolution patterns of urban groundwater driven by the natural-human dualistic water cycle and the response of land subsidence.

(1) The ERA-5 reanalysis dataset, water resources bulletin dataset and GLDAS hydrological model were combined to obtain the temporal characteristics of key water cycle elements including precipitation, runoff, evaporation, and groundwater extraction in recent twenty years. What’s more, During the process, Theil-Sen trend analysis method, wavelet analysis method variational modal decomposition, time-lag correlation analysis, and random forest methods were used to analyze the response of groundwater storage change to precipitation and groundwater extraction, and it’s seasonal and long-term trends characteristics in the North China Plain.

(2) PINNs (Physics-Informed Neural Networks) groundwater level simulation model with physical mechanisms was constructed by integrating long-term observation data from the groundwater level monitoring network, hydrogeological data, and precipitation data. Under the prior constraint of simulating groundwater level using the MLP (Multilayer Perceptron) model, the groundwater field partial differential equations were incorporated into the network. The weights of different loss function components were optimized determined. The model was successfully applied in simulating the karst and porous medium groundwater level. This study provides a powerful and reference-worthy approach for exploring the response to groundwater level.

(3) The ordinary state-based peridynamics, employing equilibrium equations based on integral formulations, was modified to simulate the evolution from continuous land subsidence to discontinuous fissure due to groundwater over-pumping. The approach is used to investigate fissure, associated to pre-existing faults and excessive exploitation of aquifer systems, occurrences in the subsiding North China Plain around Beijing. The relations between the typical hydrogeological setting and the fissures were quantified. Result highlights that wider and deeper fissures develop with increasing fault dip, aquifer thickness, piezometric decline, and stiffness of surface deposits.