One of the key reasons for the growing popularity of AI in geosciences is its unparalleled ability to efficiently analyze vast datasets within remarkably short timeframes. This capability empowers scientists and researchers to tackle some of the most intricate and challenging issues in fields like Geophysics, Seismology, Hydrology, Planetary Science, Remote Sensing, and Disaster Risk Reduction.
As we stand on the cusp of a new era in geosciences, the integration of artificial intelligence promises to deliver more accurate estimations, efficient predictions, and innovative solutions. By leveraging algorithms and machine learning, AI empowers geoscientists to uncover intricate patterns and relationships within complex data sources, ultimately advancing our understanding of the Earth's dynamic systems. In essence, artificial intelligence has become an indispensable tool in the pursuit of quantitative precision and deeper insights in the fascinating world of geosciences.
For this reason, aim of this session is to explore new advances and approaches of AI in Geosciences.
Cloud cover estimation is of crucial significance in meteorological observations and short-term/long-term weather forecasting, as it directly affects the accuracy of radiation balance assessment, precipitation prediction, and climate change modeling. Ground-based automated cloud quantification observation instruments enable continuous, high-resolution cloud monitoring with spatial-temporal continuity that satellite remote sensing cannot fully achieve, highlighting the immense value of ground-based cloud image processing for practical meteorological applications. However, existing cloud detection methods predominantly rely on supervised training with ground truth masks, which overlook the rich contextual information and inherent regularization constraints embedded in original cloud images. This oversight frequently results in mismatched cloud boundaries, inadequate model interpretability, and poor adaptability to complex cloud morphologies—particularly for thin clouds and cirrus clouds characterized by weak grayscale contrast, sparse texture, and irregular shapes. Consequently, these limitations lead to suboptimal detection performance, including under-segmentation or over-segmentation, and further induce inaccuracies in quantitative cloud cover estimation.
To address the aforementioned issues and achieve accurate cloud cover detection results, this study proposes a model-agnostic refinement method designed to optimize the coarse detection masks generated by any pre-trained cloud detection model. The framework is jointly optimized by three loss functions: a local similarity descriptor, total variation (TV) regularization, and a traditional detection loss (e.g., cross-entropy). Specifically, the local similarity descriptor is defined as the difference between two terms: the average grayscale difference of each pixel and cloud region and background pixels within a local window. This descriptor effectively enhances the discriminability between cloud and non-cloud regions at the local level. The total variation regularization term is introduced to maintain the smoothness of the detection boundary and suppress spurious noise. The cross-entropy loss ensures the overall consistency between the refined result and the ground truth.
Minimizing the combined loss function drives the coarse detection result to evolve adaptively along the actual cloud boundary, thereby achieving more precise alignment with the true cloud contours. Notably, the proposed framework elevates the detection of thin clouds and cirrus clouds, effectively mitigating missed detection areas in these tenuous cloud structures. Furthermore, the integrated loss function enhances model interpretability: the local similarity descriptor explicitly quantifies the differences within local window, and minimizing this term inherently refines the detection by strengthening the distinction between cloud and background regions. Ultimately, the refined detection results substantially improve the accuracy of cloud cover estimation, laying a solid foundation for reliable meteorological observations and weather forecasting applications.
How to cite: Hu, Y., Yang, P., Zhou, Z., Bo, R., Yang, S., and Zhang, G.: Local Similarity-Driven Refinement for Model-Agnostic Ground-Based Cloud Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16953, https://doi.org/10.5194/egusphere-egu26-16953, 2026.
Land surface temperature (LST) is a key variable for assessing crop thermal stress and supporting precision agriculture. However, thermal satellite products often involve a trade-off between spatial and temporal resolution. Sentinel-3 provides frequent LST observations, but its coarse spatial resolution limits its use for field-scale agricultural monitoring.
This study proposes a spatial downscaling approach for LST based on the fusion of Sentinel-3 thermal data with high-resolution multispectral information from Sentinel-2. The method exploits the inverse relationship between surface temperature and vegetation cover through the Normalized Difference Vegetation Index (NDVI). A linear regression model was developed to estimate LST at a spatial resolution of 10 m using Sentinel-2 NDVI as the primary predictor.
The approach was applied over the agricultural site of El Ghaba in the Marrakech–Safi region (Morocco), covering different crop types, including annual cereals (barley, wheat, and kerenza) and perennial olive orchards. Results show a clear negative correlation between NDVI and LST, confirming the regulatory role of vegetation on surface temperature. The downscaled LST maps reveal fine-scale spatial heterogeneity that is not detectable in the original Sentinel-3 product.
Quantitative evaluation indicates low absolute errors for annual crops (generally below 0.5 °C), demonstrating the robustness of the proposed method, while higher discrepancies observed for olive orchards highlight the complexity of perennial crop thermal behavior. This work enhances the spatial usability of satellite thermal data for agricultural monitoring and crop stress assessment.
How to cite: Boufous, B., Ben zhair, F., and Belaqziz, S.: Spatial Downscaling of Land Surface Temperature Using Sentinel-2 and Sentinel-3 Data Fusion for Agricultural Applications, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13366, https://doi.org/10.5194/egusphere-egu26-13366, 2026.
The scarcity and high acquisition cost of field crop samples remain a major bottleneck for applying Artificial Intelligence (AI)–driven supervised learning methods in large-scale geoscientific applications such as crop type mapping. Meanwhile, crop phenology and, consequently, spectra-temporal characteristics of the same crop type present significant interannual and regional variations due to the differences in local conditions and human activities, such as climatic, soil properties and farming practices. This causes the “domain shift” challenge. Therefore, directly applying a classification model trained in a specific region and year to a new region or year inevitably leads to poor prediction performance. The gap between the abundant availability of Earth Observations imagery and the limited accessibility of training crop samples hider efficient mapping of varied crop types across large regions. To address training sample scarcity and cross-region/year domain shift in large-scale crop type mapping, we propose a transferable crop mapping method named Global-Hierarchical-Categorical feature Alignment (GHCA). GHCA integrates unsupervised domain adaptation, contrastive learning, and pseudo-labeling to achieve multi-dimensional alignment between source domain and target domain at global, hierarchical and categorical levels. The developed method enables accurate and transferable crop mapping across diverse agricultural landscapes with minimum field survey requirements. The main contributions of our study can be summarized as follows: (1) A global feature pre-alignment mechanism is introduced by calculating the Multi-Kernel Maximum Mean Discrepancy (MK-MMD) metric across different hierarchical features to align source and target domains in global and hierarchical feature spaces. This mechanism substantially improves the initial reliability of pseudo-labels generated for the target domain, providing a reliable foundation for subsequent fine-grained categorical level feature alignment; (2) A robust pseudo-label generation strategy is developed by jointly considering prediction confidence, prediction certainty, and prediction stability. Reliable pseudo-labels for target domain are selected by calculating model prediction probabilities and predictive uncertainty estimates through teacher-student model. Moreover, the Exponential Moving Average (EMA) strategy is adopted to updated model parameters in the teacher path to enable the acquisition of obtaining more stable pseudo-labels; (3) Category-wise feature alignment is achieved by integrating pseudo-labeling with contrastive learning, which explicitly pulls intra-class feature closer for the same crop types across source and target domains, while pushing inter-class feature apart for different crop types. The effectiveness of the proposed GHCA method for both cross-region and cross-year crop mapping was evaluated across five regions in China and the U.S. over a two-year timeframe. GHCA was compared with a machine learning method (RF), supervised deep learning models (DCM, Transformer, and PhenoCropNet), and transfer learning methods (DACCN, PAN, and CSTN) for cross‑year and cross‑region crop mapping. Experimental results showed that GHCA outperformed other models in most transfer cases, with OA ranging from 0.82 to 0.95 (cross-region) and 0.89 to 0.98 (cross-year), achieving an average OA increase of 6.2% and 3.5% in cross-region and cross-year experiments, respectively. These results highlight the strong potential of advanced AI methodologies to deliver robust, quantitative, and transferable solutions for complex geoscientific problems using large Earth observation datasets.
How to cite: Yang, J., Hu, Q., Belgiu, M., and Wu, W.: A global–hierarchical–categorical alignment framework to address sample scarcity and domain shift in crop mapping, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10174, https://doi.org/10.5194/egusphere-egu26-10174, 2026.
This study introduces an innovative methodology for generating realistic soil prediction maps that visualise the spatial distribution of specific chemicals, achieved through the rigorous evaluation and comparison of advanced modelling techniques, including innovative modelling techniques based on the use of neural networks and multilayer perceptrons (MLPs). The Drava River floodplain was selected as the primary case study based on stringent criteria: a) intensive historical metal ore mining and metallurgical processing activities, which have left a legacy of contamination; b) distinctive geomorphological features, such as dynamic floodplains and sediment deposition zones; and c) diverse geological settings that facilitate reliable model calibration across transboundary reaches. Soil measurements were integrated with diverse geospatial datasets—derived from Digital Elevation Models (DEMs), land cover classifications, and remote sensing imagery—to enable high-resolution mapping of contaminant distributions via sophisticated predictive modelling powered by neural networks and MLPs. A novel, holistic approach was applied to simultaneously reconstruct multiple influencing processes, including erosion, sediment transport, and pollutant dispersion, across the entire study area. This comprehensive framework not only advances contamination mapping practices but also empowers the developed models to trace primary distribution pathways, quantify the true extent of affected zones, enhance data interpretability, and inform evidence-based decisions on land-use planning, remediation strategies, and environmental management in mining-impacted regions.
How to cite: Alijagić, J. and Šajn, R.: Advanced AI Soil Mapping Techniques and Transboundary Risk Assessment for the Drava River Floodplain , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10465, https://doi.org/10.5194/egusphere-egu26-10465, 2026.
Recent advances in artificial intelligence and geospatial data analytics have led to an increasing adoption of data-driven approaches in the identification and prediction of mineral deposits. Traditional mineral exploration methods often rely on single data sources or expert-driven interpretations and may therefore be inadequate in regions where geological information is limited or spatially complex. In contrast, artificial intelligence–based approaches enable the quantitative assessment of mineral potential and the identification of spatial patterns associated with mineralization by jointly integrating multi-source geological, geophysical, and remote sensing data. Therefore, the comparative evaluation of different artificial intelligence algorithms using approaches that account for spatial dependence is critical for selecting reliable and interpretable models in early-stage mineral exploration conducted under data-limited conditions.
This study focuses on a comparative evaluation of artificial intelligence algorithms for predicting potential iron (Fe) mineralization under limited geological data conditions in a region with metallic mineralization potential in Türkiye. The study area covers approximately 2,340 km². A total of seven predictor variables were incorporated into the modeling, classified into geological (lithology, geological age, formation type), structural (fault density), geophysical (magnetic anomaly and gravity-tilt features), and remote sensing–based datasets (iron oxide potantial zones derived from ASTER imagery). The mineralization inventory is highly sparse, comprising only 15 iron occurrences and 24 non-iron reference points selected by geologists To address this limitation, a spatially aware hard negative mining strategy was applied, in which negative samples were preferentially selected from areas spatially proximal to known mineralization occurrences. Model performance was evaluated using GroupKFold-based spatial cross-validation to minimize bias arising from spatial autocorrelation, within which the Random Forest (RF) and XGBoost (XGB) algorithms were compared. The obtained results show that the RF and XGB models achieved mean Area Under Curve (AUC) values of 0.85 and 0.89, respectively. According to the generated mineral prospectivity maps, the Random Forest model delineates approximately 207.02 km² of high-potential areas (probability ≥ 0.90), while the XGBoost model identifies high-potential areas covering approximately 404.04 km² at the same probability threshold. These results indicate that there are pronounced differences in the spatial distribution of high-potential areas depending on the algorithm used. Additionally, the feature importance analysis revealed that geological age, magnetic anomaly, formation type, and gravity-tilt features are the primary controlling factors influencing the spatial distribution of iron mineralization.
This study outcomes revealed the importance of algorithm selection and spatially aware validation strategies in artificial intelligence–based mineral exploration. The findings indicate that reliable mineral prospectivity assessments can be achieved even under limited geological data conditions. Furthermore, in early-stage exploration programs, these approaches strengthen effective target area prioritization and decision-support processes and contribute to cost reduction through more efficient planning of exploration activities.
How to cite: Karakas, G., Civci, B., Topal, B., Gokceoglu, C., Ozcan, A., Erbasli, C., Cebeloglu, F. S., Koruyucu, M., and Binal, B. E.: Performance Comparison of Some Artificial Intelligence Algorithms for Metallic Mineral Deposits: A Case from Türkiye, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20311, https://doi.org/10.5194/egusphere-egu26-20311, 2026.
Geological mapping in complex metallogenic provinces often relies on band ratios and thresholding techniques. While effective for simple targets, these traditional methods struggle to capture non-linear spectral associations inherent in natural mineral mixtures and require significant prior knowledge of the target mineralogy. This study introduces a novel, data-driven unsupervised pipeline for mineral target generation, applied to the Aït Saoun region in the Moroccan Anti-Atlas, a strategic zone characterized by polymetallic occurrences (Cu, Co, Fe, Mn).
We leverage the full spectral topology of ASTER satellite imagery (VNIR-SWIR bands) rather than reduced indices. Our approach integrates topological manifold learning to reduce the high-dimensional spectral space, followed by density-based spatial clustering to delineate mineral clusters. This combination allows for the preservation of local data structure and the automated rejection of noise without human supervision.
The pipeline successfully identified spatially coherent clusters corresponding to specific hydrothermal alteration zones. It autonomously distinguished between structural iron-manganese anomalies and lithology-controlled copper mineralization a nuance often missed by standard linear ratios. The metallogenic relevance of these spectral clusters was rigorously validated through field mapping and geochemical analysis using Atomic Absorption Spectroscopy (AAS). Results confirmed economic grades in the predicted zones, yielding Copper concentrations up to 2.60% in propylitic alteration zones and Iron-Manganese oxide grades (21.94% Fe, 1.80% Mn) in tectonic corridors. Furthermore, the detection of distal barite anomalies highlights the method’s capability to map complete hydrothermal zonations.
These findings demonstrate that topological machine learning offers a robust, superior alternative to conventional remote sensing techniques for vectoring exploration targets in arid environments. By converting raw spectral data into validated metallogenic maps, this pipeline provides a scalable tool for de-risking early-stage mineral exploration in the Anti-Atlas.
How to cite: Elomairi, M. A. and El GAROUANI, A.: Automated Mineral Cluster Detection in ASTER Data Using Topological Machine Learning: A Novel Data-Driven Approach for Geological Exploration in Ait Saoun, Anti Atlas, Morocco, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4179, https://doi.org/10.5194/egusphere-egu26-4179, 2026.
Monitoring oil and gas wells is essential for assessing environmental degradation and long-term impacts such as methane emissions from abandoned and orphaned wells. Satellite imagery combined with machine learning offers scalable capabilities for detecting and characterizing oil and gas infrastructure, yet progress remains constrained by the lack of multimodal, multiple-choice (MCQ) vision-language datasets that enable structured evaluation and post-training of vision-language models (VLMs) for oil well scene grounding. Existing resources are predominantly visual-only and therefore provide limited support for image grounding from satellite imagery.
To address this gap, we introduce SatWellMCQ, a vision-language dataset of expert-verified satellite imagery paired with natural-language descriptions and multiple-choice supervision for image-grounded identification and localization of oil wells. SatWellMCQ uses high-resolution multispectral Planet imagery (RGB and infrared) and text annotations that describe well type and spatial context. Each sample includes one expert-verified correct description and three semantically plausible distractor descriptions drawn from other samples, enabling structured MCQ evaluation. All samples were manually verified by a senior domain expert with 100% intra-expert agreement, ensuring accurate alignment between images, labels, and text. The dataset covers four categories relevant to oil well monitoring: active wells, suspended wells, abandoned wells, and control samples without visible wells, yielding a balanced distribution for training and evaluation. We publicly release SatWellMCQ to support research on image grounding and vision-language adaptation in satellite imagery of oil wells.
We evaluate SatWellMCQ across state-of-the-art VLMs in zero-shot and supervised fine-tuning (SFT) settings. In the zero-shot setup, performance is moderate only for large-scale models, with the best result achieved by Qwen3-VL-235B at 0.670 accuracy. Compact models transfer poorly in zero-shot evaluation (e.g., Granite~3.3~2B at 0.422 and Phi-4-multimodal-instruct~6B at 0.376), highlighting the difficulty of domain-specific oil well analysis without targeted supervision. Supervised fine-tuning on SatWellMCQ yields substantial gains for compact models: Granite~3.3~2B improves to 0.722 and Phi-4-multimodal-instruct~6B reaches 0.730, surpassing all zero-shot baselines. These results show that SatWellMCQ poses a challenging benchmark for current VLMs while enabling effective domain adaptation through structured MCQ supervision.
Overall, SatWellMCQ provides a resource for post-training and benchmarking VLMs on image grounding of oil wells in satellite imagery and supports geoscientific monitoring tasks relevant to environmental impact assessment and methane mitigation.
How to cite: Emam, A., Alrowili, S., Eswaran, M. K., Kinzler, R., and Samih, Y.: SatWellMCQ: A Vision–Language Satellite Datasetfor MCQ-Based Image Grounding of Oil Wells, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3405, https://doi.org/10.5194/egusphere-egu26-3405, 2026.
Cracks on retaining walls and road surfaces can reveal the early warning signs of geohazards such as landslides or slumps in rural areas. However, even today, many governments still rely on manual visual inspection to identify and evaluate cracks, which is time-consuming, subjective, and highly dependent on individual experience. Artificial intelligence (AI) applied to Earth-observation imagery not only enables the detection of potentially dangerous cracks but also makes it possible to quantify their geometric properties, providing a more objective and quantitative basis for infrastructure monitoring and geohazard risk management.
Nevertheless, several key challenges remain. First, although recent studies have developed many advanced algorithms for crack detection and segmentation, methods for measuring crack width, length ,and area are still insufficient. Second, most existing models are designed for road cracks, while cracks on retaining walls present more complex textures, illumination conditions, and background noise, requiring dedicated model fine-tuning. Third, in regions with dense vegetation, branches, leaves, and shadows often produce false detections, making it difficult for AI models to distinguish real cracks from environmental interference.
In this study, we aim to quantify crack geometry from mobile panoramic Earth-observation imagery and to develop an AI model optimized for cracks on retaining walls in complex environments. A multi-stage approach is used to combine YOLO-based crack detection with 3D geospatial information for estimating the length, width, and area of individual cracks. By focusing on real cracks under vegetation-rich and noisy conditions, this approach advances AI-based quantitative analysis of surface degradation. These crack metrics provide a foundation for future retaining wall stability assessment and risk-informed infrastructure management.
How to cite: Chiang, Y.-C., Chu, S.-C., and Lin, G.-W.: AI-Based Quantification of Crack Geometry on Retaining Walls from Mobile Earth-Observation Imagery, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20838, https://doi.org/10.5194/egusphere-egu26-20838, 2026.
Earth Observation (EO) is an essential source of information for most geosciences. However, high costs, large data volumes, and difficult access constrained its use for decades. Open data programs like Copernicus have reduced costs, and cloud access via the Copernicus Data Space Ecosystem (CDSE) has made local processing largely obsolete. In fact, API (Application Programming Interface)-based cloud access, analysis-ready mosaics and calibrated Copernicus Land Monitoring Service data products have made Sentinel data AI-ready. But despite these advances, the requirement for complex programming skills remained a significant barrier until recently. Here, we demonstrate how cloud-native processing APIs and generative artificial intelligence (AI) are removing this obstacle by enabling the "vibe coding" paradigm shift. Vibe coding is an approach to software development where the researcher focuses on the high-level logic, the functional vision, and the end product, while the syntax and code are generated and refined by AI.
Copernicus Data Space Ecosystem facilitates this transition through three key features: (1) the abstraction of EO analysis pipelines via RESTful APIs, which reduces tasks to a series of mathematical operations on pixel values; (2) the availability of intuitive web browser visualization for rapid prototyping and debugging; and (3) an extensive body of open documentation and code examples that serve as a robust training foundation for generative AI.
On CDSE, the Sentinel Hub API family utilizes "custom scripts" (or "evalscripts") — modular JavaScript files defining data inputs, outputs, calculations, and visualizations. The openEO API uses "process graphs", JSON representations of the processing steps in a unified structure as a series of nodes. Because the backend manages big data optimization and the browser handles rendering, these scripts are concise enough for AI assistants to generate, adapt, and debug effectively. The Sentinel Hub Custom Script Repository, containing over 200 community-contributed scripts, and the openEO community examples repository and CDSE "Algorithm Plaza" have laid the foundation for this approach. Neither of these advances was intentionally created to support AI, but rather to simplify programming for humans; however, combined, they enable a breakthrough in code development. We demonstrate how AI tools can efficiently adapt scripts across different satellite sensors, combine spectral indices into decision trees, and produce scalable quantitative outputs. This allows researchers not specialized in remote sensing to utilize existing code modules and natural language prompts to create meaningful results for their specific fields. Beyond the capabilities of Sentinel Hub, OpenEO supports joint analysis of data from multiple back-ends and the application of user-defined external code, such as biophysical models or pre-trained ONNX deep learning networks. While this added complexity presents a higher technical threshold, it also creates a massive opportunity for AI-driven automation. Ultimately, in combination with the public data space approach, generative AI further democratizes Earth Observation, transforming it from a specialist-only domain into an integrated component of all geoscience research workflows.
How to cite: Zlinszky, A.: From natural language to quantitative satellite imagery analysis: Copernicus Data Space Ecosystem and AI enable vibe coding of custom scripts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18394, https://doi.org/10.5194/egusphere-egu26-18394, 2026.
Fin whales (Balaenoptera physalus) produce low-frequency vocalisations that propagate efficiently through the ocean and seafloor, making them detectable on broadband ocean bottom seismometers (OBS). While primarily deployed for seismic studies, OBSs offer a unique and cost-effective opportunity for passive acoustic monitoring (PAM) of marine mammals in remote regions over extended periods. Traditional detection and classification of whale calls have relied on energy thresholding, cross-correlation, or matched filtering techniques. These approaches, however, may falter in performance in high-noise environments typical of OBS datasets and often require extensive manual post-processing, making them a labour-intensive process. These limitations motivate automated, noise-robust approaches capable of exploiting the growing volume of seismic data now available.
We present a deep learning framework for detecting fin whale calls from broadband OBSs surrounding the São Jorge Island in the Azores, as well as up to twenty stations of the wider UPFLOW array spanning the Azores–Madeira–Canaries region. Our method uses a semantic segmentation model that operates on spectrogram representations between 12–35 Hz, a frequency band encompassing the classic ‘20-Hz’ fin whale note and the lower frequency ‘backbeat’. The model architecture includes a ResNet-18 encoder pretrained on ImageNet with a U-Net decoder to identify calls in both time and frequency. Training was conducted on a dataset comprising of ~6 days of manually annotated spectrograms and an additional ~6 days of background-only spectrograms. Performance was evaluated using mean Intersection-over-Union and F1-score, achieving 0.65 and 0.80 respectively.
Once validated, the model was applied to months- to year-long OBS records across the region. Fin whale calls were detected at all stations, with clear seasonal patterns showing peak calling activity between October and February, consistent with known migratory patterns in the North Atlantic. Spatial differences in call characteristics and temporal patterns further revealed potential regional variations in vocal behaviour, offering insights into song plasticity and complexity.
By applying a deep learning-based detector on OBS data, we show that machine learning provides a powerful and efficient approach to automating fin whale call detection at scale. Our method processed hundreds of thousands of hours of OBS recordings and identified nearly a million calls across all stations. This large-scale detection unlocks detailed analyses of vocal behaviour, spatial distribution, and seasonal trends, deepening our understanding of their behaviour in the north-east Atlantic. Our findings not only highlight the interdisciplinary value of OBS datasets, but also the potential of machine learning in supporting PAM efforts for the conservation and management of wide-ranging marine species.
How to cite: Japnanto, J., Saoulis, A., Romagosa, M., Leitão, R., Silva, M. A., Graham, M., and Ferreira, A. M. G.: Detecting Fin Whale Calls from Ocean-Bottom Seismometer Data with Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-304, https://doi.org/10.5194/egusphere-egu26-304, 2026.
Magnetic and gravity surveys remain among the most cost-effective geophysical tools for investigating the subsurface. They provide information on rock geometry and bulk properties at regional to deposit scale, and they have long been used to guide mineral exploration. However, turning geophysical anomalies into reliable three-dimensional property models requires inversion, a process that is inherently non-unique: multiple subsurface distributions can explain the same anomaly. Conventional approaches, such as least-squares or Bayesian inversion, can produce valuable results; however, they remain computationally demanding for large 3D models and require strong regularization choices that may bias geological interpretation.
Over the last decade, geoscientists have explored machine learning as an alternative approach. Instead of repeatedly solving forward equations, machine learning methods learn a mapping between geophysical anomalies and subsurface properties using large training libraries of synthetic examples. Early work with convolutional neural networks (CNNs) and U-Net architectures showed the concept is viable for electromagnetic and seismic data. More recent studies have shown that deep neural networks can recover magnetic susceptibility distributions from magnetic data and, in some cases, perform joint inversion of gravity and magnetic observations. Nevertheless, purely convolutional architectures often struggle to preserve long-range spatial relationships in fully three-dimensional volumes, resulting in blurred boundaries and reduced geological interpretability.
Recent advances in deep learning offer new opportunities to address these limitations. Emerging models are designed to capture long-range dependencies and preserve sharper boundaries. They have been effective in other 3D volumetric fields, such as medical imaging and seismic interpretation, but have yet to be explored for potential-field inversion.
In this study, we develop a deep-learning-based inversion method for magnetic and gravity data aimed at critical mineral exploration. The approach targets mineral systems with distinct geophysical signatures, with a focus on volcanogenic massive sulfide (VMS) environments. By combining data-driven learning with physics-informed training, the method produces reproducible three-dimensional susceptibility and density models that reduce ambiguity in subsurface interpretation. The workflow is tested using data from the Flin Flon VMS district in Manitoba, Canada, demonstrating its potential to improve targeting of buried copper-zinc mineralization and to support the integration of advanced AI methods into geoscience workflows.
How to cite: Tirdad, S., Bellefleur, G., Yrro, F., Bavand Savadkoohi, M., and Gloaguen, E.: Toward Robust Three-Dimensional Magnetic and Gravity Inversion Using Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1760, https://doi.org/10.5194/egusphere-egu26-1760, 2026.
Automated Mineralogy – the past
The automated classification of mineral phases in rocks has been a mainstay of the Geoscience analytical community for over 40 years. While we have seen great leaps forward in AI in µCT and light microscopy/petrography, the automated capabilities for the SEM have progressed and changed very little in decades, relying heavily on outdated methods that were available at the time.
The technology come with several significant problems moving forward, including excessive hardware-software dependencies, complex mineral libraries and classifications, inconsistent user experience, and difficult workflows outside their intended use.
Recent technological advances
There are two broad shifts that are taking place across a number of microscopy and microanalysis techniques – the acquisition of more quantitative data, and the application of deep learning neural networks. As a general trend this can be thought of as building better datasets, and building bigger datasets.
EDS as a SEM-based technique is fertile territory for both of these shifts. As an analytical technique EDS is commonly applied qualitatively, or as an image based method for distinguishing regions based on chemical maps. In recent years it has become easier than ever before to calibrate systems and detectors for concentration data, meaning the SEM can generate more robust datasets without having to fall back on other techniques.
Deep Learning is a topic that covers a broad range of mathematical applications to everything from the acquisition of microscopy datasets, through to data processing and interpretation across almost all sciences. There are many different flavours of deep learning neural network (DLNN) and each type lends itself to different applications, particularly in the varied data rich environments of microscopy. DLNN are inherently hard to track exactly how they operate, but at their best should be easy to use, and easy to understand how they’ve been applied to a scientific problem.
Automated Mineralogy – the future
The introduction of both quantitative mineral chemistry and DLNN to automated mineral classification is a huge leap forward, solving many of the problems of traditional software. Detaching data acquisition from processing removes software dependencies and frees users to build their ideal system. An DLNN-driven, unsupervised data processing approach can be data led rather than user led, making it more robust and consistent across instruments and facilities. Quantitative analysis can build on the DLNN approach by allowing a “best fit” classification, removing the need for constant modification of mineral libraries, and simply allowing “textbook” globally consistent mineral compositions to drive the labelling of segmented data.
How to cite: Taylor, R.: Why Automated Mineralogy needed an upgrade, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13941, https://doi.org/10.5194/egusphere-egu26-13941, 2026.
The Gross Calorific Value (GCV) indicates coal quality by measuring the total heat released during the complete combustion of the coal. Accurate GCV estimation is crucial for efficient pricing, processing, and energy performance assessment in industries. Conventional oxygen bomb calorimetry, though precise, is relatively slow and expensive for large-scale analyses. Since coal’s organic and elemental composition strongly affects its heating value, understanding this relationship can help with reliable GCV evaluation. In this study, we analyzed the mid-infrared FTIR spectra of coal and selected 56 absorption bands associated with the relevant organic and elemental constituents of coal. These were used as input features for various machine learning (ML) models to predict the GCV of coal from the Johilla coal basin in India. The ML models tested included piecewise linear regression (PLR), partial least squares regression (PLSR), support vector regression (SVR), random forest regression (RFR), artificial neural networks (ANN), and extreme gradient boosting regression (XGB). By combining the predictions from the three models (PLSR, RFR, and XGB) through a simple average, we achieved the highest accuracy (R² = 0.951, RMSE = 19.05%, MBE = 1.42%, MAE = 4.053 cal/g), indicating strong agreement between the predicted and measured values. Overall, the FTIR-based method yields results that match or surpass those of traditional laboratory techniques reported in earlier research. The GCV values predicted from the FTIR models were statistically tested using t-tests (test for mean) and F-tests (test for variance) at a 1% significance level and were found to be statistically similar to the results from the standard bomb calorimeter method. The study demonstrates that the FTIR-based approach is independent and reliable and can be used as a faster and more convenient alternative method for determining GCV, making it highly useful for quick coal quality analysis in industry.
How to cite: Vinod, A., Prasad, A. K., and Varma, A. K.: A novel method for rapid and reliable estimation of Gross Calorific Value (GCV) of Coal using mid-infrared FTIR Spectroscopy and a multi-model Machine Learning Approach, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1075, https://doi.org/10.5194/egusphere-egu26-1075, 2026.
Geochemical prospecting is a fundamental tool in mineral exploration. Traditionally, the interpretation of geochemical data has relied on classical statistical methods, which in many cases are univariate or linear in nature and may fail to adequately capture the complex multivariate relationships among geochemical parameters. In this context, machine learning approaches offer an alternative framework for the integrated analysis of multivariate data and the identification of hidden patterns.
This study evaluates the application of a Gaussian Mixture Model (GMM) as an unsupervised method for the identification of geochemical anomalies of potential geological interest. The analysis was conducted on a dataset of 114 soil samples collected from the southwestern sector of the province of Ciudad Real. Before the application of the GMM, an exploratory statistical analysis was performed, including the Kaiser–Meyer–Olkin (KMO) test and the Measure of Sampling Adequacy (MSA), aimed to assess the suitability of the variables for multivariate analysis.
After conducting several experiments, the results indicate that the Gaussian Mixture Model can identify zones with anomalous values consistent with geological interest, highlighting its potential as a supportive tool in geochemical prospecting.
How to cite: Jaeger, J., Barquero, J. I., López-Gómez, J. A., and Higueras, P.: Application of Gaussian Mixture Models for Geochemical Anomaly Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5687, https://doi.org/10.5194/egusphere-egu26-5687, 2026.
The increasing volume and complexity of geochemical literature pose major challenges for the sustainable curation of domain-specific databases such as GEOROC (Geochemistry of Rocks of the Oceans and Continents), the world’s largest repository of geochemical and isotopic data from igneous and metamorphic rocks and minerals, aggregating more than 41 million values from over 23,000 publications. Although GEOROC underpins a wide range of geoscientific research, the extraction and harmonization of metadata from publications still relies heavily on manual effort, which significantly limits the scalability.
In this contribution, we present a novel information extraction architecture that moves beyond linear processing pipelines toward an Large Language Model (LLM)-based multi-agent system combining document layout analysis, schema-driven reasoning, and modality-aware extraction. Unlike generic LLM approaches that treat documents as continuous text streams, our architecture adopts a "Visual-First" strategy. We utilize a layout-aware backbone (MinerU, Niu et al., 2025) to decompose PDF manuscripts into a sequence of geometrically grounded primitive blocks, each representing a localized document region with associated visual and typographic features, preserving the geometric grounding essential for interpreting complex data tables. A routing agent subsequently validates and refines the initial layout classification, dynamically dispatching blocks to specialized downstream agents for text, table, or figure processing. This adaptive routing strategy improves robustness against layout variability across journals, publication years, and formatting styles.
Central to the framework is an active schema agent that operationalizes the GEOROC metadata model. Rather than treating the database schema as a static template, this agent continuously provides extraction targets, normalization rules, unit standards, and conflict-resolution policies that guide all subsequent processing steps. Text blocks are handled by an Optical Character Recognition (OCR) driven information extraction agent, table blocks by a table parsing agent capable of reconstructing complex table structures, and figure blocks by a visual reasoning agent designed to interpret diagrams and digitize plotted values. Each agent produces structured candidate values enriched with confidence estimates and fine-grained provenance, including page-level and bounding-box references to the original document.
The outputs of these modality-specific agents are consolidated by a merge-and-judge agent, which goes beyond simple aggregation. This agent performs cross-modal arbitration, unit harmonization, and deduplication, resolving conflicts between heterogeneous sources according to schema-defined priorities and data-quality criteria. The final result is a machine-readable JSON representation that preserves both extracted values and their evidential context.
By combining layout grounding, adaptive routing, schema-driven reasoning, and judgment-based integration, this system delivers a robust and extensible approach to large-scale metadata extraction. The framework substantially supports the curation process and strengthens GEOROC’s role as a FAIR-compliant reference infrastructure by enabling more efficient reuse of published geochemical data in future geochemical research.
References:
Niu, J., Liu, Z., Gu, Z., Wang, B., Ouyang, L., Zhao, Z., ... & He, C. (2025). Mineru2. 5: A decoupled vision-language model for efficient high-resolution document parsing. arXiv preprint arXiv:2509.22186.
How to cite: Yang, T., Elezabawy, K., Kurzawe, D., Kallas, L., Traun, M., Sarbas, B., Sturm, A., Möller-McNett, S., Willbold, M., and Wörner, G.: Multi-agent Geochemical Literature Data Mining System, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14415, https://doi.org/10.5194/egusphere-egu26-14415, 2026.
Around the globe we experience a significant biodiversity loss, mainly driven by direct anthropogenic exploitation, land use changes, and climate change. The most effective strategy to limit biodiversity loss is the designation and management of protected areas. Consequently, the European Union has adopted the EU Biodiversity Strategy for 2030, aiming to protect 30% of aquatic and terrestrial ecosystems by 2030. However, a consistent framework to designate protected areas across all EU member states is lacking. Additionally, the monitoring of biodiversity is challenged by the dynamic nature of the biological system, exacerbated by ongoing climate change, putting additional pressure on the member states in the identification of suitable areas for conservation.
In contrast, the increasing amount of detailed geospatial and climatic data contains valuable information that can be used to optimise protected area designation. Recent developments in artificial intelligence and machine learning now provide us with powerful tools to best utilise these vast amounts of data. In this study, we develop a transparent and reproducible framework to prioritise protected areas in forests. Here we apply the CAPTAIN framework based on reinforcement learning (RL) to identify valuable forest habitats for conservation in the federal state of North Rhine-Westphalia (NRW), Germany. First, we model habitats of ten forest bird indicator species across the period of 2016-2024. Second, we use the changing habitat patterns to train a RL model that identifies 30% of the most valuable forest sites in the federal state. Finally, we model valuable forest sites under different policies (e.g., including or excluding opportunity costs for nature conservation) to illustrate how potential limitations of nature conservation management can be addressed. Our results indicate that forest sites in the south-east of NRW are most suitable for conservation. Furthermore, we find that including opportunity costs for nature conservation in the model predictions produces similarly strong outcomes for safeguarding the most endangered bird species. The framework makes use of open-source data and can be applied to any other region or country to support strategic nature conservation management.
How to cite: Horn, K., Silvestro, D., Wallis, C., Leitao, P. J., Daldaban, E., and Rudolph, A.: Identifying valuable forest habitats for conservation in north-western Germany using AI and citizen science, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4011, https://doi.org/10.5194/egusphere-egu26-4011, 2026.
As similar disasters and accidents continue to occur, public concern about the limitations of existing disaster response systems and the need for institutional improvement is increasing. The National Disaster Management Research Institute of Korea conducts disaster cause investigations as part of its statutory responsibilities, examining problems observed before and after disasters, institutional weaknesses, and public demands for improvement. In this context, news data provide valuable unstructured information that reflects on-site conditions, response activities, policy debates, and public opinion, and thus complement official investigation records in understanding institutional and managerial factors related to disasters.
This study aims to develop a media analysis framework based on big data and text mining for use in disaster cause investigations. Disaster-related news articles were first collected, and a large language model (Gemini) was applied to identify and extract sentences that describe problems and suggested improvements in the stages of disaster occurrence and response. The extracted sentences were then processed using natural language processing techniques, including stopword removal and the merging of duplicate and semantically similar sentences. Based on semantic similarity, the remaining sentences were grouped to organize major issues. In addition, nouns were extracted and their frequencies were analyzed by year to identify key terms and to examine changes in topics emphasized in media coverage.
Applying the proposed framework to the disaster cause investigation of the 2023 Osong Underpass Flooding Disaster conducted in 2025, we identified 21 problem items grouped into seven categories, such as insufficient pre-closure of the underpass and inadequate maintenance of river embankments. In addition, 17 improvement measures were derived in six categories, including improvements to underpass closure criteria and flood risk grading, as well as the strengthening of river management practices, and were systematically organized and proposed. The results indicate that combining news big data, text mining, and large language models can effectively structure key issues and institutional weaknesses, and can serve as a useful analytical tool for strengthening the evidence base and explanatory power of disaster cause investigations.
How to cite: Kim, J. E., Shin, H., and Choi, S.: A Big Data and Text Mining–Based Media Analysis Framework for Disaster Cause Investigation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8542, https://doi.org/10.5194/egusphere-egu26-8542, 2026.
Wildfires are increasingly reshaping landscapes across the U.S., disrupting hydrogeologic processes such as runoff, infiltration, and sediment transport—posing major challenges for streamflow prediction and water resource management. Traditional conceptual and physically based hydrologic models often struggle to capture these disturbance-driven dynamics. In this study, we explore the potential of long short-term memory (LSTM) networks, a type of recurrent neural network, to simulate post-fire streamflow across 1,082 fire-affected basins spanning the contiguous U.S.—representing the first near-continental-scale application of LSTMs for wildfire-related hydrologic prediction.
Three LSTM models were trained on different temporal splits of fifteen-year datasets containing wildfire events: one using pre-fire data, one using post-fire data, and one using the full dataset. Models were evaluated on unseen basins in both pre- and post-fire windows. Results show that the model trained on the full dataset consistently outperformed the others, underscoring the importance of temporally diverse training data that include disturbance events. Importantly, LSTMs demonstrated strong generalization across disturbed and undisturbed environments, highlighting their ability to learn hydrologic patterns beyond the constraints of traditional process-based modeling frameworks.
Feature importance analysis revealed that topographic variables (e.g., elevation and slope) were most influential, followed by soil/geologic and vegetation characteristics, while fire-specific indicators (e.g., burn severity) ranked surprisingly low. This suggests that the LSTMs internalized key controls on streamflow response without heavy reliance on the explicit disturbance metrics included. To further isolate the model’s learned response to wildfire, simulations were performed with synthetic unburned conditions for each disturbed basin and compared against burned scenarios. Spatial analysis by EPA Level II ecoregion revealed that in the Southeastern U.S., Ozark/Appalachian Forests, and Mediterranean California, the model identified a persistent, multi-year increase in streamflow-lasting up to three years after wildfire. These regions share ecological characteristics such as high vegetation biomass, seasonal climate regimes, and terrain-driven hydrologic gradients that collectively amplify post-fire reductions in evapotranspiration and enhance runoff generation. In contrast, no significant streamflow change was detected in the Western Cordillera, South Central Prairies or Cold Desert ecoregions, where water-limited climates and lower fuel loads results in a dual-action response of hydrologic buffering and constrained post-fire increases in water yield.
Together, these findings demonstrate that LSTMs can detect regionally coherent hydrologic responses to wildfire even in the absence of strong dependence on explicit disturbance features, highlighting the promise of AI-driven, data-centric approaches for modeling hydrologic change in an era of increasing disturbances. As wildfires and other extreme events become more frequent, integrating machine learning into hydrologic prediction frameworks offers a powerful pathway toward adaptive water resource management and improved resilience across diverse ecohydrologic settings.
How to cite: Hogue, T., Moon, C., and Corona, C.: Quantifying Post‑Wildfire Hydrologic Response Using LSTMs: Ecoregion Patterns Across the Contiguous United States, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15295, https://doi.org/10.5194/egusphere-egu26-15295, 2026.
Earth system is characterized by intricate interactions between human activities and natural processes, where stochastic dynamics, nonlinear feedbacks, and emergent behaviors collectively determine system evolution and sustainability outcomes. Despite significant advances in Earth system science, two fundamental challenges persist: the insufficient integration of physical process models with observational data, and the lack of interpretable frameworks for simulating coupled human-Earth dynamics and optimizing governance strategies. These limitations critically impede our ability to conduct effective Earth system governance and guide human-environment interactions toward sustainable development pathways. To overcome these challenges, this study proposes an innovative framework that synergistically integrates data assimilation and reinforcement learning to enhance both predictability and decision-making capabilities in the complex Earth system. Data assimilation, as a well-established methodology in Earth system science, systematically combines dynamic models with multi-source observations to improve system observability and forecast accuracy. Reinforcement learning, grounded in the Bellman equation and Markov decision processes, provides a natural paradigm for modeling adaptive human-environment interactions and deriving optimal strategies through sequential decision-making under uncertainty. Building upon these complementary methodologies, we develop a Multi-Agent Deep Reinforcement Learning (MADRL) framework that employs the Markov decision process as the theoretical foundation, integrates agent-based modeling to represent heterogeneous stakeholder behaviors across multiple organizational levels, utilizes deep neural networks to handle high-dimensional state-action spaces, and incorporates data assimilation techniques to continuously update system states and reduce forecast uncertainties. This integrated framework is specifically designed to address fundamental Earth system governance challenges by capturing emergent phenomena arising from complex human-environment interactions, enabling the exploration of intervention mechanisms such as economic incentives, regulatory policies, and cooperative arrangements, and providing interpretable decision pathways that balance economic development with environmental sustainability. Through this integration, our framework offers a systematic approach to tackle classical problems in Earth system governance, from the tragedy of the commons to planetary boundaries, ultimately advancing our capacity to navigate toward sustainable development trajectories in an increasingly coupled human-Earth system.
How to cite: Yuan, S. and Li, X.: Generalizing human-Earth systems modeling and decision-making: A multi-agent deep reinforcement learning framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19409, https://doi.org/10.5194/egusphere-egu26-19409, 2026.
This presentation outlines a comprehensive framework of multi-scale digital solutions designed to address Africa's pressing water challenges. We explore the integration of advanced physical modelling with a diverse suite of next-generation hydrologic observations from remote sensing and in-situ networks to crowd-sourced data. The core of our approach lies in automated systems for data fusion, processing, and assimilation, leveraging machine learning and hybrid techniques to enhance model accuracy. Critically, we incorporate robust uncertainty quantification to ensure reliable outputs. These integrated components enable the development of actionable, real-time forecasting and decision support systems for water resources allocation and disaster management. We will demonstrate practical applications, including autonomous processes and embedded devices, showcasing a transformative pathway towards proactive, data-driven water governance across the African continent.
How to cite: Berchie Morfo, S. and Bamfo, N. K. O.: An Integrated Digital Framework for Multi-Scale Water Security in Africa., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-716, https://doi.org/10.5194/egusphere-egu26-716, 2026.
Posters on site: Fri, 8 May, 14:00–15:45 | Hall X4
Defining subsurface soil conditions in complex coastal settings requires the use of both geophysical and geotechnical datasets, each with different resolution and sensitivity. This study combined helicopter-borne electromagnetic (HEM) data, where large areas are spatially covered with limitations to vertical resolution, with cone penetration test (CPT) data, where high resolution can be achieved while the spatial resolution often is very sparse due to drilling associated costs. Τo formulate a continuous three-dimensional model of subsurface soil properties for levee risk assessment, these datasets were integrated. HEM data provides extensive covering resistivity profiles, while CPT provides high resolution, spatially limited measurements of mechanical soil behaviour.
It is known that resistivity as a soil property depends on many parameters (mostly water quality and soil type), and there is no straightforward method to directly translate it to soil, hence the use of ML. To deal with these complexities, we employed machine learning methods – Random Forests and neural networks – to merge heterogeneous datasets and predict continuous soil behaviour indices and discrete lithological types. We propose the use of multiple features, such as spatial coordinates, depths, distance from coast, soil types and local geological conditions. After pre-processing, machine-learning models were trained to fuse the datasets to ensure spatial consistency in the coastal environment. Afterwards, the Soil Behaviour Type Index (SBT) (Robertson, 1990) was calculated using the CPT measurements and then was discretized into lithological units.
A classical machine learning algorithm (Random Forest) and a PyTorch-based neural network were trained for regression (predicting the continuous SBT index) and classification (predicting soil types) tasks, and their performance was evaluated using standard statistical and visual metrics. Final models were retrained on the full dataset to increase generalizability and robustness. The final product is to map 𝐼𝑐 values and lithological classes at every HEM point and ultimately to make a 3D subsurface soil model. The outcome for each process was validated against an 80%-20% test to ensure reasonable results.
While regression models had similar RMSE scores, classification models generally produced models with greater accuracy of dominant soil types but captured fewer underrepresented mixed lithologies. This work focuses on the interpretability of soil models through integrating data (i.e., not just purely statistical but spatial output) and ultimately continuity in the spatial domain (where engineers are most concerned). The goal of this study is to develop a framework where continuous geophysical data, collected either by helicopters or drones can be combined with additional geological boreholes and CPTs and other geotechnical information, to enable us to image the subsurface beyond resistivity. One of the products of this study serves to represent an approach to providing a better product to those grappling with levee design and safety.
How to cite: Madelis, D., Karaoulis, M., and De Smedt, P.: Electromagnetic & Cone Penetration Test Data Fusion on Soil Characterization, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4956, https://doi.org/10.5194/egusphere-egu26-4956, 2026.
Fractional Vegetation Cover (FVC) is a key ecological variable for monitoring ecosystem health, land degradation, and vegetation dynamics in dryland environments. While satellite and UAV observations enable scalable FVC estimation over large spatial extents, the accuracy and robustness of these models remain strongly dependent on high-quality field-based reference data for calibration and validation. Traditional in-situ methods, including visual estimates using transect-based surveys, remain widely used but are labor-intensive and inherently subjective. Digital photography has emerged as a practical alternative, typically analyzed using index-based computer vision techniques or deep learning models. However, these methods are highly sensitive to background variability and therefore rely on massive labeled datasets. Recent advances in multimodal large language models (MLLMs) suggest a potential paradigm shift, as these models combine visual perception with high-level reasoning and benefit from diverse pre-training that enables conceptual knowledge transfer across tasks. In this study, we evaluate the feasibility of using MLLMs for direct estimation of FVC from ground-level photographs without task-specific training. We collected and compiled a dataset of more than 1,100 quadrat pictures from across 26 dryland sites in Saudi Arabia, spanning a wide range of surface conditions from bare soil to sparsely vegetated rangelands. Each picture corresponded to a 1 m × 1 m quadrat with FVC estimated independently by two experts, whose average was used as reference data for assessment of model predictions. Six state-of-the-art multimodal large language models, including Qwen2.5-VL, Mistral-Small-3.2, LLaMA-4-Maverick, LLaMA-4-Scout, and two Gemma-3 variants, were evaluated using four prompt designs that varied in length, ecological context, and methodological detail. Across all models and prompts, MLLMs achieved a mean absolute error of approximately 7.8%, demonstrating competitive performance relative to traditional image-based methods. The best-performing model-prompt combinations achieved mean absolute error values below 5%, with low systematic bias. Short and ecologically explicit prompts consistently outperformed more complex prompt designs, achieving an average reduction in mean absolute error (MAE) of approximately 1.3–1.4 percentage points compared to visually guided or highly structured prompts (MAE ≈ 6.9% versus 8.2–8.4%). Overall performance was more sensitive to model choice than to prompt structure, with mean MAE varying from approximately 5.6% to 10.0% across models, compared to a narrower range across prompts. The highest accuracy was obtained using the Qwen2.5-VL model with an ecologically detailed prompt, which achieved a mean absolute error of 4.9%, near-zero bias, and an RMSE of 8.4%. Across all prompt designs, Qwen2.5-VL and Mistral-Small-3.2 consistently delivered the best overall performance, both maintaining mean MAE values below 6% and exhibiting stable behavior across prompt variations, indicating robustness to prompt design. These results demonstrate that MLLMs can provide accurate and scalable FVC estimates directly from field photographs, without requiring specialized training datasets. This approach offers a promising alternative for rapid field surveys and reference data generation, particularly in dryland ecosystems where background complexity and data scarcity limit the effectiveness of conventional methods.
How to cite: Lopez Camargo, O. A., Elias Lara, M., El Hajj, M., Cheng, H., Scilla, D., Angulo, V., Al wahas, A., Johansen, K., and McCabe, M. F.: Evaluating Fractional Vegetation Cover using Multimodal Large Language Models: A Comparative study with Human Observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18939, https://doi.org/10.5194/egusphere-egu26-18939, 2026.
Quartz grain surface microtextures observed by scanning electron microscopy (SEM) provide important information on sediment transport history, depositional processes and sediment provenance. Traditionally, the interpretation of these features has relied upon qualitative visual assessment—an approach deeply rooted in expert judgement and cumulative experience. While fundamental, this methodology is inherently susceptible to subjectivity and inter-analyst variability. To counter balance this problem, we explore image-based classification approaches (utilizing Deep Learning frameworks) as a tool to support quartz microtextural analysis and assist in the identification of likely depositional environments thus establishing sediment provenance relationships.
A dataset of 3 367 SEM images was compiled, spanning a diverse range of sedimentary contexts: aeolian dunes, beach faces’, alluvial systems, basal sands, and nearshore, alongside with high-energy deposits from storm and tsunami events. Based on this dataset, five classification models were developed. Three were designed to discriminate between the full set of seven depositional classes, while two focused on a reduced classification scheme comprising four classes (alluvial, beach, dune and nearshore). All models were optimised using an increasing number of training epochs to assess the stability and evolution of classification performance. The results obtained were further examined in comparison with SandAI, an existing tool for microtexture classification, to evaluate its behaviour when applied to new sedimentary contexts and datasets acquired under different conditions.
The most consistent classification results were obtained for environments characterised by well-preserved and distinctive mechanical microtextures (e.g. aeolian sediments). Conversely, while environments defined by overlapping processes occasionally yielded higher nominal accuracies in QzTexNet (CNN-based models developed within the scope of this work), this is potentially attributed to their over-representation in the dataset. Analysis of classification outcomes indicates that microtextural overprinting, dataset imbalance and variations in image quality reduced the visibility of diagnostic features, thereby complicating the differentiation of depositional settings. Nevertheless, the data suggests that our models successfully capture sedimentologically meaningful patterns when surface textures remain clear. While SandAI showed stable performance within its original scope, its accuracy was limited, peaking at 47% for its target environments and dropping significantly when faced with complex deposits like tsunami or nearshore grains. In contrast, the newly developed QzTexNet models showed slightly more encouraging results, reaching accuracies of around 55% and demonstrating a steady improvement through successive refinements.
Ultimately, these findings demonstrate that automated classification offers a powerful complement to traditional analysis, particularly in ensuring reproducibility across large-scale datasets. Solely based on our database, it was observed that challenges regarding dataset equilibrium and textural complexity persist, targeted methodological refinements and supervised training hold significant potential. Such advancements represent a promising frontier in sedimentary provenance studies, particularly for the rigorous identification of deposits linked to extreme geological events.
This work is supported by FCT, I.P./MCTES through national funds (PIDDAC): LA/P/0068/2020 (https://doi.org/10.54499/LA/P/0068/2020), UID/50019/2025(https://doi.org /10.54499/UID/PRR/50019/2025), UID/PRR2/50019/2025). Finally this work is a contribution to project iCoast (project 14796 COMPETE2030-FEDER-00930000).
How to cite: Marques, N., Costa, P., and Pina, P.: Quartz grain microtexture analysis using Artificial Intelligence: application to tsunami and storm deposits provenance studies, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14874, https://doi.org/10.5194/egusphere-egu26-14874, 2026.
- Zircon trace element geochemistry is a pivotal tool for unraveling petrogenesis and the evolutionary history of the Earth’s crust. While two-dimensional (2D) discriminant diagrams are conventionally used to identify parent rock types, the emergence of machine learning (ML) has introduced a transformative research paradigm. ML not only enhances classification accuracy but also resolves the inherent ambiguities found in traditional geochemical diagrams. However, the reliability of current ML models typically depends on the vast archives of labeled samples from the Phanerozoic. When extending research to “deep-time” samples, such as Hadean zircons, the scarcity of labeled data often forces researchers to rely on models trained exclusively on Phanerozoic datasets. This approach is prone to misclassification due to “domain shift,” caused by systematic variations in zircon trace element distributions across different geological eons. To address this challenge, we propose a Domain Adversarial Neural Network (DANN) framework tailored for zircon trace element analysis. By aligning the feature distributions of the source domain (Phanerozoic) and the target domain (Precambrian), the DANN extracts “domain-invariant yet geologically significant” high-dimensional feature representations, effectively mitigating the effects of temporal data bias. Our results demonstrate that DANN significantly outperforms traditional machine learning methods across multiple performance metrics. Furthermore, t-SNE visualization confirms that the source and target domains are effectively aligned within the feature space. When applied to ~4.3 Ga zircon samples from the Jack Hills, the model achieved a classification accuracy of 0.923. This high level of performance underscores the framework’s exceptional generalization capability for identifying unlabeled deep-time samples and its potential for broader applications in Precambrian geology. This study develops a transferable, data‑driven framework for inferring deep‑time geological processes, providing a novel methodology to address the limitations inherent in the traditional principle of uniformitarianism. Furthermore, the framework is extensible to other mineral systems (e.g., apatite, monazite), thereby opening new avenues for quantitatively reconstructing the dynamic evolution of the early Earth.
How to cite: Zhang, M., Chen, G., Kusky, T., Harrison, M., Cheng, Q., and Wang, L.: Identifying zircon provenances using domain-adversarial neural network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12550, https://doi.org/10.5194/egusphere-egu26-12550, 2026.
This study presents a comparative analysis of time-series forecasting models to predict caisson tilt using early-stage monitoring data. To establish a training dataset that accounts for inherent geotechnical uncertainty, 1,000 2D numerical simulations were performed using PLAXIS2D, based on an actual design case in South Korea. To incorporate spatial variability, the subsurface was discretized into 61 independent zones: Deep Cement Mixing (33 zones), foundation rubble (6 zones), backfill rubble (10 zones), and underlying heaving soil (12 zones). Geotechnical parameters including elastic modulus (E), undrained shear strength (Su), and interface strength reduction factor (Rinter), were varied by up to 50% of their design values. Latin Hypercube Sampling (LHS) was used to assign geotechnical properties to each zone. Each case simulated a 28-stage construction sequence, with caisson tilt extracted at each stage to generate time-series data.
Four forecasting models such as ARIMA, LSTM, Temporal Convolutional Network (TCN), and an encoder-only Transformer, were evaluated. The dataset was split into 680 simulations for training, 170 for validation, and 150 for testing. Forecasting performance was assessed across varying initial observation lengths (cut = 3, 5, 10, 15, and 20 stages) to predict all remaining future stages. Results indicate that while the statistical baseline (ARIMA) showed consistently high errors regardless of observation length, with RMSE values of approximately 0.09 at cut = 3 and 0.08 at cut = 10. In contrast, deep learning models exhibited clear error reductions as more initial observations became available. Among the tested models, the TCN achieved the highest accuracy, with RMSE values of approximately 0.006 at cut = 10 and 0.004 at cut = 15. The encoder-only Transformer model also maintained stable performance for cut ≥ 10, with RMSE values below 0.01.
Acknowledgements This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2023R1A2C1007635).
How to cite: Kim, S., Hong, J., Huh, I., and Youn, H.: Forecasting Offshore Caisson Tilt via Deep Learning: A Numerical Simulation-Based Approach Accounting for Geotechnical Uncertainty, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6107, https://doi.org/10.5194/egusphere-egu26-6107, 2026.
Bayesian inversion provides a rigorous framework for uncertainty quantification in geophysics， but is often computationally prohibitive due to the reliance on Markov Chain Monte Carlo (MCMC) sampling, which requires massive numbers of forward simulations. While deep learning surrogate models offer acceleration, existing architectures (e.g., CNNs, FNO, DeepONet) often struggle with fixed discretization constraints and cannot flexibly handle the irregular observation coordinates typical in field surveys.
To address these challenges, we propose the General Geophysical Neural Operator (GGNO), a novel Transformer-based architecture designed for mesh-independent operator learning. This design fulfills three fundamental requirements for forward solvers in the context of practical inversion: (1) Discretization-invariant, allowing the processing of input models with different mesh resolutions; (2) Prediction-free, enabling direct solution querying at arbitrary spatio-temporal coordinates; and (3) Domain-independent, decoupling input and output discretizations.
We validate GGNO on Magnetotelluric (MT) forward modeling, demonstrating exceptional generalization while achieving accuracy two orders of magnitude higher than traditional methods. By integrating GGNO into a Bayesian framework, we achieve highly efficient MCMC sampling, reducing the computational time from tens of days to a few minutes, which allows for a comprehensive exploration of the posterior distribution. Applied to field data, this approach successfully recovers complex subsurface resistivity structures with rigorous uncertainty bounds. These results highlight GGNO's potential to enable high-precision subsurface imaging and robust probabilistic interpretation for complex geophysical exploration.
How to cite: zhang, H. and xu, Y.: Rapid Bayesian Geophysical Inversion Using General Geophysical Neural Operator, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6766, https://doi.org/10.5194/egusphere-egu26-6766, 2026.
Satellite Image Time Series (SITS) are a cornerstone of Earth observation, enabling long-term monitoring of environmental processes such as vegetation dynamics, land-use change, and natural hazards. However, optical satellite time series, including Sentinel-2, are frequently irregular and incomplete due to cloud cover, atmospheric effects, and acquisition constraints, which strongly limit their usability in operational monitoring systems. In contrast, Sentinel-1 Synthetic Aperture Radar (SAR) provides regular observations for any weather condition and offers complementary information for mitigating optical sensor limitations. Generating dense and reliable Sentinel-2 time series from multi-sensor observations therefore remains a critical challenge.
This work investigates Gaussian Process (GP) based statistical models for the reconstruction and densification of Sentinel-2 image time series by jointly exploiting Sentinel-1 and Sentinel-2 data. Gaussian Processes offer a flexible Bayesian framework for pixel interpolation and extrapolation. We explore GP formulations capable of handling irregular temporal sampling, multi-output dependencies, and latent variable structures, enabling the fusion of heterogeneous optical and radar observations.
An in-depth analysis of the state-of-the-art is conducted, covering multi-output Gaussian Processes, sparse and variational approximations for scalability, latent variable models (including hierarchical GP-LVMs), and inverse GP approaches based on shared latent spaces. These methods are evaluated with respect to three key challenges: ensuring spatio-temporal coherence of reconstructed images, fusing asynchronous multi-sensor observations, and maintaining computational tractability for large-scale satellite datasets.
To support experimental investigations, a representative multi-regional dataset is constructed over mainland France and overseas territories, capturing diverse climatic patterns, land-cover types, and cloud conditions, including extreme events such as flooding.
This study establishes the methodological foundations for reconstructing dense Sentinel-2 time series conditioned on Sentinel-1 observations, with explicit uncertainty quantification. By leveraging Sentinel-1 data, the approach effectively imputes missing Sentinel-2 values while providing consistent average pixel estimates with associated uncertainty, which is critical for geoscience applications. The proposed framework contributes toward more robust Earth observation monitoring systems and the development of reliable geospatial digital twins.
How to cite: Nespoulous, B., Constantin, A., Derksen, D., and Defonte, V.: Probabilistic Reconstruction of Sentinel-2 Satellite Image Time Series Using Multi-Sensor Gaussian Process Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7774, https://doi.org/10.5194/egusphere-egu26-7774, 2026.
Sentinel-2 optical image time series are a key source of information for many Earth observation applications, including climate monitoring, agriculture, ecosystem dynamics, and land surface change analysis. Dense and regular observations are essential to accurately capture seasonal patterns, abrupt events, and long-term trends. However, in practice, Sentinel-2 time series are often sparse and irregular due to cloud cover and varying acquisition conditions. These limitations significantly complicate continuous monitoring and the analysis of surface dynamics. Moreover, beyond time series densification, there is a growing need to anticipate future optical observations to support scenario analysis, early warning systems, and predictive environmental monitoring.
To address these challenges, we propose a deep learning–based framework for densifying Sentinel-2 time series by generating plausible optical images at arbitrary past or future dates. The approach relies on multimodal satellite observations, jointly exploiting optical Sentinel-2 and radar Sentinel-1 data. Indeed, SAR measurements are insensitive to cloud cover and provide complementary structural and temporal information. This multimodal setting enables the reconstruction of missing observations and the prediction of future optical states while preserving realistic spatio-temporal dynamics.
From a methodological perspective, the model is explicitly designed to handle sparse, incomplete, and temporally misaligned multimodal time series. It operates on temporal sets of Sentinel-2 and Sentinel-1 images acquired at irregular dates around a target time. A cross-attention mechanism is used to explicitly model interactions across time and modalities, allowing the network to identify and weight the most relevant observations for generating a Sentinel-2 image at a given target date.
In addition, the proposed framework incorporates a probabilistic decoder that estimates not only the predicted Sentinel-2 image but also an associated uncertainty map. This uncertainty estimation provides valuable insight into the confidence of the generated pixels, which is particularly important for downstream applications such as anomaly detection, risk assessment, and decision-making support.
The model is evaluated across multiple geographical regions and land-cover types, demonstrating strong performance in both densification and forecasting tasks. Results show that the proposed approach successfully preserves the temporal dynamics of the scenes, notably by accurately reproducing vegetation phenology as reflected in NDVI time series. Forecasting experiments further highlight the importance of radar information: Sentinel-1 observations close to the target date allow the model to detect surface changes occurring after the last available optical image, thereby improving future predictions. Overall, the proposed method represents a step towards the densification and forecasting of Sentinel-2 time series, offering a promising direction for future methodologies aimed at continuous Earth surface monitoring and predictive analysis.
How to cite: Defonte, V., Derksen, D., Constantin, A., and Nespoulous, B.: Densification and forecasting of Sentinel-2 time series from multimodal SAR and optical satellite data using deep generative models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7860, https://doi.org/10.5194/egusphere-egu26-7860, 2026.
Autonomous spacecraft operations are increasingly important as missions grow more complex, ground contact opportunities remain limited, and the number of LEO satellites continue to rise. Reliable onboard orbit determination (OD) and orbit prediction (OP) are essential for mission planning, resource allocation, and communication scheduling. Operational OD/OP typically relies on physics-based models that estimate parameters (initial state, drag coefficient, etc.) from tracking data. However, environmental modeling is not perfect, and uncertainties in atmospheric density can cause prediction errors to grow rapidly. This limits OP reliability.
We present an onboard-oriented hybrid OD/OP concept that augments a classical physics-based OD/OP chain with a lightweight machine-learning (ML) correction module to compensate for systematic OP errors in real time. While data-driven correction of propagator errors has been explored previously, this work emphasizes the tight integration of a compact correction model into an operational workflow under onboard constraints. The implementation is based on the Python OD/OP toolbox Artificial Intelligence for Precise Orbit Determination (AI4POD) and targets deployment within the Autonomous Space Operations Planner and Scheduler (ASOPS) experiment, that is planned for validation on the ATHENE-1 satellite.
The approach is demonstrated using simulated GPS-like tracking data generated with a high-fidelity reference model, while OD/OP are performed with a reduced-complexity model representative of onboard settings. A compact artificial neural network (ANN) is trained to predict OP errors in the RSW frame from available onboard data, reducing the maximum three-day along-track error from ~5 km to ~1.2 km.
To assess operational robustness, we complement the baseline results with a statistical consistency check of the residuals across all prediction cases and outline planned tests with additional ML/DL correction models.
How to cite: Aigner, B., Dallinger, F., Andert, T., and Haser, B.: Onboard Hybrid Orbit Prediction with Lightweight Machine-Learning Error Correction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19439, https://doi.org/10.5194/egusphere-egu26-19439, 2026.
Orbit Determination (OD) is commonly addressed with classical estimators such as Weighted Least Squares, which are statistically well founded but can be sensitive to poor initialization and may degrade when the initial state is weakly known. Physics-Informed Machine Learning offers an alternative by embedding orbital dynamics directly into the estimation process. In this work, Physics-Informed Extreme Learning Machines (PIELMs) are investigated as fast OD models that do not require a high-quality initial guess, since the output layer is obtained from a physics-based training objective that enforces consistency with both measurements and dynamics.
While single-layer PIELMs can achieve high accuracy, they may exhibit reduced stability in regimes with limited measurement support. To improve representational capacity and generalization, the Deep PIELM augments the model with an autoencoder-based feature hierarchy that is pretrained efficiently via the Moore–Penrose pseudoinverse, followed by physics-informed nonlinear least-squares optimization of the final layer.
Comparative results highlight the trade-offs among classical least squares, single-layer PIELM, and Deep PIELM in terms of OD accuracy, robustness under poor initialization, and computational efficiency under sparse optical and range measurements from a limited set of ground stations. For suitable hyperparameter configurations, the multilayer architecture provides improved stability and accuracy over the single-layer variant while retaining low training times, positioning Deep PIELMs as an effective complement to classical least-squares OD when robust performance without reliable initial guesses is required. The presented work is part of the Artificial Intelligence for Precise Orbit Determination project.
How to cite: Dallinger, F., Aigner, B., Andert, T., and Haser, B.: Single- vs. Multilayer Physics-Informed Extreme Learning Machines for Orbit Determination, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19927, https://doi.org/10.5194/egusphere-egu26-19927, 2026.
The Meteosat Third Generation (MTG) mission represents a major step forward in geostationary meteorological observation by combining, onboard Meteosat-12, multiple instruments with highly complementary characteristics. Among them, the Flexible Combined Imager (FCI) provides multispectral images of the full Earth disk every ten minutes with a spatial resolution reaching 1 km at nadir, while the Lightning Imager (LI) observes the same scene at a much higher temporal sampling, but with a coarser spatial resolution of approximately 4.5 km at nadir. Although designed for distinct operational purposes, these two sensors offer a unique opportunity for joint exploitation, as they observe identical atmospheric phenomena under fundamentally different spatio-temporal trade-offs. In this context, Thales investigates the use of artificial intelligence techniques to leverage this complementarity and generate enhanced observation products from existing MTG-I data.
The core hypothesis of this work is that the high temporal density of LI observations implicitly encodes fine-scale spatial information. In other words, temporal correlations within LI time series can partially compensate for the sensor’s lower spatial resolution. By exploiting these correlations, fine spatial features can be reconstructed from high temporal frequencies. The availability of reference matching high resolution data enables to consider this process without the need for artificially degraded training data.
To implement this hypothesis, a hybrid deep learning architecture combining convolutional neural networks (CNNs) and Transformers is proposed. CNN components are used to efficiently extract local spatial structures, such as gradients, cloud edges, and internal texture patterns, while Transformer-based attention mechanisms model short- and long-range temporal dependencies across successive LI acquisitions. This combination enables a joint representation of spatial detail and temporal coherence, while remaining compatible with large data volumes and near-operational processing constraints.
The proposed approach is evaluated along two complementary scientific tasks. The first focuses on spatial super-resolution of LI images using LI temporal sequences alone. The second addresses the fusion of FCI and LI data to generate a product combining high spatial resolution with high temporal frequency. In both cases, the results are conclusive. The use of FCI images as a cross-reference makes it possible to assess the physical consistency of reconstructed features and to prevent the introduction of spurious, non-physical details. The super-resolved products remain radiometrically consistent with the input observations, with low radiance discrepancies (RMSE below 1), while recovering finer spatial structures than those achievable through conventional interpolation methods. Compared to standard SISR (Single Image Super Resolution), CNN + Temporal Conv1D, CNN + sparse Conv3D approaches, the hybrid CNN–Transformer model achieves the best overall performance.
As a perspective, the proposed method shows strong potential for operational deployment. Its computational efficiency allows approximately one hour of MTG data—corresponding to about sixty full-disk Earth images—to be processed in less than five minutes on standard computing infrastructure with one Nvidia H-100 configuration, paving the way for the routine generation of high-resolution, high-frequency products from existing geostationary missions.
How to cite: Dublé, N., Tanguy, S., Arsene, L., Poulain, V., Puechmaille, D., Hinojo Comellas, O., and Stoicescu, M.: AETHER: AI Enhancement for Third-gen Earth observing ImageR. Reaching 3x spatial upsampling and 10x temporal upsampling from existing MTG-I products., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22777, https://doi.org/10.5194/egusphere-egu26-22777, 2026.
Rapid and objective mapping of co-seismic surface ruptures is essential for post-earthquake impact assessment and for improving our understanding of fault geometry, stress transfer, and rupture processes that inform longer-term seismic hazard analyses. However, rupture mapping has traditionally relied on manual interpretation of field observations or remote-sensing data, which is time-consuming and difficult to extend consistently to large spatial extents, multiple earthquakes, and diverse data sources. Here we present an automated deep-learning framework—the Deep Rupture Mapping Network (DRMNet)—a convolutional neural network designed for end-to-end, high-precision detection of co-seismic surface ruptures from multi-sensor imagery. DRMNet is applied to four large continental earthquakes: the 2021 Mw 7.4 Maduo, 2022 Mw 6.9 Menyuan, 2001 Mw 7.8 Kokoxili, and 1905 Mw ~8 Bulnay (Mongolia) events. The framework consistently delineates both primary and subsidiary rupture structures across centimetre-scale drone imagery and metre-scale satellite data. Across diverse tectonic settings, image resolutions, and preservation states, DRMNet achieves precisions approaching or exceeding 90%. By enabling consistent rupture recognition across multiple events, sensors, and timescales, the proposed framework overcomes the event-specific and local-scale limitations of previous approaches, supporting both rapid post-earthquake response and retrospective rupture reconstruction, and laying the groundwork for standardized global surface-rupture inventories.
How to cite: Liu, X., Wang, S., Shi, X., Su, C., Klinger, Y., Delorme, A., Li, H., Pan, J., and Chen, H.: Deep-learning based large-scale automated observation of earthquake surface ruptures, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11012, https://doi.org/10.5194/egusphere-egu26-11012, 2026.
Natural (geological) hydrogen refers to molecular hydrogen produced in the subsurface through abiotic and biogenic pathways, which may migrate, accumulate transiently, be consumed by secondary reactions, or escape to the surface. Increasing evidence indicates that such systems could be a strategic low-carbon energy source, but their exploration is limited as regional-scale, data-driven approaches to identify mechanisms of active or fossil migration in geologically complex environments are lacking. Surface expressions such as circular and sub-circular depressions associated with soil and vegetation anomalies have been reported worldwide as indirect indicators of hydrogen migration and leakage. However, their detection remains limited to either local reconnaissance of the field or manual interpretation of remote-sensing data. In this research, we present an AI-assisted remote sensing framework to conduct a regional screening based on the potential for natural hydrogen seepage patterns to enhance early-stage exploration and improve the quantitative characterization of surface indicators linked to subsurface energy systems. Deep-learning–based computer vision models are used to study high-resolution satellite imagery and automatically identify and classify circular and sub-circular geomorphological features that could correspond to hydrogen exudation. The resulting detections are integrated into a GIS framework for the extraction of morphometric and spatial statistics, providing a formal analytical benchmark to relate surface structures to lithology, structural configuration, and the regional tectonic setting. The workflow is applied to the Alta Guajira region (in northern Colombia), a geologically complex segment of the Caribbean margin characterized by accreted oceanic crust, major fault systems, and sedimentary depocenters that may favor hydrogen generation and migration. Using an AI-based approach allows the construction of a regional inventory of candidate seepage-related structures while significantly reducing false positives associated with purely morphology-based analyses. The results support the prioritization of targets for future field verification, geochemical sampling, and subsurface investigations. Beyond its implications for natural hydrogen prospectivity, the proposed methodology demonstrates how artificial intelligence can translate qualitative geological observations into quantitative, reproducible screening tools. By providing a transparent and spatially explicit representation of subsurface energy systems, AI-assisted screening also facilitates communication with stakeholders and local communities, contributing to informed public perception of emerging sustainable subsurface energy resources in data-limited regions such as Alta Guajira.
The researchers thank the SHATKI Research Project (code 110563), Contingent Recovery Contract No. 112721-042-2025, funded by the Ministry of Science, Technology and Innovation (Minciencias) and the National Hydrocarbons Agency (ANH).
How to cite: Monterroza Montes, M. A., San Martín Cañas, S., Lora-Ariza, B., and Donado, L. D.: AI-assisted Remote Sensing Screening of Potential Natural Hydrogen Seepage Features in Alta Guajira, Northern Colombia, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15913, https://doi.org/10.5194/egusphere-egu26-15913, 2026.
The volcanic island of Tenerife (Canary Islands, Spain) is characterized by low-magnitude background seismicity associated with local hydrothermal and volcano-tectonic processes. The island has been experiencing, since 2016, a slight increase in seismic activity, with earthquakes generally having magnitudes below 2. For this reason, we are revising the seismic catalogue using deep learning tools to improve its completeness.
Over the last decade, machine learning methods—particularly deep learning approaches—have gained traction across multiple disciplines due to their increased computational efficiency, high accuracy, and reduced need for manual supervision. One such method, PhaseNet [1], is a deep convolutional neural network based on the U-Net architecture [2] that has shown strong performance in waveform-based seismic phase detection. Its ability to process large volumes of seismic data and automatically identify relevant signal features represents a significant opportunity to enhance the quality and completeness of seismic catalogs. Nevertheless, applying a neural network to data with a different nature from that used for its training phase can lead to a substantial decrease in performance. In particular, PhaseNet was primarily trained on tectonic seismicity, whereas seismic events in Tenerife are predominantly volcanic-hydrothermal. Consequently, retraining the network on waveforms representative of the target seismicity is essential to ensure a reliable inference.
Using PhaseNet as a baseline, we conducted an extensive comparative analysis of several training configurations to adapt the original network to the seismic data from the Canary Islands (Tenerife). Our study focused on four key aspects: model initialization, learning rate selection, data clustering strategies, and model partitioning. The model initialization strategies include fine-tuning from pre-trained weights and training from randomly initialized weights. Regarding model partitioning, we evaluated a global model (a single model trained on all data), local models (one model per station), and cluster-based models (trained on groups of stations with similar characteristics). The performance of each configuration was evaluated on an independent dataset using multiple metrics to provide a comprehensive assessment. Specifically, we analyzed precision, recall, and ROC curves to identify suitable trade-offs between detection sensitivity and specificity.
These preliminary results will be beneficial for subsequent analysis aimed at a better characterization of the island's microseismicity and its relationship with the activity of its volcanic-hydrothermal system.
References:
- [1] Zhu and G. C. Beroza, “PhaseNet: a Deep-Neural-Network-Based seismic arrival time picking method,” Geophysical Journal International, Oct. 2018, doi: 10.1093/gji/ggy423.
- [2] O. Ronneberger, P. Fischer, and T. Brox, “U-NET: Convolutional Networks for Biomedical Image Segmentation,” in Lecture notes in computer science, 2015, pp. 234–241. doi: 10.1007/978-3-319-24574-4_28.
How to cite: Calderón-Delgado, M., D’Auria, L., Álvarez-Hernández, A., García-Hernández, R., Ortega-Ramos, V., M. van Dorth, D., de Armas-Rillo, S., López-Díaz, P., and M. Pérez, N.: Improving the seismic catalogue completeness of Tenerife (Canary Islands, Spain) through deep learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19427, https://doi.org/10.5194/egusphere-egu26-19427, 2026.
