We investigate the spatial and temporal dynamics of destruction across the Gaza Strip during the Middle East conflict that escalated sharply after the Hamas incursion into southern Israel on 7 October 2023 and subsequent Israeli airstrikes. Leveraging Synthetic Aperture Radar time series compiled from Sentinel-1 imagery, we derive data-driven assessments of conflict-related damage in an exceptionally hostile and data-scarce environment. Our primary objectives are to map the distribution of destroyed structures and reconstruct the timeline of damage progression. We employ coherence loss analysis to identify structural damage based on the satellite-derived temporal signatures. The workflow encompassed systematic data preprocessing, spatial analysis, and result validation against UNOSAT datasets to ensure reliability.

Pre-conflict analysis indicated that more than half of all structures were undamaged or only lightly affected, with 31% showing major damage. By late 2023, this distribution had shifted markedly: the proportion of undamaged or lightly affected buildings dropped to 22%, while severely damaged structures rose to 32% and completely destroyed buildings accounted for 10%. The damage further intensified through mid-2025, with severely damaged and destroyed buildings collectively representing over 80% of all assessed structures—highlighting a sustained and accelerating pattern of devastation.

The analysis reveals that the entire Gaza Strip experienced extensive structural loss, with densely populated urban areas emerging as persistent damage hotspots. By May 2025, all five districts displayed comparable destruction levels, though with distinct temporal trajectories. The near-total absence of intact or lightly damaged structures in multiple urban cores underscores the systematic and prolonged nature of bombardments, reflecting a transformation of the urban fabric unprecedented in recent conflict-driven damage assessments.

How to cite: Geiß, C., Andersch, E., Huber, M., and Taubenböck, H.: Sentinel-1 coherence loss analysis for damage assessment in conflict areas: Evidence from the Gaza strip following October 7th 2023, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7601, https://doi.org/10.5194/egusphere-egu26-7601, 2026.

Rapid situational awareness is essential for seismic resilience in tectonically active regions such as Taiwan. For critical maritime infrastructure, traditional post-earthquake reconnaissance is often constrained by limited accessibility and safety concerns, leading to delays in disaster response. This study presents an automated disaster monitoring framework that integrates UAV remote sensing and Geospatial Artificial Intelligence (GeoAI) to quantify seismic impacts on wharf facilities. High-resolution aerial imagery and multi-temporal geospatial data are combined to establish a processing pipeline for identifying disaster footprints, with particular attention to the spatial distribution of structural fissures and surface deformations. A YOLOv11-based deep learning model is employed for automated damage detection and segmentation. To enable quantitative assessment, morphological skeletonization and three-dimensional spatial analysis are applied to derive geometric characteristics of damage features. The extracted information is further used to compute the Pavement Condition Index (PCI) as an indicator of facility serviceability. Experimental results show that the proposed framework achieves mAP and Recall values exceeding 90%, with a spatial localization accuracy of ±2 cm. The results demonstrate the capability of the proposed approach to reduce the time required for post-earthquake damage assessment and to support disaster monitoring and infrastructure management in seismically active maritime environments.

How to cite: Huang, C.-J. and Cho, Y.-C.: Integration of UAV remote sensing and GeoAI for rapid post-earthquake disaster monitoring, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9181, https://doi.org/10.5194/egusphere-egu26-9181, 2026.

Landslides in the Western Corinth Rift reflect a mix of long-term “set-up” conditions—such as terrain, rock type, and fault-related structure—and short-term triggers linked to transient deformation and changing rainfall patterns. To represent these interacting processes in a clear and interpretable way, we propose a two-phase, multi-scale landslide susceptibility workflow based on explainable XGBoost. At Phase 1 (watershed scale) we develop a baseline susceptibility model using a standardized set of conditioning factors. These include (i) terrain and geomorphometric variables (elevation, slope, aspect, profile curvature, plan curvature and topographic wetness index (TWI) and (ii) lithological and structural controls (lithology and hydrolithology classes, distance from river network and fault-influence proxies such as distance to faults). The model is trained using historical landslide inventories, whereas interpretability was built in through explainable AI tools, such as SHAP, allowing us to quantify both global and site-specific contributions of conditioning factors, including key interactions. The result is a set of susceptibility maps paired with readable diagnostics that explain why certain areas are critical. At Phase 2 (local refinement and activity confirmation) focuses on the Krini–Gkrekas–Pititsa sector, where observations are denser and more reliable. Here, we evaluate whether susceptibility hotspots from Phase 1 align with evidence of ongoing or emerging instability. We add dynamic indicators and independent validation using: European Ground Motion Service InSAR ground motion, SBAS historical InSAR data; GNSS trend metrics and antecedent precipitation indices from station data. The goal is not just to refine local interpretation, but to test whether predicted patterns make physical sense, by checking consistency between (a) areas predisposed by lithology and structure and (b) present-day deformation signals and rainfall forcing. The workflow aims to produce decision-ready, interpretable outputs at two complementary scales: (1) watershed-scale susceptibility that highlights where failures are more likely based on relatively stable controls, and (2) a localized assessment that strengthens confidence where susceptibility coincides with measured deformation and hydrometeorological conditions. This improves trust and usability of AI-assisted landslide hazard assessment in tectonically active landscapes.

Keywords

Landslide susceptibility; XGBoost; explainable AI; SHAP; multi-scale modeling; watershed analysis; lithology; active faults; EGMS; InSAR ground motion; GNSS; antecedent precipitation index; Western Corinth Rift.

How to cite: Madonis, N., Ganas, A., and Tsangaratos, P.: Multi-Scale, Explainable XGBoost Landslide Susceptibility Mapping: From Watershed-Scale Controls to EGMS–GNSS–Rainfall Validation of Active Instabilities in the Western Corinth Rift, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13368, https://doi.org/10.5194/egusphere-egu26-13368, 2026.

Understanding the influence of topographic parameters on landslide susceptibility (LSM) is crucial for risk management in regions where landslides are recurrent and potentially catastrophic. Although landslides are often triggered by short-term external forcings such as intense rainfall, the expansion of human settlements onto steep slopes greatly amplifies their impacts, making prediction and mitigation increasingly urgent and challenging.

The Serra do Mar is a mountain chain extending over 1,500 km along the southeastern coast of Brazil, separating the inland plateau from the coastal plain and characterized by rugged relief strongly controlled by geological structures, including faults and steep escarpments. High seasonal rainfall combined with intense weathering makes this region naturally prone to landslides, as dramatically illustrated in February 2023, when extreme rainfall triggered widespread slope failures in the municipality of São Sebastião (São Paulo State), causing severe damage and loss of life.

Despite the importance of such events, traditional landslide susceptibility mapping approaches, largely based on field surveys and geotechnical analyses, are costly and time-consuming. Remote sensing combined with explainable machine learning offers a powerful alternative for large-scale spatial hazard assessment.

This study investigates how different Digital Elevation Model (DEM) resolutions affect predictive landslide susceptibility modeling using machine learning and explainable artificial intelligence (XAI) techniques. A multiscale set of topographic predictors was derived from airborne lidar and Copernicus DEMs. These predictors were integrated with a landslide inventory from the February 2023 event (1,070 mapped scars), which served as the reference dataset for training and spatially validating Random Forest susceptibility models, enabling a direct comparison of how different DEM resolutions reproduce observed landslide patterns. Model interpretability was then assessed using SHAP (Shapley Additive Explanations) to quantify scale effects and the relative contribution of topographic controls on landslide susceptibility.

How to cite: Ferreira, B., Viana, C., Coelho, R., Grohmann, C. H., Brenning, A., and Strohmaier, F.: Influence of topographic parameters on landslide susceptibility using machine learning: A case study in the municipality of São Sebastião, São Paulo, Brazil., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9970, https://doi.org/10.5194/egusphere-egu26-9970, 2026.

How to cite: Arellano, C. M., Hölbling, D., Nafieva, E., van Ek, J. J., Henriod, S., Rebois, Y., Schwingshandl, A., Forcieri, S., Heidrich, R., Hörbe, I., and Abad, L.: Interpreting landslide susceptibility models using explainable machine learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13408, https://doi.org/10.5194/egusphere-egu26-13408, 2026.

Groundwater pollution is a persistent, largely hidden risk in Mediterranean farming basins such as Western Thessaly (Greece), where heavy irrigation, seasonal recharge pulses, and highly variable geology can speed up the movement of contaminants from the land surface into aquifers, making intrinsic vulnerability maps essential for early warning, land-use decisions, and risk-aware governance; however, the widely used DRASTIC index—despite its practicality—relies on fixed weights and linear scoring, which limits its ability to capture nonlinear relationships and changing, time-dependent exposure. To overcome these constraints, we present a hybrid, explainable framework that strengthens the classic DRASTIC structure by introducing an eighth factor, Transit Time (TT), and pairing the resulting parameter set with a tree-based machine learning approach—centered on Random Forest (RF)—to improve predictive skill, spatial detail, and interpretability. We build and compare four configurations: a baseline 7-parameter DRASTIC map (DRASTIC A), an extended DRASTIC map with TT (DRASTIC B), an RF model trained on the original seven DRASTIC layers (RF A), and an RF model trained on the seven layers plus TT (RF B). The models draw on thematic raster layers (e.g., depth to groundwater, recharge, soil, aquifer media, vadose zone characteristics) sampled at nitrate monitoring locations, with TT included as a practical proxy for travel-time delay and attenuation processes that influence when and how strongly pollution signals reach the aquifer. Because spatial autocorrelation can inflate performance when using ordinary random splits, we adopt spatial cross-validation (block- and buffer-based schemes) to better test real-world transferability, address class imbalance with SMOTE, and evaluate outcomes using accuracy, F1-score, class-wise precision/recall, ROC-AUC, and confusion matrices, with special attention to correctly identifying high and very-high vulnerability areas. Among all approaches, RF B performs best (accuracy 0.8214; F1 0.8788), indicating that the combination of nonlinear learning and transit-time information yields clearer, more reliable discrimination of vulnerable zones than either index mapping alone or RF without TT. To make the models transparent and defensible for stakeholders, we apply explainable AI methods—permutation importance and SHAP—to reveal both overall driver rankings and local, pixel-level contributions; consistently, depth to groundwater, vadose zone influence, and recharge stand out as the strongest controls, while TT, although not always dominant in global importance, meaningfully sharpens the spatial tracing of vulnerable corridors and pathways. Finally, to support risk-informed planning under uncertainty, we produce confidence maps based on maximum predicted class probability and normalized entropy maps that summarize ambiguity across classes, clearly separating areas where the model is both confident and vulnerable from areas where predictions are uncertain and additional monitoring or field verification is justified; these layers are masked for nodata regions and designed for direct integration into management workflows. Overall, the proposed Random Forest–DRASTIC–Transit Time framework demonstrates how a spatially validated, explainable ML extension of DRASTIC can deliver more detailed, decision-ready vulnerability maps by blending static hydrogeologic controls with dynamic travel-time behavior, offering a scalable pathway for more sustainable groundwater protection as environmental pressures intensify.

How to cite: Tsangaratos, P., Matiatos, I., Ilia, I., and Markantonis, K.: Explainable Machine Learning for Spatio-Temporal Groundwater Vulnerability Mapping: A Random Forest-DRASTIC-Transit Time Framework for Western Thessaly, Greece, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21505, https://doi.org/10.5194/egusphere-egu26-21505, 2026.

Detailed information on building attributes, such as construction materials and structural types, is a fundamental prerequisite for accurate natural hazard risk assessment. Recent deep learning approaches based on convolutional neural networks (CNNs) have demonstrated the effectiveness of extracting such exposure-related information from street-level imagery, establishing a solid foundation for data-driven building characterization.

This study is motivated by the emerging capabilities of vision language models (VLMs), which leverage large-scale pretraining and generalized visual semantic reasoning to provide a unified framework for interpreting complex urban scenes. To assess their effectiveness in structural exposure modeling, we conducted comparative experiments using zero-shot inference and fine-tuning strategies. The dataset consists of over 29,000 annotated street-level façade images from the earthquake-prone region of Santiago, Chile.

The zero-shot results indicate that general-purpose off-the-shelf VLMs (e.g., InternVL2-8B) struggle to accurately infer complex structural engineering attributes due to insufficient domain-specific knowledge. In contrast, fine-tuning based on InternVL3-2B yields a substantial performance improvement: the model achieves high accuracy in building height estimation (90.6%) and roof shape classification (87.0%), and demonstrates strong performance in predicting lateral load-resisting system materials (78.8%) and complex seismic building structural types (SBST, 72.6%). These results suggest that, fine-tuned VLMs can effectively acquire domain expertise, enabling scalable and low-cost exposure modeling. Future work will further investigate the potential of VLMs to infer latent structural characteristics through semantic reasoning.

How to cite: Sun, Y., Abdelsalama, A., Xue, X., Aravena Pelizari, P., and Geiß, C.: Vision-Language Models for Structural Exposure Modeling from Street-Level Imagery, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-23030, https://doi.org/10.5194/egusphere-egu26-23030, 2026.

The Noto Peninsula, Japan, experienced two strong earthquakes within a short interval of approximately eight months in 2023 and 2024; the first event triggered only a limited number of landslides (28), whereas the second event resulted in widespread slope failures, with more than 2,300 landslides identified. This rare sequence provides a unique opportunity to investigate how landslide susceptibility and triggering mechanisms evolve under repeated seismic loading within the same tectonic and geomorphological setting. However, conventional landslide susceptibility studies typically treat successive earthquakes as independent events, overlooking the potential influence of prior seismic damage on subsequent slope failures.

In this study, we propose an interpretable, SHAP-based machine learning framework to analyze the temporal evolution of earthquake-induced landslide susceptibility during the 2023–2024 Noto earthquake sequence. An XGBoost model was first trained using landslide data from the 2023 event, during which landslide occurrences were sparse, and transfer learning was employed to enhance model robustness under small-sample conditions. SHAP-based interpretation indicates that landslide susceptibility in 2023 was primarily controlled by topographic and long-period seismic factors, with the top five contributors being elevation, surface roughness, slope gradient, long-period spectral acceleration (PSA at 3.0 s), and the topographic position index (TPI), reflecting a preconditioning process that brought slopes close to instability. The resulting susceptibility map was then compared with the spatial distribution of landslides triggered by the 2024 earthquake, revealing a pronounced spatial overlap between the 2023 high-susceptibility (potentially unstable) zones and the 2024 observed landslide locations. In contrast, SHAP analysis for the 2024 event shows a shift in dominant controlling factors toward roughness, peak ground velocity (PGV), TPI, mid-period spectral acceleration (PSA at 1.0 s), and slope gradient, indicating a release process in which pre-weakened slopes were driven beyond their stability thresholds by stronger and more velocity-dominated ground motion.

The results indicate a pronounced spatial correspondence between high-susceptibility areas identified after the 2023 earthquake and landslide occurrences in 2024, with a lift value of 2.80 for the top 5% susceptibility class. SHAP-based interpretation reveals a clear transition in dominant triggering factors between the two events. In 2023, landslide susceptibility was primarily controlled by long-period ground motion and topographic framework, reflecting a preconditioning process that brought slopes close to failure. In contrast, the 2024 earthquake activated widespread landslides through velocity-related and mid-period seismic components, representing a release process that pushed pre-weakened slopes beyond their stability thresholds.

These findings demonstrate that earthquake-induced landslides in the Noto Peninsula may follow a slope-state–controlled evolutionary pattern, in which earlier seismic events systematically modify slope conditions and strongly influence the spatial and mechanistic characteristics of subsequent failures. This study highlights the importance of incorporating inter-event interactions into landslide susceptibility modeling and provides new insights for post-earthquake hazard assessment in regions affected by sequential seismic events.

How to cite: Ye, C. and Oguchi, T.: Legacy Effects of Earthquake-Induced Landslides under Sequential Seismic Events:SHAP-Based Interpretation of Preconditioning and Release Processes during the 2023–2024 Noto Earthquakes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9306, https://doi.org/10.5194/egusphere-egu26-9306, 2026.

Landslides are one of the most destructive natural hazards, causing significant loss of life, extensive damage to infrastructure, and long-term disruption to socioeconomic development, particularly in rapidly urbanizing regions. Consequently, accurate landslide susceptibility mapping is a critical tool for effective hazard assessment and risk management. Although the extensive use of machine learning algorithms for landslide susceptibility mapping, the black-box nature of the models often limits the acceptance of model results by decision-makers. This study presents an explainable artificial intelligence framework for landslide susceptibility mapping that integrates SHapley Additive exPlanations (SHAP) with Recursive Feature Elimination (RFE) to optimize ensemble machine learning models. The proposed framework was tested on an arid and rapidly developing region in East Cairo, Egypt. A landslide inventory of more than 180 events was compiled from field surveys and satellite imagery, and fourteen conditioning factors representing topographic, geological, and anthropogenic controls were initially considered. Unlike traditional feature selection approaches that rely mainly on statistical importance, the proposed framework selects predictors based on their physical and geological contribution to slope instability. The results show that SHAP-based feature selection significantly reduces model complexity while maintaining high predictive performance, with only five predictors for Random Forest and nine for XGBoost. Beyond predictive performance, the framework provides clear physical and geological explanations for slope failure processes. SHAP interaction analysis identified two dominant instability mechanisms: human-induced factors within a 200 m buffer around the road cuts, as well as structural instability on slopes with orientations ranging from 225° to 320°, as expected from kinematic conditions for daylighting within the area of study. These findings demonstrate that explainable AI can move beyond black-box prediction by linking machine learning outputs to geological ground truth. Overall, this proposed framework offers a practical and interpretable tool for landslide hazard assessment and sustainable land-use planning, particularly in data-scarce and rapidly developing environments.

Keywords: Explainable AI, SHAP, feature selection, landslide susceptibility, Hazard assessment

How to cite: Abdelkader, M., Abriha, D., and Csámer, Á.: From Black-Box Predictions to Trustworthy Landslide Susceptibility Mapping Using Explainable AI, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13884, https://doi.org/10.5194/egusphere-egu26-13884, 2026.

Landslide susceptibility mapping is widely used for risk reduction, yet many high-performing deep models remain hard to interpret and rarely communicate where predictions are reliable. We present an explainable, confidence-mapped workflow that combines remote sensing/GIS-derived conditioning layers with modern deep tabular architectures (FT-Transformer, ResMLP, and TabNet). To test the developed methodology, a case-study area in the Regional Unit of Magnesia (Zagora–Mouresi, Greece) was selected. Conditioning factors describing terrain, hydrology, proximity, and geology are Frequency Ratio–weighted, then used to train probabilistic susceptibility models evaluated with discrimination and calibration metrics. Spatial confidence is mapped using normalized predictive entropy to identify zones where susceptibility estimates are less decisive. Explainability is achieved with SHapley Additive exPlanations (SHAP), consistently highlighting elevation as the dominant control, followed by aspect, with lithology and slope also exerting strong influence; proximity to the river network and faults and curvature-related metrics contribute secondarily. The resulting susceptibility and confidence products improve transparency for decision support and provide a scalable template for large-area hazard assessment.

How to cite: Chrysafi, A.-A.: Explainable, Confidence-Mapped Deep Learning for Remote-Sensing–Driven Landslide Susceptibility Mapping, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22979, https://doi.org/10.5194/egusphere-egu26-22979, 2026.

Flooding remains among the most damaging hydro-meteorological hazards in the Mediterranean. On Euboea Island (Greece), steep terrain, ongoing land-use change, and highly connected transport corridors can intensify both flood occurrence and potential consequences. This study presents an integrated, decision-oriented framework that jointly maps (a) flood susceptibility and (b) flood impact potential (exposure), combining Google Earth Engine (GEE) for predictor generation with Python-based machine learning, explainability, and uncertainty analytics.

A multi-source predictor database is assembled in GEE from satellite and ancillary datasets to represent key topographic, climatic, geological, and pedological controls on flooding. Terrain and morphometric predictors are derived from the ALOS 12.5 m DEM, including elevation, slope angle, plan and profile curvature, Topographic Wetness Index (TWI), and Topographic Position Index (TPI). Hydrologic connectivity is captured through distance to the river/stream network. Climatic forcing is represented using the Modified Fournier Index (MFI) from WorldClim v2.0 as a predictor variable for rainfall influence. Subsurface controls are incorporated via lithology (geological map) and topsoil texture (LUCAS database; sand, silt, and clay content), which modulate infiltration, storage, and runoff generation. Land-surface conditions affecting runoff are characterized using CORINE Land Cover 2018, reflecting vegetation cover and imperviousness patterns. In parallel, exposure is quantified using land-use intensity, building footprint/coverage metrics, and road-network descriptors (density, proximity, connectivity) to identify areas where flood impacts are likely to be most severe.

Flood occurrence labels are derived from an event inventory, and spatially explicit sampling and partitioning are applied to reduce spatial autocorrelation and improve generalization. Susceptibility is modeled using tree-based ensembles (Random Forest and XGBoost), trained and evaluated in Python with spatial cross-validation and metrics capturing both discrimination and reliability (AUC, F1/TSS, Brier score, and calibration diagnostics). To explicitly communicate confidence and reveal spatial weaknesses, we generate uncertainty and entropy maps: (a) predictive uncertainty estimated from ensemble dispersion and calibrated probabilities, and (b) Shannon entropy of class probabilities to highlight ambiguous transition zones, data-sparse areas, and geomorphologically heterogeneous corridors. Explainability is delivered via SHAP (global and local), supported by interaction and partial dependence analyses to identify dominant controls and to attribute exposure hotspots to drivers such as building and road concentration.

The resulting susceptibility, exposure/impact, and uncertainty–entropy maps provide transparent, decision-relevant information to support mitigation prioritization and strengthen trustworthy flood-risk screening on Euboea Island.

Keywords: flood susceptibility; exposure; impact mapping; Google Earth Engine; Python; tree-based ensembles; uncertainty; predictive entropy; SHAP; explainable AI; Euboea; Greece

How to cite: Papadomarkakis, D., Frousiou, M.-S., and Notas, A.: Explainable, Uncertainty-Aware Flood Susceptibility and Impact Mapping on Euboea Island (Greece) Using Google Earth Engine and Tree-Based Ensemble Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13768, https://doi.org/10.5194/egusphere-egu26-13768, 2026.

Heatwaves represent one of the most severe climate-related threats to European cities, where their impacts are intensified by urban heat island effects, aging populations, and uneven access to cooling resources and green infrastructure. Although heat-related risks are increasingly acknowledged in urban policy, many existing assessment frameworks continue to rely on conventional formulations that combine hazard, exposure, and vulnerability, grounded in the Intergovernmental Panel on Climate Change (IPCC). Such approaches inadequately capture the complex and dynamic interactions among climate processes, urban morphology, and socio-demographic vulnerability, thereby limiting their usefulness for designing locally targeted and context-specific adaptation strategies.

This study presents a spatiotemporal machine-learning framework for assessing heatwave risk in Bologna, Italy, following the IPCC risk concept. High-resolution environmental, infrastructural, and socio-demographic datasets covering the period 2014–2023 were compiled at the census-tract level. A Long Short-Term Memory (LSTM) neural network was developed to capture temporal dependencies in heatwave risk and optimized using the Hippopotamus Optimization Algorithm to improve predictive performance. The model integrates diverse set of 14 climatological, demographic, economic, and environmental indicators.

Examination of the results indicates a strong spatial agreement between observed and predicted heatwave risk patterns, with classification accuracies exceeding 77% for both low- and high-risk categories. Explainability analysis based on Partial Dependence Plots identifies temperature, vegetation cover, proximity to cooling and healthcare facilities, and the density of elderly female populations as the most influential determinants of heatwave risk. Future projections under RCP 4.5, 6.0, and 8.5 scenarios suggest a substantial expansion of high and very high heatwave risk classes by 2050. This expansion is most pronounced under the RCP 8.5 scenario, where areas classified as very high risk increase from approximately one-third of the urban area to nearly two-thirds.

The findings further highlight the mitigating role of urban green infrastructure, showing that higher vegetation density and improved proximity to green spaces can substantially reduce heatwave risk, albeit with spatially uneven benefits. By combining predictive capability with transparent interpretation, this framework offers practical, fine-scale evidence to support climate adaptation, nature-based solutions, and more equitable heat-resilient urban planning.

How to cite: Saber, A., De Luca, C., Pourzangbar, A., and L. Bell, M.: Mapping Urban Heatwave Risk with Explainable Spatiotemporal AI: Evidence from Bologna under Climate Change Scenarios, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14058, https://doi.org/10.5194/egusphere-egu26-14058, 2026.

Wildfires in Mediterranean ecosystems are increasing in frequency, extent, and severity under the combined influence of climate change and human pressure. This trend is intensifying the need for operational hazard products that are not only accurate, but also transparent, auditable, and easy to justify to decision-makers. Here we present an open-access, fully reproducible remote-sensing workflow for (i) pre-fire danger mapping and (ii) post-fire burn severity assessment, explicitly designed around explainability rather than black-box prediction. The workflow is implemented in Google Earth Engine using only freely available data sources: Sentinel-2 and Landsat 8 optical imagery, ERA5-Land meteorological reanalysis, and OpenStreetMap ancillary layers.

Post-fire impacts are standardized through NBR (Normalized Burn Ratio) and dNBR (difference Normalized Burn Ratio), converted into burn-severity classes using established USGS-style thresholds. Pre-fire danger is mapped using a physically interpretable, rule-based score derived from six binary, pixel-level indicators representing necessary conditions for elevated danger: (1) fuel availability (vegetation presence), (2) fuel dryness (SWIR-based moisture proxies), (3) heat (2 m temperature and/or LST), (4) atmospheric dryness (relative humidity), (5) wind speed, and (6) antecedent moisture deficit (recent precipitation/soil moisture). This structure provides built-in explainability, because each pixel’s class is directly traceable to the specific conditions that triggered it.

We demonstrate the workflow through a comparative analysis across four major Greek wildfire contexts, Attica, Euboea, Rhodes, and Evros, spanning different seasons and synoptic regimes. Using consistent pre-fire (multi-week) and post-fire compositing windows, we quantify how danger conditions co-occur prior to ignition, assess concordance between high-danger classes and observed fire perimeters, and relate pre-fire signatures to subsequent dNBR patterns, including differences associated with fuel structure, topography, and human exposure (proxied by proximity to roads and settlements from OpenStreetMap).

To move beyond qualitative map interpretation, we complement the rule-based danger score with two lightweight, fully explainable modeling layers that quantify driver effects and test cross-region generalization. First, we fit generalized additive models (GAMs) using continuous satellite- and reanalysis-derived predictors to recover nonlinear response curves and threshold-like behavior. Second, we use a hierarchical ordinal logistic regression in which baseline levels and selected driver effects can differ by region, enabling us to identify which driver–severity relationships are consistent across Mediterranean landscapes and which are site-specific.  We keep the models fully interpretable by reporting GAM response curves and logistic-regression odds ratios (with uncertainty), so predicted danger can be directly linked to physical drivers rather than opaque feature-importance scores. We generate all satellite/reanalysis-derived layers and danger/severity maps in Google Earth Engine, then export pixel-level predictor and outcome samples to fit the GAM and hierarchical logistic models in open-source Python, enabling transparent estimation of driver effects with uncertainty. Finally, we evaluate transferability using leave-one-region-out validation to identify where learned driver–danger relationships remain robust under differing regimes and where localized recalibration may be required for operational deployment.

Keywords

wildfire danger; burn severity; Google Earth Engine; Sentinel-2; Landsat 8; ERA5-Land; dNBR; generalized additive models; hierarchical logistic regression; explainable AI; transparent hazard mapping; Mediterranean ecosystems

How to cite: Notas, A., Frousiou, M.-S., and Papadomarkakis, D.: An Open and Explainable Google Earth Engine Workflow for Wildfire Danger and Burn Severity Mapping in Mediterranean Ecosystems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6499, https://doi.org/10.5194/egusphere-egu26-6499, 2026.

Landslide dams represent a major geomorphic hazard in the seismically active Himalayan belt, where temporary river blockages can lead to catastrophic outburst floods that impact downstream communities and infrastructure. Despite their importance, landslide dam susceptibility remains underexplored compared to conventional landslide hazard assessment. This study addresses this gap by developing a machine learning-based susceptibility model specifically targeting landslide dam formation, with the evaluation of spatial transferability between adjacent river basins. The following fifteen conditioning variables was compiled from diverse geospatial datasets: slope, aspect, elevation, plan curvature, relative relief, Topographic Wetness Index (TWI), distance to stream, distance to fault, distance to lineament, lithology, geomorphology, land use land cover (LULC), and median values of Normalised Difference Vegetation Index (NDVI), Normalised Difference Moisture Index (NDMI), and Normalised Difference Water Index (NDWI). A Random Forest (RF) classifier was implemented and trained exclusively on the Alaknanda basin and then applied to the neighbouring Bhagirathi basin for external validation, ensuring strict spatial separation between the training and test domains. The RF model achieved strong internal performance in the Alaknanda basin, and external validation in the Bhagirathi basin demonstrated robust transferability, with only modest performance degradation. Feature importance analysis revealed that elevation, NDMI, aspect and relative relief were the primary controls on dam formation. Susceptibility maps identified high-risk zones concentrated along deeply incised river valley segments, fault intersections, and areas underlain by high-grade metamorphic rocks. This susceptibility map may provide actionable information for disaster risk assessment, infrastructure planning, and the development of early warning systems in the Alaknanda–Bhagirathi river system and similar mountain regions worldwide.

How to cite: Joshi, S. and Siva Subramanian, S.: Landslide Dam Susceptibility Mapping in the Indian Himalayas: A Random Forest Approach with Cross-Catchment Validation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16199, https://doi.org/10.5194/egusphere-egu26-16199, 2026.