GeoAI for Disaster Response:
From Research to Practice

Authors: Alireza Niknam Khajepasha, Ayyoob Sharifi, Bisser Raytchev, Fuko Nakai

Abstract:In recent years, due to severe climate change, water-related disasters have increased significantly. Water-related disasters can be divided into two categories: rapid-onset, such as flooding (Hwang, Kim et al. 2024), hurricanes (Nelson, Poleacovschi et al.), and snowmelt (Pinzner, Sturm et al. 2024). And slow-onset events such as droughts(Ansari, Liaqat et al. 2024) and water insufficiency(Pandey, Mohapatra et al. 2025). Either of these events poses a risk to human health and the environment, requiring short- or long-term actions to prevent damage. However, records show that flooding is among the most destructive and recurring water-related disasters. Between 1993 and 2024, flood events have affected more than 2.3 billion people(Adil 2025). Also, flood events between 2010 and 2014 caused damage totaling 400 billion USD. Urban areas, due to their high population density, are more vulnerable to flooding. Factors such as unplanned urban sprawl, rapid urbanization, and the destruction of natural floodplains (deforestation) reduce the absorption capacity of cities during floods. Moreover, insufficient infrastructure and its unequal distribution led to vulnerability among marginalized communities(Andreasen, Agergaard et al. 2022). These problems have made the flood one of the most destructive natural disasters, not just among water-related disasters but also among all types of catastrophes. Significant progress has been made in predicting flood risks, primarily driven by advancements in hydrological modeling, remote sensing, and Geographic Information Systems. However, these tools often treat urban flooding as a static phenomenon and often miss the dynamics of floods, which could result in inaccurate risk assessments. With recent advances in computer science, Artificial Intelligence has been increasingly adopted across various disciplines to enhance predictive capabilities, with flood risk management emerging as one of its most prominent application areas. Accurate and timely flood prediction is essential for effective mitigation planning and for minimizing damage to human life, infrastructure, and property. Despite extensive research on AI-based flood prediction, limited attention has been paid to the comparative suitability of different AI models under varying road network characteristics and real-time operational constraints. This study presents a systematic comparison of three widely adopted AI-based models across three regions in Japan with different road configurations to determine which performs best at predicting both floods and road damage simultaneously. This multi-regional approach enables the identification of (1) the most suitable model for flood risk prediction when road systems are explicitly incorporated, and (2) the most effective model for real-time flood prediction applications.

Authors: Darius Couchard , Charles Hamesse, Rob Haelterman, Sam Leroux

Abstract: The recent escalation in global conflicts has led to a proportional increase in the deployment of landmines. Although these devices can be rapidly dispersed across expansive territories, the subsequent demining process is significantly more resource-intensive and time-consuming. For instance, the Republic of Croatia required 31 years to clear 1,174 square kilometers of land contaminated during the War of Independence, officially achieving mine-free status in 2026 (1). In total, 107’000 mines and 407’000 unexploded ordnances have been removed. To enhance the efficiency and automation of landmine detection, recent research has been focusing on the analysis of spectral signatures (2) (3). Given that landmines are frequently partially buried or covered by overgrown vegetation, detection based on spatial features is often unreliable. However, multispectral imaging enables the differentiation of the chemical properties of a mine's surface from its background environment, even when the target is represented by a limited number of pixels (4). The TaMaCare project focuses on the development of a prototype designed for landmine detection via an integrated suite of electro-optical and thermal sensors. The system's payload spans the visible (VIS), short-wave infrared (SWIR), and thermal spectra, utilizing both panchromatic and multispectral modalities integrated onto a unified sensing platform. Data acquisition tasks are managed by an NVIDIA Jetson Orin Nano computational module. The sensor platform includes an AlliedVision 1800 U-130 panchromatic sensor for broad VIS-SWIR coverage, a WEOM Thermal Core for thermographic analysis, and a 16-band MQ013RG-E2 multispectral camera focused on the near-infrared (NIR) spectrum. Dual Ximea MU051CG-SY RGB cameras are employed for binocular depth estimation. An external signal generator is also included to perform hardware-level triggering, achieving near-perfect acquisition synchronization for the RGB cameras. Finally, an OAK-1 W camera is integrated to provide wide-field reference imagery and to facilitate auxiliary sensor processing. This specific model was selected for its embedded image processing pipeline, enabling effortless acquisition of high-quality imagery without the need of parameter tuning. Beyond the challenges of hardware integration and synchronous acquisition, significant complexity arises in the post-processing of heterogeneous sensor data. Because the sensors on the platform are not co-axial, the resulting imagery exhibits distinct perspective disparities. To generate a coherent spatiotemporal datacube suitable for target detection, the images must be spatially aligned via image registration. As the project aims at detecting landmines from the ground, as opposed to aerial remote sensing, registration is particularly challenging. The parallax effect induced by the proximity of the target to the sensors make homography estimation based on extrinsic parameters insufficient. Robust and non-rigid registration techniques are therefore required to mitigate the significant misalignments. Furthermore, inter-view occlusions are introduced by the parallax effect, as features visible in one sensor's perspective are not present in another. Consequently, the registration framework must consider non-correspondence regions using occlusion masks. Given that the sensor suite captures different wavelengths (RGB, SWIR, and long-wave thermal), the registration technique must also be multi-modal. To address these complexities, the RoMa (Robust Matcher) dense feature-matching architecture is employed (5). This model generates dense correspondence maps across modal pairs, complemented by a confidence map that serves as a proxy for estimating occlusion regions. Once a registered datacube is obtained, a supervised pixel-wise classifier is trained. By operating at the pixel level, the model uses the spectral signatures derived from the integrated sensor suite, thereby ignoring the impact of incomplete landmine spatial features. This methodology demonstrates that the integration of multi-modal sensor data significantly enhances the detection robustness of landmines compared to single-modality systems.

Authors: Esra Ozaydin

Abstract: Earthquakes lead to significant human and economic losses due to their unpredictability and destructive impact. Rapid and reliable mapping of building damage is a critical component of disaster response, enhancing situational awareness and directly supporting decision-making processes for emergency teams and recovery planning. Recent advances in GeoAI have enabled automated analysis of satellite imagery for damage assessment. However, the effectiveness of these approaches is critically dependent on the availability and quality of training data. Existing datasets, while valuable, are often limited in scale, geographic diversity, or representativeness of local building characteristics, reducing their applicability in practical deployments. The 2023 Turkey earthquakes provide a representative real-world scenario in which these challenges become particularly evident. Existing global benchmark datasets often fail to capture localized structural typologies and environmental patterns, leading to domain shift and reduced generalization capability. To address this gap from research to practice, models must be trained not only on clean and isolated features, but also on the inherent spatial complexity and radiometric challenges of real-world disaster scenes. In this study, we propose a difficulty-aware learning strategy for GeoAI-based change detection models. Using the publicly available KATE-CD Dataset as a case study, we investigate real-world constraints that affect model reliability.

Authors: Mohammad Sharif, Dirk Wittowsky

Abstract: Transportation networks are essential for the movement of people and goods and for maintaining access to critical resources and emergency services during disruptive events. Yet, many transportation systems include aging, deteriorated, or otherwise vulnerable components that may fail under extreme conditions such as natural hazards or human-induced disasters. As these events become more frequent and intense, increasing attention has been directed toward understanding and improving the performance of transportation systems under stress. In this context, resilience has emerged as a central concept. Transportation resilience describes the ability of a system not only to withstand and absorb the impacts of disruptions, but also to continue operating under adverse conditions and recover rapidly afterward (Diab & Shalaby, 2020). Assessing resilience can therefore support safer and more reliable infrastructure, better investment and management strategies, and broader socioeconomic stability (Esmalian et al., 2022). Consequently, the development of robust and operational methods for resilience assessment has become a pressing research need. To address this need, resilience assessment in urban transportation networks has generally evolved along two complementary lines: infrastructure-oriented assessment (Ashja-Ardalan et al., 2025) and operations-oriented assessment (Lu et al., 2024). The infrastructure perspective focuses on the structural robustness and physical condition of network assets, whereas the operational perspective emphasizes system functionality, including traffic performance and vulnerability to disturbances such as interruptions, shocks, and disasters. Because these perspectives target different aspects of resilience, the methods used also vary according to research objectives, data availability, and the complexity of the network under study. Existing approaches can broadly be classified into simulation-based methods, optimization models, graph-theoretical approaches, probabilistic techniques, and big-data-driven frameworks (Serdar et al., 2022). Together, these approaches have contributed a range of quantitative resilience metrics that support evidence-based planning and decision-making. However, their effectiveness increasingly depends on the availability of high-resolution data and methods capable of capturing the dynamic and interconnected nature of urban transport systems. This methodological demand has stimulated growing interest in the integration of new geospatial data sources with artificial intelligence (AI). The availability of large-scale and diverse data enables more adaptive and data-driven resilience models that incorporate a wider range of relevant variables and can be updated more efficiently as conditions evolve. For example, climate-related tools such as the Coupled Model Intercomparison Project (CMIP) processing framework provide access to regional projections of precipitation and temperature (Ashja-Ardalan et al., 2025), while GPS trajectories from smart devices offer valuable information on travel behavior and traffic dynamics during disruptive events (Alizadeh & Dodge, 2025). On the methodological side, graph neural networks (GNNs) are particularly well suited to transportation applications because they can capture both the topological  Correspondence: mohammad.sharif@uni-due.de structure of road networks and the spatial-temporal dependencies of traffic states (Fan et al., 2023; Wang et al., 2020). Recent studies have shown that neural-network-based methods can support resilience analysis across pre-disaster, disruption, and recovery phases. In this way, AI does not simply improve prediction accuracy; it also creates new possibilities for identifying critical components, modeling cascading effects, and designing more responsive resilience strategies. The recent shift from traditional AI to generative AI may further expand these capabilities in future transport resilience research (Wu et al., 2026). Against this background, the present study has two objectives. First, it reviews traditional and AI-based metrics and methods for assessing urban transportation network resilience, with particular emphasis on current developments, research gaps, and unresolved methodological challenges. Second, it demonstrates the practical relevance of these approaches by investigating how extreme rainfall affects the resilience of the urban road network in Duisburg, Germany. Specifically, the study considers rainfall intensities between 40 and 90 mm/h, representing progressively severe disruption scenarios. These conditions lead to lower travel speeds, traffic detours, and partial road closures, thereby degrading the functional performance of the traffic system and providing a realistic basis for resilience analysis under weather-related stress. To operationalize this analysis, we apply a spatio-temporal graph neural network (Huang et al., 2024) to identify critical road links in Duisburg’s urban network under normal and rainfall-affected conditions. Using average link speed and physical adjacency, the model learns spatial-temporal dependencies among road segments, while attention weights and an influence propagation mechanism are used to estimate each link’s network-wide importance. The results show that under normal conditions, the most critical links are concentrated on major arterials and key urban corridors. As rainfall intensity increases, however, criticality shifts toward surrounding roads and bottlenecks due to speed reductions and partial closures. Some links remain consistently important across scenarios, whereas others become critical only under extreme rainfall, highlighting the context-dependent nature of vulnerability. Overall, the framework captures both direct and cascading effects of disruption and provides an efficient basis for resilience-oriented planning and adaptation.

Authors: Michiel De Baets, Brian G. Booth, Hiep Luong, Michiel Vlaminck

Abstract: Floodwater segmentation on UAV imagery is an essential part of cost-effective forecast models and disaster assessment. Current techniques analyse individual RGB images with convolutional neural networks (CNN), but the variable appearance of floodwater leads to inaccurate and inconsistent results. This work describes a hybrid AI model that increases segmentation accuracy on flood scene videos by leveraging both temporal consistency and physical domain knowledge. CNNderived activation maps are projected on a high resolution digital elevation model (DEM) and aggregated within watershed basins to decide local floodwater levels. Validation on a novel set of four flood scene videos in Flanders, Belgium, shows significant performance improvements over competing methods, highlighting the critical role of incorporating physical domain knowledge and temporal consistency into floodwater segmentation.

Authors: Andreas Luyts, Jasper Wouters

Abstract:Timely and reliable flood extent mapping from satellite imagery is a critical capability for disaster response, supporting rapid situational awareness, resource allocation, and humanitarian decision-making. In operational contexts, however, satellite observations are often incomplete, asynchronous, and affected by cloud cover or revisit gaps. Most existing fusion approaches rely on near-simultaneous Sentinel-1 (S1) and Sentinel-2 (S2) acquisitions, limiting their robustness when data are missing and underutilizing valuable historical context. We propose a multimodal flood mapping framework designed explicitly for such imperfect, real-world conditions. The method requires only a flood-time S1 image—ensuring baseline operability—and can optionally incorporate floodtime S2, pre-flood S1, and pre-flood S2 observations when available. Each modality is encoded into a shared feature space using a remote-sensing foundation model. A lightweight context module and pixel-wise gating mechanism then adaptively weight the contribution of each modality before fusion. This design enables permutationinvariant aggregation and naturally handles missing inputs. Experiments on an extended Kuro Siwo dataset demonstrate that the approach effectively leverages historical and multimodal data when available, while maintaining strong performance when auxiliary inputs are absent. The framework is therefore well-suited for operational disaster response scenarios where data availability is uncertain.