This session welcomes contributions exploring the integration of AI and data-driven methodologies in geomorphology. We invite studies applying supervised and unsupervised learning, neural networks, ensemble models, object detection, clustering, or dimensionality reduction techniques to analyze geomorphic features and processes. We particularly encourage submissions focusing on innovative workflows, reproducible pipelines, transfer learning, hybrid modeling, and the integration of remote sensing or DEM-derived datasets. Contributions addressing methodological challenges, model interpretability, and equity in data and tool access across different geographic regions are also welcome. Furthermore, contributions on remote sensing studies and AI applications in interplanetary geomorphology represent a step forward for our session.
The session aims to foster interdisciplinary dialogue among geomorphologists, data scientists, and Earth system modelers, and to outline future research directions toward a more transparent, inclusive, and methodologically robust use of AI in the geomorphological sciences.
Orals: Thu, 7 May, 08:30–10:15 | Room G1
Recent advances in Artificial Intelligence (AI) and computer vision have opened new opportunities for automated monitoring and analysis in geomorphology. Among these, optical flow methods represent a powerful approach for quantifying surface velocity fields in rivers using video data. This work presents the development and application of a low-cost optical flow tool designed to estimate river surface velocity from fixed monitoring points, offering a practical and scalable solution for hydrological and flood-risk management applications.
The proposed method relies on the implementation of an AI-assisted optical flow algorithm capable of tracking the motion of water surface patterns (e.g., ripples, foam, floating debris) in standard RGB video sequences. By leveraging dense flow estimation and adaptive filtering, the tool produces high-resolution velocity maps that can be continuously updated in near real time. The system has been tested in different riverine environments, showing robust performance under varying lighting and flow conditions, and demonstrating its ability to capture both steady and transient flow dynamics.
One of the key strengths of this approach lies in its low operational cost and flexibility. The method can be implemented using conventional cameras and open-source software, eliminating the need for expensive. This makes it particularly suitable for establishing permanent observation points in critical areas, such as flood-prone zones or regions with limited monitoring infrastructure. Continuous optical monitoring of surface velocity provides valuable information for calibrating hydrodynamic models, identifying changes in river morphology, and supporting early-warning systems for extreme hydrological events.
Beyond the technical development, this research emphasizes the importance of integrating AI-based monitoring tools within broader frameworks for territorial management and risk mitigation. Establishing collaborations with stakeholders such as Basin Authorities, local governments, and civil protection agencies—can significantly enhance the effectiveness of these systems. Shared data platforms and automated AI-driven analytics could enable more proactive responses to extreme events, improving preparedness and resilience in flood-prone communities.
In future developments, the integration of deep learning models for feature detection and noise reduction could further enhance the accuracy and robustness of surface velocity estimation. Combining optical flow data with other remote sensing sources (e.g., UAV imagery, satellite observations) could also provide a multi-scale understanding of fluvial dynamics. This research thus contributes to the growing field of AI applications in geomorphology, highlighting how intelligent, low-cost monitoring systems can play a crucial role in sustainable river management and flood risk assessment.
How to cite: Sabato, G., Luparelli, A., Chirivì, M., and Lupi, A.: Optical Flow-based Tool for Surface Velocity Monitoring in River Systems: A Step Toward AI-driven Flood Risk Management, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1142, https://doi.org/10.5194/egusphere-egu26-1142, 2026.
Badland landscapes, characterised by intensely dissected slopes on unconsolidated sediments or soft rocks, are crucial hotspots for understanding soil erosion and sediment transport dynamics. Consequently, alongside their rapid climatic responses and links to anthropic land use, the geomorphological processes driving badlands morphogenesis are widely studied; recently, approaches combining multiplatform remote sensing and Machine Learning (ML) have been proposed due to their superior performance compared to other statistical models.
This study develops an integrated and multi-source approach using detailed geomorphic and hydrological parameters through a Random Forest (RF) algorithm to obtain a high-resolution land cover classification and erosion susceptibility maps of badland landscape in Basilicata Region (Italy). The workflow analyses geomorphological data at the micro-topography scale (3 cm/px) for geomorphometric landscape classification. Topographic and hydrological predictors were extracted from high-resolution Digital Elevation Models (DEMs) and orthomosaics were derived from optical images acquired in a drone survey conducted in May 2025. Spanning 0.025 km², the study area exhibits characteristic badland morphologies, located on poorly cemented silty clays from the Lower Pleistocene. Two different experiments were conducted. In the first one, ten topographic and hydrological predictors (e.g. Topographic Position Index, aspect, profile and tangent curvatures, Stream Power Index) were computed using open-source GDAL and GRASS GIS tools to assemble a multi-layer spatial dataset. In the second experiment, the R, G and B bands from the optical orthomosaics were also considered and included as three additional predictors. In both the experiments, 9,900 training points and 3,000 test points were extracted from the dataset to conduct a spatial cross-validation. Following the accuracy assessment, the algorithm was retrained on the full dataset to generate: i) a land cover map of three features: ‘Badland’, ‘Vegetation’ and ‘Pediment’; and ii) an erosion susceptibility map based on the probability of a pixel belonging to the ‘Badland’ class. The first experiment using only morphometric predictors showed a global accuracy of 82.49%, while the second experiment integrating the three RGB bands increased accuracy to 97.43%.
How to cite: Sozio, A., Marsico, A., Colacicco, R., La Salandra, M., Muscillo, S., Refice, A., and Capolongo, D.: Machine Learning techniques for the detection of geomorphological features in badland landscapes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7293, https://doi.org/10.5194/egusphere-egu26-7293, 2026.
In formerly glaciated regions, high resolution multibeam bathymetry is crucial to understand submarine glacial landforms and the processes that formed them. Using submarine glacial landforms to study past glacial dynamics contributes to the understanding of how glacial margins will change in the future. The increase in mapped seafloor coverage also creates an opportunity to develop and deploy automated models for recognizing and classifying glacial and glacimarine features. Datasets of glacial and glacimarine features were developed from manual interpretation of multibeam bathymetry from the deglaciated shelves of Svalbard and Antarctica. Features were assigned to one of 10 classes: crag and tails; flutes and drumlins; grounding zone wedges; mega-scale glacial lineations; channels; iceberg ploughmarks; large moraines; small moraine ridges; planes; and bedrock structures. Additional morphological analysis was performed for each feature in both datasets. For each sample, inputs to the model included bathymetry, slope, aspect, profile curvature, tangential curvature, and the rasterized outline of the feature of interest. We performed classification using six established Convolutional Neural Network (CNN) architectures, including ConvNeXtTiny, DenseNet201, EfficientNet, MobileNet, ResNet50, a baseline Simple CNN in addition to a Genetic Conditioned Convolutional Neural Network (GC-CNN). Three approaches were taken using the datasets: 1) Training separate models using the Svalbard and Antarctica datasets with testing only on the same dataset. 2) Training separate models using the Svalbard and Antarctica datasets and testing on the other dataset. 3) Training a combined model using all available data from both datasets and testing on a separate dataset from North Greenland. Our results show that morphological differences between features from different regions have a significant effect on machine learning model accuracy. Developing robust glacial landform classification models that can be applied to features from all regions require data that capture the variability of a particular feature class.
How to cite: Danielson, M., Truong, T., and Jakobsson, M.: Comparison of Machine Learning Results of Glacial Landform Classification of Datasets From Svalbard and Antarctica, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10389, https://doi.org/10.5194/egusphere-egu26-10389, 2026.
Recent advances in deep learning, particularly convolutional neural networks (CNNs), together with the increasing availability of high-resolution multibeam bathymetry data, have made automated classification of glacial landforms increasingly feasible. However, robust and scalable classification remains challenging. The main challenge is not only computational scale, but the mismatch between data-driven learning and the physics-governed nature of glacial terrain. To address this gap, we propose a physics-informed deep learning framework for automated classification of submarine glacial landforms from multibeam bathymetry. Unlike conventional CNNs that rely purely on data-driven features, our approach integrates physically meaningful constraints reflecting glacial geomorphology. Specifically, we develop a Physics-Informed Genetic-Conditioned Network (PI-GCNet) by integrating a genetic-conditioned layer into a standard CNN and introducing a physics-guided loss that enforces geomorphological consistency. Each landform class is represented by a learnable genome vector initialized from class-wise statistics of depth, slope, and curvature and trained jointly with the embeddings. A genetic attraction and repulsion mechanism structures the latent space, and classification is performed via an energy-based distance between embeddings and genomes. We further optimize a composite objective combining genetic Cross-Entropy, genome regularization for stable and interpretable representations, and a class-wise slope deviation loss penalizing departures of predicted mean slope from expected class-specific values. Together, these components enhance robustness and interpretability, enabling scalable and physically consistent mapping of submarine glacial landforms. We validate the proposed model using high-resolution multibeam bathymetry acquired in northern Greenland. We prepared a dataset of 1515 samples across 10 classes of glacial and glacimarine features identified through manual interpretation. Model inputs include bathymetry, slope, aspect, profile curvature, tangential curvature, and the rasterized outline of each feature. In addition, we compare PI-GCNet with state-of-the-art deep learning baselines to demonstrate reliability and improved generalization.
How to cite: Truong, T., Danielson, M., and Jakobsson, M.: Physics-Informed Genetic-Conditioned Network for Glacial Landform Classification, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10851, https://doi.org/10.5194/egusphere-egu26-10851, 2026.
Drumlins are key indicators of subglacial processes and former ice-flow patterns, but regional drumlin inventories are often compiled manually, limiting scalability and reproducibility. This contribution presents a workflow for drumlin segmentation, prediction, and GIS-ready analysis using deep learning applied to digital elevation models (DEM) and an existing, manually mapped, drumlin reference shapefile from Latvian drumlin fields. The DEM is tiled into 640×640 image patches, and reference polygons are converted into You Only Look Once (YOLO)(Ultralitics S.a.) segmentation labels. A YOLOv11 segmentation model is trained for drumlin delineation and then used to generate predictions across the tiled DEM. Predicted outputs (TXT masks) are converted back into georeferenced polygons and exported as GeoPackages, enabling immediate integration with standard GIS-based morphometric analysis and mapping methods.
Model performance is evaluated using both standard YOLO metrics and an inventory comparison against the control dataset. The statistics for bounding boxes, precision and recall reach 0.808 and 0.652, with mAP50 of 0.755 (mAP50–95: 0.514). For segmentation masks, precision and recall are 0.757 and 0.601, with mAP50 of 0.672 (mAP50–95: 0.275). Inventory comparison yields 1190 predicted vs 1146 control drumlins, with 906 true positives, 284 false positives, and 247 false negatives, corresponding to precision 0.761, recall 0.786, and F1 0.773 (nMCC: 0.387).
The results demonstrate that YOLO-based segmentation can produce georeferenced drumlin polygons at scale with quantified uncertainty, providing a practical route toward repeatable drumlin inventories and downstream geomorphological analyses.
This research was funded by the Latvian Council of Science, project "Reconstruction of ice stream dynamics and deglaciation of the SE sector of the Scandinavian Ice Sheet in Latvia", project No.lzp-2024/1-0193
How to cite: Puriņš, H. H.: ML assisted drumlin inventory: YOLOv11 segmentation, polygon reconstruction, and accuracy assessment, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11575, https://doi.org/10.5194/egusphere-egu26-11575, 2026.
Accurate characterisation of nearshore bathymetry is crucial for modelling waves, monitoring hazards, and managing coasts.
Ship-based technologies produce accurate bathymetry measurements but are expensive, time-consuming, and dangerous to use in relatively shallow coastal terrain, areas with rock outcrops, and intense waves. This results in a lack of bathymetry data in shallow coasts, creating a data void in nearshore zones.
Several studies have used remotely sensed imagery for inverting nearshore bathymetry in shallow, homogeneous sandy coastal systems. However, existing empirical models often fail to estimate accurate bathymetry in energetic rock coast environments. To date, little or no attention has been placed on deriving nearshore bathymetry in rock coastlines due to high turbidity and heterogeneous bottom substrates. Empirical models do not perform effectively in determining nearshore water depth, which explains the greater focus on soft-coast systems and the neglect of rocky coastlines.
Multispectral (MS) optical sensors are preferred for bathymetry studies. The blue and green bands of the optical satellite image are limited in depth by turbidity, large waves, and currents. In contrast, Synthetic Aperture Radar (SAR) can track surface roughness and infer water depths in active waves and swells. By integrating passive optical and active SAR data, the limitations of relying solely on multispectral images to derive bathymetry in complex coastal areas can be overcome.
Rocky coasts are characterised by morphological heterogeneity that promotes increased turbulence under wave forcing. This affects light penetration and increases the complexity of optical water properties, reducing the accuracy of projected nearshore water depth using only optical sensors. While the western seaboard of Ireland is dominated by hard, rocky, cliffed coastline and extreme wave climates, it has a significant data gap in nearshore bathymetry.
In a multi-sensor approach, this research used machine learning techniques to combine multispectral Sentinel-2 and Sentinel-1 SAR data with multibeam data to provide a comprehensive dataset for projecting nearshore bathymetry. This was carried out using supervised machine learning to train known depth values in a given coastal area, and unsupervised machine learning to predict unknown water depths in another nearshore zone. Projected water depth was validated using single-beam sonar data collected in Ballard Bay.
Random Forest, XGBoost, and LightGBM, techniques were used to train models. Models were then applied to generate nearshore bathymetry maps at three locations on the west coast of Ireland: Ballard Bay, Farrihy Bay, and Loop Head. The results indicated that Random Forest outperformed the XGBoost and LightGBM models across all sites, with R² values of 0.782, 0.765, and 0.740, respectively, with corresponding Root Mean Square Error (RMSE) of 0.578m, 0.698m, and 0.756m. A stacking ensemble model was built by combining the three models, which improved the R² for bathymetry prediction by 10% at all sites.
This research represents one of the first applications of machine learning–based nearshore bathymetry reconstruction focused specifically on rocky coastlines. The proposed ensemble method can produce precise bathymetric maps for nearshore areas across diverse regions and time periods. This will enable frequent assessment of rock coast evolution in relation to potential climate-driven impacts.
How to cite: Arhin, M., Cullen, N., Hegarty, S., and Bourke, M.: Multi-Sensor Based Nearshore Bathymetry Projection in Rocky Coastlines Using Machine Learning Techniques, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13506, https://doi.org/10.5194/egusphere-egu26-13506, 2026.
Bathymetric mapping of the Arctic seafloor remains challenging due to persistent sea ice, which limits systematic surveys and degrades multibeam echosounder (MBES) data through reduced swath width, ice interference, and vessel-induced noise. As a result, Digital Terrain Models (DTMs) derived from MBES data in ice-covered regions are often fragmented, coarse, and incomplete, obscuring bottom morphological features. These includes submarine glacial landforms that are informative of past glacier extents and ice-sheet dynamics. Standard interpolation is commonly used for upsampling and gap-filling but systematically oversmooths seafloor morphology, removing the small-scale variability central to glacial geomorphological interpretation.
Here, we investigate whether domain-informed generative super-resolution can recover geomorphologically meaningful structure in degraded Arctic bathymetry. We target upscaling from 100–200 m grid cell resolution to 25 m (4×–8×), with explicit emphasis on preserving glacial landforms rather than optimizing pixel-wise fidelity. We compile (i) 25 m MBES-derived bathymetry from surveys near northern Greenland and around Svalbard, which is downsampled to 100 m and 200 m to create controlled low-resolution inputs, and (ii) a larger set of terrestrial post-glacial Digital Elevation Models (DEMs) from Norway, Iceland, and the Hudson Bay region derived from airborne LiDAR and satellite products. The terrestrial DEMs provide a geomorphological prior without hydroacoustic artifacts and are used for training, while MBES data are reserved exclusively for evaluation in the Arctic bathymetry use case, acknowledging the domain shift between terrestrial and submarine environments.
We train a conditional diffusion model with a U-Net backbone to generate 25 m terrain conditioned on low-resolution inputs. In controlled downsampling experiments, conventional super-resolution metrics show limited separation from deterministic baselines; however, distributional similarity, quantified using the Wasserstein distance of elevation-value distributions, consistently improves. Qualitative assessments in regions such as Svalbard, Nares Strait, and Victoria Fjord show that the diffusion model produces sharper glacial lineations, more distinct retreat moraines, and clearer iceberg scour patterns than interpolation-based methods. To better quantify these geomorphological improvements, we introduce a Fourier-domain evaluation based on radial power spectral density and cross-correlation. Frequency-domain analysis shows that diffusion outputs more closely match the spectral characteristics of the 25 m reference data and tend to restore mid-wavelength power associated with glacial bedforms. Overall, the results suggest that domain-informed generative super-resolution can produce more interpretable bathymetric grids, while underscoring the need for evaluation metrics aligned with geomorphological realism.
How to cite: Matwiejczuk, J., Jasche, J., Mayer, L., Mohammad, R., and Jakobsson, M.: Diffusion-based Super-Resolution of Arctic Bathymetry for Glacial Geomorphology, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18865, https://doi.org/10.5194/egusphere-egu26-18865, 2026.
Artificial Intelligence is increasingly reshaping geomorphological research by enabling scalable, data-driven analyses of complex Earth surface processes. In this contribution, we present a supervised machine-learning framework for reconstructing Late-Quaternary coastal paleo-landscapes, applied to the rocky coasts of the Cilento Promontory (southern Tyrrhenian Sea), a tectonically quasi-stable sector preserving well-constrained sea-level indicators.
We trained a Random Forest classifier on an expert-labelled geomorphological dataset integrating DEM-derived morphometric parameters, lithology, distance from the coastline, and field-validated paleo-environmental markers. The model was developed within a fully reproducible workflow and validated against independent geomorphological mapping and sea-level proxy datasets.
Results demonstrate high classification performance and the ability to automatically discriminate between Last Interglacial paleo-sea cliffs and polycyclic, currently active coastal cliffs across different lithological contexts. The AI-based approach overcomes key limitations of traditional “bathtub” methods, allowing the detection of relict and partially buried landforms and extending paleo-landscape reconstructions into areas lacking direct field evidence.
Beyond the specific case study, this work illustrates how machine-learning approaches can be effectively integrated with geomorphological knowledge to reconstruct complex coastal paleo-landscapes. The proposed framework allows the identification of inherited and partially obscured landforms that are difficult to detect through traditional methods alone, offering a transferable tool for investigating long-term coastal evolution. This integration of AI and geomorphology provides new insights into the geomorphic response of rocky coasts to Quaternary sea-level fluctuations and climatic forcing.
How to cite: Sorrentino, A., Mattei, G., Pappone, G., Ciaramella, A., and Aucelli, P. P. C.: Machine learning–based reconstruction of Late-Quaternary coastal paleo-landscapes: an AI framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21186, https://doi.org/10.5194/egusphere-egu26-21186, 2026.
Human activities are increasingly degrading European coastal seabed, highlighting the need for efficient monitoring tools. To address these impacts, the Marine Strategy Framework Directive (MSFD) was established in 2008 with the aim of protecting and preserving marine ecosystem while ensuring their sustainable use.
Within this framework, high-resolution seafloor mapping represents a fundamental tool for coastal governance, habitat monitoring and marine geological studies. Fine-scale survey using Side-Scan Sonar (SSS), often integrated with AUV systems, provides detailed information on seabed morphology and acoustic facies, supporting habitat mapping and coastal management.
Recent advances in artificial intelligence (AI) and machine learning (ML), combined with Geographic Information Systems (GIS), have significantly improved the automated interpretation and mapping of submerged morphological features from marine geophysical data. In this context, this study investigates the application of computer vision techniques to nearshore environments along the Italian coastline, with a specific focus on the Segment Anything Model (SAM) framework.
We evaluate both the foundation-model implementation of SAM (SAM2) and its prompt-driven variant (SAM3) for the detection and classification of seafloor features in high-resolution SSS mosaics. The analysis was conducted on three nearshore areas in the Apulia region (southern Italy): Torre Guaceto (BR), Leporano (TA) and Zapponeta (FG). Using SAM2, we implemented a supervised workflow in which manually segmented features, including ripple marks, coralligenous formations, seagrass beds, sandy plains and rocky outcrops were used to train a Contrastive Captioner (CoCa) classifier via transfer learning. This approach achieved a test Macro F1 score of approximately 0.90.
In parallel, the text-promptable SAM3 model was employed for zero-shot segmentation of small, high-backscatter point-like features (“white dots”). These features were validated through an independent AUV-based video survey and confirmed to correspond to coralligenous formations. This result demonstrates the capability of SAM3 to rapidly extract geologically meaningful targets without requiring prior training.
Future developments will focus on alternative strategies and model refinement to further improve detection accuracy and robustness.
How to cite: Parisi, F., Scarrica, V. M., De Giosa, F., and Staiano, A.: Automated seafloor feature recognition in Side-Scan Sonar data using Segment Anything Models (SAM): a case study from Apulian coastal nearshores, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3878, https://doi.org/10.5194/egusphere-egu26-3878, 2026.
Posters on site: Wed, 6 May, 10:45–12:30 | Hall X3
The Mediterranean Sea is often affected by tropical-like cyclones, which cause heavy rainfall, strong winds, storm surges and flooding. The accurate classification of precipitation into convective and stratiform within these systems is essential for understanding storm dynamics and improving predictive models. In this study, we developed a deep learning approach based on U-Net architecture to classify convective and stratiform precipitations during Mediterranean cyclones using the Global Precipitation Measurement (GPM) IMERG product. We derived physically consistent labels for training through an exponential distribution-based thresholding of rainfall intensities. The trained U-Net model effectively reproduced the spatial structure of convective rainbands and surrounding stratiform regions within the cyclone structure. In addition to validating the convective precipitation detection using brightness temperature satellite observations and ERA5 reanalysis, we also incorporated pluviometer records. These ground-based measurements confirmed the model’s strong capability to identify areas affected by convective precipitation. This study demonstrates the potential of integrating a physics-based approach with deep learning for high-resolution characterization of precipitation in Mediterranean cyclones. While the segmentation of convective precipitation alone does not directly quantify coastal hazard, these results provide essential input layers for downstream coastal-impact assessments. In particular, the high-resolution identification of convective rainfall can be integrated into hydrological and hydraulic models (e.g., HEC-RAS or similar) to simulate surface runoff, flash-flood dynamics, and related coastal impacts under a changing climate.
How to cite: Kushabaha, A., Alemán, J. J. G., Miglietta, M. M., Mastrangelo, D., and Panegrossi, G.: Evaluation of a deep learning model to classify convective and stratiform precipitation patterns , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1804, https://doi.org/10.5194/egusphere-egu26-1804, 2026.
Alluvial-coastal plains and shallow-marine areas are facing increasing hydrological pressure from rising sea levels, changing precipitation patterns, and more frequent extreme events that alter their water balance. Prolonged droughts and unsustainable freshwater withdrawals further exacerbate the problem, causing saline intrusion into already vulnerable aquifers. In this context, adjusting drainage channel levels has become a key climate adaptation strategy. Our work combines long-term forecasting of aquifer dynamics using hydro-climatic data, including meteorological forecasts, with near-real-time forecasting to optimize the operation of high-capacity pumping stations that safeguard the drainage network of San Rossore Migliarino Massaciuccoli Regional Park (Pisa, Italy). The study addresses two complementary forecasting approaches: 1. Extended-range forecasting of aquifer levels through the integration of historical groundwater measurements and meteorological forecasts, with particular emphasis on precipitation and temperature projections over a two-week period. This approach provides weekly predictive outputs essential for determining optimal operational thresholds for pump activation, accurately tailored to varying seasonal and meteorological conditions throughout the year. 2. An operational, near-real-time forecasting framework designed to support daily management decisions. This system incorporates real-time data on meteorological conditions, groundwater levels, channel hydrometric levels, and drainage system activity to serve as an early warning mechanism during exceptional events requiring prompt intervention. Overall, preliminary results unequivocally highlight how integrated long-term and near-real-time operational forecasting systems are now indispensable for sustainable and resilient water resource management. Such systems allow for proactive anticipation of critical conditions, optimized drainage system use, reduced energy consumption, and preservation of the hydro-saline equilibrium in coastal aquifers. This emphasizes the necessity for continuous monitoring and technologically advanced solutions in fragile alluvial-coastal plains facing intensifying climatic and anthropogenic pressures.
How to cite: Lupi, A.: Integrated forecasting approaches for optimizing alluvial-coastal drainage systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4986, https://doi.org/10.5194/egusphere-egu26-4986, 2026.
Obtaining information on the size and shape of individual sediment grains is fundamental to many geoscientific applications, as these properties provide insights into sediment transport and depositional processes. Conventional approaches for grain size and shape analysis rely on manual or semi-automated workflows and are therefore labor-intensive and time-consuming. Recent advances in deep learning, particularly in image segmentation and object detection, have enabled the development of automated methods for measuring grain size and shape. However, existing approaches typically are trained on homogeneous, task-specific datasets, which limits their ability to generalize across different data types and settings. Additionally, challenging image characteristics often compromise the segmentation accuracy.
We present an upgraded version 2.0 of the ImageGrains framework (Mair et al., 2024), which leverages a recently published next-generation segmentation model, Cellpose-SAM (Pachitariu et al., 2025). We re-trained this model on our newly released open-access dataset comprising 243 images and more than 29,000 manually annotated sediment grains. This dataset consists of images from various settings, including photographs of fluvial gravel, coarse pro-glacial deposits, and X-ray computed tomography scans of glacial till and marine sand. We use these data to benchmark the segmentation performance of the method against ground-truth annotations and to compare it to the performance of existing segmentation methods.
The results show that ImageGrains 2.0 achieves higher segmentation accuracy and improved generalization to previously unseen data compared to current state-of-the-art methods. When comparing the size and shape of grains predicted by the model with the ground truth annotations, we find that the increase in segmentation accuracy of our upgraded framework directly translates to more precise and realistic morphometric results, such as grain size distributions. We make our framework available to the community as a free and open-source installable Python package, as well as through interactive computing environments such as Jupyter Notebooks and a graphical user interface.
References
Mair et al. (2024): Automated detecting, segmenting and measuring of grains in images of fluvial sediments: The potential for large and precise data from specialist deep learning models and transfer learning, ESPL, 49, 1099–1116, https://doi.org/10.1002/esp.5755.
Pachitariu, et al. (2025): superhuman generalization for cellular segmentation, https://doi.org/10.1101/2025.04.28.651001.
How to cite: Mair, D., Witz, G., Do Prado, A., Garefalakis, P., Wild, A., Ville, F., Schuster, B., Horn, M., Österle, J., Fabbri, S., Litty, C., Achleitner, S., Leistner, S., Hiller, C., and Schlunegger, F.: Introducing ImageGrains 2.0 for improved grain size and shape measurements in 2D and 3D data from images of sediment, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10505, https://doi.org/10.5194/egusphere-egu26-10505, 2026.
Badlands are landscapes that develop on poorly consolidated bedrock under high erosion rates and their spatial distribution in Italy is related to marine clays outcrops. These landscapes are vulnerable to dynamic changes driven by geological processes, biological and anthropic factors at different spatial and temporal scales. This characteristic, along with the rapid transformation of landforms makes them ideal open-field laboratories, suitable for testing and improving means for predictive analysis. The study focuses on Badlands landscapes of the Basilicata region (Southern Italy) to explore the potential of the integration of different remotely-sensed data, from ground-based to satellite-based technologies, in the definition of a probabilistic model for the evaluation of erosion trends. The integrative approach is essential for a comprehensive study of landscape dynamics, accounting for the complex interactions between top-down drivers (climatic and anthropogenic) and bottom-up drivers (biotic and geomorphological factors) of landforms evolution. In this perspective, Machine Learning (ML) techniques are effective tools for analysing the spatially heterogeneous responses of different morphologies in order to study their evolution and, therefore, their susceptibility to various processes (landslides, soil erosion). Maximum Entropy (MaxEnt) distribution models estimate a target probability distribution as a function of environmental predictors based on presence-only data. This approach has found several applications in the field of geomorphology, providing a more user-friendly and accessible way to perform studies based on statistical predictions. In contrast, MaxEnt has proven to be less robust in accuracy than other ML models. To overcome this issue while still guaranteeing accessibility, the study proposes the use of an advancement of the classical MaxEnt software called spatialMaxEnt, more sensible to the spatial distribution of data due to spatial cross-validation and a series of optimized functionalities to minimize overfitting. Presence data comprehends sites of occurrence of erosion processes identified through high-resolution topographic data within three study areas (Aliano, Tursi and Montalbano Jonico), externally grouped based on spatial autocorrelation, while environmental variables include topographic, climatic and anthropic attributes. A multicollinearity analysis using Pearson’s correlation coefficient is conducted prior to the MaxEnt modelling to identify and exclude highly correlated variables. This procedure ensures the selection of the most explanatory predictors and reduces the risk of model overfitting. The output data of the study is represented by Badlands susceptibility maps. This study further recommends comparisons with other AI-based models, as well as their possible combination, to enhance the robustness and significance of the results. Despite the challenges, the layers produced and the implementation of methodologies for modelling and identifying erosion-prone areas remains a fundamental tool for environmental monitoring and decision aimed at the conservation of the geological heritage.
How to cite: Santoro, G., Colacicco, R., Capolongo, D., and Marsico, A.: Integration of Remote Sensing techniques to assess badlands susceptibility in the Basilicata region (Southern Italy), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12111, https://doi.org/10.5194/egusphere-egu26-12111, 2026.
Large scale landslide susceptibility modelling (LSM) has rapidly evolved with the availability of large landslide inventories and high-resolution global environmental datasets. However, two fundamental issues continue to undermine the reliability and interpretability of such models: spatial autocorrelation (SAC) in landslide occurrences and strong feature redundancy (FR) among terrain-derived and environmental predictors. Although both factors are known to affect model outcomes, their relative and distinct impacts on predictive performance and model interpretability at large to global scales are still poorly disentangled. Here, we present a comparative modelling study aimed at systematically evaluating how SAC and FR differently influence large-scale landslide susceptibility models outcome. Using the UGLC (Unified Global Landslide Catalog, https://essd.copernicus.org/preprints/essd-2025-482/), which includes more than one million landslide records and an equal number of geomorphologically constrained non-landslide samples, we implement Random Forest (RF) models under three experimental settings: (i) conventional random train–test splitting, (ii) spatial k-means clustered train–test splitting to mitigate SAC-induced bias, and (iii) random train–test splitting combined with feature de-correlation to mitigate multicollinearity among 60 global geomorphological, hydrological, geological, and soil predictors. Our results show that random splitting leads to strongly optimistic performance (accuracy ≈ 0.96), dominated by spatial dependence between training and testing samples. When spatial clustering is applied, model performance decreases markedly (accuracy ≈ 0.82), exposing the true predictive capability under spatial independence. Feature de-correlation does not address SAC-related bias but produces measurable gains in robustness (accuracy ≈ 0.94), improving model interpretability across diverse climatic and geomorphological settings. These findings highlight that SAC and FR affect global LSM in fundamentally different ways. A spatial train–test splitting is indispensable for unbiased performance assessment, whereas feature de-correlation serves as a complementary strategy to enhance model stability and interpretability. This distinction is critical for the development of scientifically interpretable and operationally reliable global landslide susceptibility models.
How to cite: Capolongo, D., Mancino, S., and Amatulli, G.: Beyond optimistic accuracy: effects of spatial autocorrelation and feature redundancy in large-scale landslide susceptibility models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12758, https://doi.org/10.5194/egusphere-egu26-12758, 2026.
The soil weathering rate of karst carbonate rocks is very low, resulting in scarce and thin soil layers. Under the subtropical monsoon climate, a well-developed epikarst zone is formed in the upper part of the vadose zone, with abundant fractures and pores filled with soil. The karst hillslopes surface bedrock is exposed, forming a mosaic soil landscape with soil.However, traditional profile surveys make it difficult to quantitatively determine the spatial distribution of soil and the epikarst zone with high precision; additionally, complex lithological conditions and strong spatial heterogeneity of carbonate rocks further limit the accurate quantification of soil thickness (ST) and epikarst thickness (EkT).
Therefore, this study investigated the soil-epikarst structures and their spatial distribution at different topographic locations (including different hillslope positions, ridges, saddles, and valleys) using Electrical Resistivity Tomography (ERT, 5268 sampling points) in a peak cluster-valley catchment in Southwest China. Furthermore, an interpretation method was established, where the application of revised inflexion points in 1D resistivity vertical profiles for improving ST and EkT characterization accuracy was assessed, with interpretations validated against borehole data.
Results showed that between hillslopes, the average ST ranges from 0.42 to 0.52 m, and the average EkT ranges from 3.38 to 4.58 m. The average ST in the valley (3.23 m) is significantly greater than that on hillslopes (0.49 m). Although there are some scattered, fragmented areas with EkT exceeding 20 m in the valley, both the average and median EkT in the valley (3.77 m and 2.98 m) are slightly smaller than those on hillslopes (3.93 m and 3.63 m).
This study integrated high-density ERT observations with a 1-m UAV LiDAR DEM and interpretable machine learning to predict karst soil and epikarst thickness. Important topographic controlling factors were screened out by machine learning, including those affecting ST (i.e., slope position (SP), relative elevation (RElev), slope gradient (S), slope roughness (SR), hillslope shape (HS), slope aspect (SOS), profile curvature (PrC)) and those affecting EkT (i.e., plan curvature (PLC), profile curvature (PrC), flow length (FLU), flow direction (FLD), aspect (A), relative elevation (RELE)).
Moreover, machine learning has made it possible to predict the spatial distribution of soil and the epikarst zone in the catchment with high precision, thereby providing structural information for studies such as soil erosion investigation, hydrological models, and material transport in porous media.
Keywords
Karst; Epikarst zone; Soil thickness (ST); Epikarst thickness (EkT); Electrical Resistivity Tomography (ERT); Machine learning; Southwest China
How to cite: Peng, T., Jiang, W., and Dai, B.: ERT and Interpretable Machine Learning Integrated Prediction of Karst Soil and Epikarst Thickness in a Peak Cluster-Valley Catchment, Southwest China, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15480, https://doi.org/10.5194/egusphere-egu26-15480, 2026.
Digital Elevation Models (DEMs) are a fundamental representation of terrain surfaces and are widely used in geomorphological analysis and terrain modeling. Traditional interpolation-based techniques are commonly used to address missing or unobserved regions in DEMs. However, these methods can struggle to recover coherent terrain structures when the gaps are large or irregular.
In this work, we investigate DEM completion by adapting model representations pretrained on large-scale image datasets to elevation data, which provides an effective initialization for Transformer models without requiring training from scratch. A Transformer-based architecture is employed to capture long-range spatial dependencies and global terrain structure, which are essential for reconstructing missing elevation regions. Geomorphological constraints are incorporated into the learning process to guide the reconstruction toward structurally consistent and physically plausible terrain surfaces.
Experimental results on a subset of the swisstopo DEM dataset demonstrate that the proposed approach achieves a mean absolute error (MAE) of approximately 6 m and a root mean squared error (RMSE) of approximately 11 m on the validation. Given that the average per-sample local topographic amplitude is approximately 256 m, an MAE of 6 m corresponds to less than 3% of this scale. The proposed approach leads to more coherent and structurally consistent DEM reconstructions compared to traditional interpolation-based methods, particularly in regions with missing or sparsely sampled elevation data. While the framework is developed for DEM completion, similar modeling principles could potentially be explored for other spatial surface reconstruction problems involving incomplete geospatial data.
How to cite: Lu, T., Jaboyedoff, M., Lin, R., and Wu, J.: Learning DEM Completion with Pretrained Spatial Representations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17529, https://doi.org/10.5194/egusphere-egu26-17529, 2026.
This abstract examines the role of Artificial Intelligence (AI) and Machine Learning (ML) in the geomorphological surveying of plunging cliffs, which represent some of the most dynamic and hazardous landforms on Earth. These steep and inaccessible environments are shaped by complex interactions between marine erosion, tectonics, weathering, and gravitational processes. Due to the complex logistics, traditional field-based surveying can be both risky and limited in spatial coverage. AI and ML techniques provide powerful tools to overcome these constraints by enabling the automated analysis of large, multi-source geospatial datasets, such as images, physical-chemical data, etc.
Data collected via swim surveys, drones, satellite imagery, and photogrammetry can be integrated into AI-driven workflows. Convolutional neural networks and other deep learning architectures can automatically detect coastal landforms, allowing detailed mapping of geomorphological features at unprecedented scales. Change detection algorithms applied to time-series datasets identify subtle deformation, rockfall precursors, and erosion patterns that may not be visible in the field or through manual interpretation. In parallel, ML-based classification and clustering methods help differentiate cliff characteristics, such as lithological units and surface conditions, improving the understanding of cliff geomorphic behaviour.
Overall, AI and ML can guide the transformation of plunging cliff geomorphological surveying from a largely manual and episodic practice into a continuous, high-resolution and, in the near future, predictive science. This approach not only enhances scientific insight into sea cliff dynamics but also provides practical tools for early warning systems, land-use planning, and the long-term management of these environments.
How to cite: Furlani, S.: AI and ML applications on plunging cliffs, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19667, https://doi.org/10.5194/egusphere-egu26-19667, 2026.
Effective monitoring of hydro-geomorphological processes such as sediment displacement and channel migration using UAV-captured RGB images requires accurate, temporally consistent 3D reconstructions that enable the estimation of height/volume changes across multiple acquisitions epochs, despite variations in imaging conditions, flight paths, and camera geometries. Neural Radiance Fields (NeRF) offer a powerful framework for reconstructing complex fluvial environments from multi-view imagery. However, standard NeRF training relies on equal‑probability random pixel sampling, a strategy that introduces stochasticity into the reconstruction process and undermines the temporal repeatability required for change detection. When separate NeRFs are trained for different epochs the random sampling of pixels – combined with differing camera poses – leads to inconsistent ray distributions and surface sampling, and reduced comparability of resulting reconstructions. These effects are particularly detrimental in dynamic river corridors where low-texture surfaces, water reflections and surface motions already limit reconstruction coverage and stability.
We propose Coverage-Efficient Non-redundant Sampling (CENS) as a grid-based deterministic alternative to random pixel sampling in NeRF training, combined with a pose-consistent inference framework, to improve temporal reconstruction repeatability for hydro-geomorphic change analysis. Instead of randomly sampling pixels per iteration, we impose a spatially deterministic sampling grid that ensures intra-epoch spatial coverage and consistency. To address the challenge of differing camera poses between epochs, we introduce a cross‑epoch evaluation strategy in which all epochs are queried using a shared set of reference camera poses – either drawn from one epoch’s camera set or defined as a synthetic virtual camera lattice – after alignment to a common world coordinate frame. This enables pose‑consistent rendering across time even when the original acquisition geometries differ.
Rendered depth maps from these shared viewpoints are then projected onto a fixed world-space grid, enabling cell-wise comparison of height and volume change. This two-layer determinism—fixed sampling during training and fixed camera/grid sampling during inference—decouples temporal analysis from the stochastic and pose‑dependent variability inherent in standard NeRF pipelines.
We will evaluate the method on multi‑epoch UAV imagery of an active river reach, comparing reconstructions trained with traditional random pixel sampling against those produced using the proposed CENS approach. With improved intra-epoch spatial coverage and geometric stability, combined with pose-consistent inference, we anticipate that the proposed approach will yield more coherent inter-epoch estimates of sediment elevation, channel-bed change and/or bank erosion, with lower uncertainty in low-textured or reflective regions. We expect corresponding 3D points from different epochs to align more consistently in world space, enabling more reliable detection of subtle geomorphic adjustments like bar migration, scour-fill cycles, and/or sediment redistribution.
This study aims to demonstrate that sampling strategy and inference geometry are critical, yet previously overlooked, determinants of temporal repeatability in NeRF-based hydro-geomorphological monitoring. By replacing stochastic pixel sampling with deterministic grids and enforcing shared inference poses across epochs, we introduce a simple but potentially powerful modification that may enhance the reliability of NeRF-derived height/volume change estimates in dynamic river environments. The approach could open new possibilities for long‑term environmental and hydro-geomorphological research where spatiotemporal consistency is essential.
How to cite: Akwensi, P., Schulte, F., and Winiwarter, L.: Enhancing Multi-Temporal 3D Reconstruction of River Corridor Dynamics with Deterministic NeRF Sampling Strategies, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12102, https://doi.org/10.5194/egusphere-egu26-12102, 2026.
Posters virtual: Tue, 5 May, 14:00–18:00 | vPoster spot 3
EGU26-5775 | ECS | Posters virtual | VPS26
Fully Automated Unsupervised Machine Learning Framework for Mapping Erosion Hotspots in Quick Clay Areas Using Remote Sensing–Derived DataTue, 05 May, 14:00–14:03 (CEST) vPoster spot 3
Quick clays are fine-grained, highly sensitive marine deposits that are widespread across formerly glaciated regions, including Norway, Sweden, Finland, and Canada. The low remoulded strength of the quick clays makes them particularly susceptible to extensive retrogressive landslides, which pose serious challenges to society. Erosion is recognized as one of the most important pre-conditioning and triggering factor for quick clay landslide. Therefore, identification of the erosion hotspots is essential for understanding landslide initiation processes and for effective hazard mitigation in quick clay terrains. Machine learning has emerged as an effective tool for erosion hotspot mapping, allowing complex spatial patterns and nonlinear interactions among erosion-controlling factors to be identified from remote sensing–derived data. Recent studies have demonstrated that Deep Neural Networks can be effectively employed to identify erosion-prone zones in quick clay environments when sufficient labelled data are available. This study investigates whether unsupervised machine learning applied to remote sensing–derived data can effectively identify erosion hotspots in quick clay areas. A fully automated, Python-based workflow was developed for erosion hotspot mapping in quick clay areas using remote sensing–derived data. The dataset includes terrain, hydrological, environmental, and anthropogenic parameters relevant to erosion and slope instability. Initially, a total of twenty input parameters were considered. Pearson correlation coefficients were computed to assess inter-feature dependencies, and principal component analysis (PCA) was employed to evaluate feature importance. The unsupervised analysis was performed using multiple clustering techniques to capture different structural characteristics of the data where each cluster represents a distinct level of erosion susceptibility. The results suggest that the proposed unsupervised framework can effectively delineate erosion hotspots in quick clay areas and constitutes an initial step toward the development of early warning systems.
Acknowledgements
This work was supported by the Research Council of Norway through the SAFERCLAY project (Grant No. 352887). Orkun Türe was supported by the Council of Higher Education of Türkiye under the DOSAP scholarship programme and served as a visiting researcher at NGI and NTNU.
How to cite: Türe, O., Tao, R., L’Heureux, J.-S., Oguz, E. A., and Tyagi, A.: Fully Automated Unsupervised Machine Learning Framework for Mapping Erosion Hotspots in Quick Clay Areas Using Remote Sensing–Derived Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5775, https://doi.org/10.5194/egusphere-egu26-5775, 2026.
