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.
Preserving clean water resources and efficiently treating wastewater is critical for ensuring human survival on Earth and in extraterrestrial environments. Major pollutants, including ammonium, heavy metals, industrial dyes, and chemicals, threaten limited clean water supplies and soil. Among the renowned absorbent materials, natural zeolite minerals have demonstrated their effectiveness for pollutant removal compared to clay minerals and synthetic equivalents like biochar, activated carbon, and MOFs, owing to their relatively extensive reserves and eco-friendly nature. Recently, investigating and optimizing pollutant removal rates from water without conducting laboratory experiments is getting more crucial, considering the time-consuming, expensive, and error-prone nature of laboratory testing due to human factors and potential calibration issues among the chosen analytical techniques.
This study aims to forecast the ammonium removal efficiency (% adsorption) and capacity (mg/g) of natural and modified zeolites from aqueous solutions using the regression ensemble LSBoost (MATLAB R2024b) machine learning (ML) algorithm, which is equivalent to XGBoost open-source library. A total of 527 experiments on 15 different zeolite compositions were gathered from a combination of 14 suitable moderately recent (≥ 2005) and highly referred studies to assess the performance of zeolite minerals on ammonium removal rates from aqueous solutions. The LSBoost algorithm achieved over 0.99 R2 fitting for training and overall, 0.95 R2 for prediction on the quarterly partitioned testing data for both efficiency and capacity. Throughout the improvement of the ML models using different random forest ML approaches, the number of predictors was successfully reduced to 8 based on importance rates among 31 different features in the initial dataset, with a negligible accuracy loss (<0.1 R2) on both training and testing. This research provides a valuable contribution to optimizing applicable experimental parameters in water treatment processes by effectively identifying the significance of predictors within a comprehensive data set. In addition to this, our model not only provides a robust predictive tool for optimizing zeolite performance in water treatment but also represents the first open-sourced web application in the literature to estimate the water treatment performance of zeolites.
How to cite: Akkaş, E., Ünal, B. C., and Ersoy, O.: AI-Based Prediction and Optimization of Ammonium Removal Efficiency and Capacity of Natural Zeolites Using LSBoost (XGBoost) for Sustainable Ecosystems, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-809, https://doi.org/10.5194/egusphere-egu25-809, 2025.
Artificial Intelligence and Machine Learning (AI/ML) are gaining increasing interest due to their capacity to increase precision and productivity in the current big data era. Machine learning has indicated its robustness in geosciences, particularly rock-type classification. Lithological classification in the traditional way has raised critical concerns and the need to curb the limitations it breeds, such as time consumption and subjective results. The gold mineralisation occurrence is structurally controlled in the Obuasi Gold district of Ghana. It exhibits complex patterns and relationships that may not be readily discernible through traditional methods, leading to missing out on discovering new resources or potential exploration targets. Consequently, this work attempts to create a predictive model by exploring the best machine-learning algorithms to predict rock types in the Obuasi Gold District using X-ray fluorescence (XRF) geochemical data. Here we established comparative predictive modelling using four supervised classification algorithms: Gradient Boosting (GBoost), Adaptive Boosting (AdaBoost), Support Vector Machine (SVM) and Random Forest (RF). The acquired XRF data was integrated with the model using the Google Collaboratory cloud-based platform. Results show that the performance evaluation of the models indicated SVM as the best algorithm for deployment with a Classification Accuracy (CA) of 0.902. Therefore, ML algorithms have been a great tool in rock-type classification, whereby SVM emerged as the best in the case of the Obuasi Gold District. However, it is encouraged to understand the geology of a particular area before employing the tool and the datasets must be balanced to yield good results and avoid model overfitting.
Keywords: Artificial intelligence; Machine learning algorithm; Support vector machine; Lithogeochemistry; Rock-type classification; Obuasi Gold District
How to cite: Muhammed, A. B., Sunkari, E. D., and Basit, A. W.: Application of Machine Learning Algorithms to Predict Rock Types Using Geochemical Data: A Case Study from the Obuasi Gold District, Ghana, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-275, https://doi.org/10.5194/egusphere-egu25-275, 2025.
Artificial Intelligence (AI) is revolutionizing the field of geomorphology, offering a robust tool for objective and quantitative analyses. This pioneering study proposes an innovative framework based on Machine Learning clustering techniques, capable of classifying drainage patterns into multiple morphological classes. This work follows up on a related study in which an attempt at classifying 156 terrestrial and extraterrestrial (Mars and Titan) river networks was made. Rivers’ outlines are intrinsically noisy, difficult to isolate from the background, and can be ambiguous for the human eye. The previous works have been focused on accurately classifying patterns, using the expertise of morphologists, thus introducing a weak link, the human eye, in the chain. This time, a reliable, automatic, and scalable methodology has been obtained, leveraging computers’ precision, objectivity, and computational power. The HydroRIVERS dataset, a publicly available data bank containing vector data, was utilized in this study. All HydroRIVERS data layers are provided in a geographic projection (latitude/longitude), referenced to the WGS84 datum. Each data layer includes an attribute table with information on the morphometric characteristics of each river reach. The input parameters for the clustering models included morphometric features such as LENGTH_KM, DIS_AV_CMS, ORD_STRA, ORD_CLAS, and ORD_FLOW.
During a preliminary experiment, a local convexity test was conducted to determine the optimal number of clusters (k) to identify the best metric values. This test made sure that the number of clusters with the highest evaluation metric was selected, varying in a closed numeric interval. Each cluster corresponds to a specific river class. Significant results were obtained with k = 6, k = 8, k = 10, and k = 12. Subsequently, the K-Means algorithm was applied, grouping the dataset into distinct clusters based on the morphometric parameters. The results were remarkable, with 10 being the best value for k. The results indicate that the clustering algorithm is able to optimally separate the dataset, producing a high inter-cluster distance and a low intra-cluster distance. The dataset points along the features, as highlighted by the three principal components obtained by performing PCA on the final five-dimensional clustering resulting vector space, are well grouped in relatively small clusters, far away from each other. The next step involves using the centroids obtained from the analysis of the large dataset as a reference for classifying the 155 rivers. In general, the centroids obtained from this kind of Learning could be of great value to the scientific community, establishing a new and innovative way of discerning between different classes of rivers without having to manually analyze and inspect images. This approach promises efficient and accurate classification of both terrestrial and extraterrestrial drainage patterns.
How to cite: D'Aniello, M. and Donadio, C.: A novel approach using Machine Learning to objectively classify terrestrial and extraterrestrial river networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19838, https://doi.org/10.5194/egusphere-egu25-19838, 2025.
Recent trends in climate change and global warming have amplified the occurrence of severe convective initiation (CI) and other extreme weather events, underscoring the importance of high-frequency remote sensing. Modern satellite constellations such as EUMETSAT, GOES, Himawari, and GK2A now offer global monitoring intervals of 10 minutes and regional updates as frequent as 1–2 minutes. However, these capabilities are not universally accessible—many developing countries lack in-orbit assets or historical high-frequency data archives.
This study presents a zero-shot video frame interpolation (VFI) approach to generate high-frequency (1–10 minute) satellite imagery from legacy or sparsely sampled observations. Leveraging a flexible Many-to-Many Splatting VFI model, our framework avoids domain-specific retraining while delivering reliable intermediate frames. We validate the method using overlapping data from the KMA GK2A (10-minute full-disk, 2-minute Asia-Pacific) and KMA COMS (3-hour full-disk, 15-minute Asia-Pacific) satellites over the period July 25, 2019, to March 31, 2020.
Our results indicate significant improvements in both PSNR and SSIM metrics, confirming the model’s efficacy in three critical applications:
-
Up-sampling Archived Geostationary Data
- Enhancing the temporal resolution of older satellite imagery (e.g., COMS 30-minute or 3-hour intervals) to match or approximate modern satellite capabilities (e.g., GK2A 10-minute intervals). This harmonization facilitates unified climatology analyses spanning multiple generations of instruments.
-
Sensor-Error Correction and Gap Filling
- Recovering missing or corrupted frames resulting from attitude adjustments, sensor calibrations, or malfunctions on geostationary satellites. Ensuring a continuous record provides more robust inputs to operational forecasting and climate assessments.
-
Delayed High-Frequency Observation Services
- Enabling resource-constrained meteorological agencies to retrospectively produce and disseminate high-frequency satellite products (e.g., near 1-minute intervals) to improve nowcasting, risk assessment, and disaster preparedness.
Our preliminary findings show minimal computational overhead per inference step, making this cost-effective method feasible for near-real-time deployment and post-event analyses alike. By bridging temporal gaps in global satellite datasets, this technique supports advanced level-2 products such as cloud-tracking and convective-initiation alerts, thereby driving broader socioeconomic and scientific benefits in both developed and developing regions.
How to cite: Ryu, H. and Choi, Y.: Toward High-Frequency Satellite Observations in Data-Sparse Regions: A Zero-Shot Interpolation Framework for Missing and Historical Imagery, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17456, https://doi.org/10.5194/egusphere-egu25-17456, 2025.
In the age of big data, artificial intelligence (AI) is transforming Earth sciences by enabling efficient analysis and visualisation of complex datasets and fostering innovative approaches to solve age-old geoscientific challenges. Kalpa, a Python-based free and cross platform software, represents a pioneering step in this direction. Built with versatility at its core, Kalpa seamlessly integrates AI and machine learning workflows into geoscience applications, offering customization through its Python plugin architecture. What sets Kalpa apart is its ease of use, even for non-experts. Its intuitive interface lowers the learning curve, enabling a broader audience—including researchers, professionals, and enthusiasts—to leverage advanced geospatial and AI tools without requiring extensive technical expertise.
Kalpa's capabilities span advanced 3D visualization, geospatial data processing, and machine learning model development. It supports global and regional raster and vector dataset visualisation and processing, allowing for interactive analysis in both geographic and cartesian coordinates. With tools to process satellite imagery, geological and geophysical data, uncrewed aerial vehicle (UAV) data, and digital geological maps, Kalpa caters to a wide range of applications, from mineral exploration to natural hazard forecasting. Its machine learning integration supports supervised and unsupervised algorithms for applications such as lithological mapping, mineral prospectivity mapping, land cover and land usage studies, agricultural productivity mapping and natural disaster management. In this study, we demonstrate Kalpa’s transformative potential through three case studies:
- Lithological Mapping in Ladakh, India: Utilizing LANDSAT and SRTM data, we produced accurate lithological maps for this geologically complex region.
- Copper Prospectivity Mapping in Northwest India: Combining remote sensing, geophysical, and geological data, Kalpa predicted copper mineralization zones, with all known deposits falling within areas of predicted probabilities exceeding 0.70.
- Landslide Susceptibility Mapping in Uttarakhand, India: Using remote sensing datasets, Kalpa identified high-risk landslide zones, supporting disaster management efforts.
Kalpa’s user-friendly interface, robust machine learning integration, and publication-ready export capabilities position it as a powerful tool for advancing geoscience research and practical applications. By bridging the gap between domain expertise and cutting-edge AI methodologies, Kalpa empowers Earth scientists, environmental researchers, and GIS professionals to analyze, model, and predict with unprecedented efficiency and precision. This software marks a new frontier in the application of AI to Earth sciences, enabling multidisciplinary research and fostering innovative solutions to pressing geoscientific challenges.
How to cite: Sukhbir, S., Singh, S. P., Singh, U., Kumar, M., and Goyal, T.: Kalpa: Empowering Artificial Intelligence-Driven Geospatial Analysis for Multidisciplinary Applications, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9671, https://doi.org/10.5194/egusphere-egu25-9671, 2025.
Shale oil and shale gas are important unconventional resources. Organic matter serves as the primary source of shale oil and gas generation, and high TOC values typically indicate better oil and gas reservoir conditions and production. Therefore, an accurate TOC prediction model is conducive to low-cost evaluation of reservoir hydrocarbon potential and improvement of development efficiency. However, geochemical experimental measurements are costly, and the data obtained is discrete. It is unable to meet the requirements for fine-scale assessment of shale reservoirs. The multiple regression method and ΔlogR method, when directly applied to shale reservoirs, often result in significant errors. In this study, we propose a composite model for accurate TOC prediction in shale reservoirs based on data enhancement and empirically driven. We first address the issue of poorly characterized logging responses and discrete experimental data. The features and quantities of the dataset are enhanced by introducing reconstruction curves and generative adversarial networks (GAN). The validity of the synthesized data is then verified by plotting the data density. In the empirically-driven module, we optimize a density-gamma modified method on traditional ΔlogR method according to the characteristics of shale reservoirs. The modified ΔlogR method will be integrated into the GWO-SVR model as an empirically driven subject in the form of a fitness function. Above, a composite model with both empirical and data-driven components is constructed. We use the Dongying Depression in China as an example for model experiments. The composite model was generalized to wells X and Y. The R² (coefficient of determination) was 0.95 and 0.97, the RMSE (Root Mean Square Error) was 0.31 and 0.29, and the MAE (Mean Absolute Error) was less than 0.3, which indicated a high degree of consistency between the model predictions and the experimental values. Further controlled experiments revealed that the composite model predicted better than the ΔlogR method and the GWO-SVR model alone. Finally, we also performed SHAP interpretability analysis on the model. By revealing the decision-making mechanism inside the model, we verified the rationality of the empirical drive and enhanced the credibility of the model. This provides strong technical support and decision-making basis for the subsequent oil and gas exploration and development work.
How to cite: Hong, Y., Deng, S., Li, Z., and Wei, Z.: TOC Intelligent Prediction Model in Shale Reservoir: Integrating Data Enhancement with Empirically Driven Algorithm, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2156, https://doi.org/10.5194/egusphere-egu25-2156, 2025.
The shear wave (S-wave) velocity is a key basis for shale reservoir development, particularly for fracability evaluation. Additionally, S-wave velocity also plays a significant role in prestack seismic inversion and amplitude versus offset (AVO) analysis. However, the actual logging data often lack S-wave velocity data, so it is of significant importance for S-wave velocity prediction. We propose a rapid and precise prediction method for the S-wave velocity in shale reservoirs based on class activation maps (CAM) model combined with physically constrained two-dimensional Convolutional Neural Network (2D-CNN). High sensitivity curves related to S-wave velocity are selected as the foundation. Meanwhile, based on the petrophysical theory of pore medium, the petrophysical model of complex multi-mineral components is established. The dispersion effect is reduced to a certain extent and the results are used to constrain the model. The Adam optimization algorithm is used to construct a 2D-CNN model under the constraint of petrophysical model. The CAM is obtained by replacing the global average pooling (GAP) layer with a fully connected layer, which in turn leads to interpretable results. Then, the model is applied to wells A, B1, and B2 in the southern Songliao Basin. Afterwards, comparisons are made with unconstrained model and petrophysical model. The results show that the correlation coefficients and relative errors in the three test wells are 0.96 and 2.14%, 0.97 and 2.35%, and 0.97 and 2.9%, respectively. The higher prediction accuracy and generalization ability of the new method is confirmed. Finally, we present the defined C-factor as a means of evaluating the extent of concern regarding CAMs in regression problems. The C-factor confirms that the focus of 2D-CNN can be significantly enhanced by incorporating the petrophysical model, thereby imposing physical constraints on the 2D-CNN. In addition, we establish the SHAP model to assist in proving the importance of constraints.
Keywords: S-wave velocity prediction; Physically constrained 2D-CNN; Petrophysical model; Class activation mapping technique; Explainable results
How to cite: Li, Z., Deng, S., and Hong, Y.: S-wave velocity prediction of shale reservoirs based on explainable physically-data driven model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4699, https://doi.org/10.5194/egusphere-egu25-4699, 2025.
In the context of CO₂ storage, cost-effective monitor methods are essential to ensure safe and long-term storage. This work explores the use of seismic time-lapse monitoring, combined with deep learning (DL) techniques, to assess potential leakage and migration pathways. The goal is to develop a cost-effective monitoring method while guaranteeing the safety of storage operations. To this end, we propose a Siamese Neural Network (SNN)-based framework to analyse shot gathers, designed to detect and localize changes within the storage complex. We aim to address the challenges of working with large seismic datasets, enabling the identification of significant events with high confidence, while avoiding the need for event-by-event processing. This framework can allow experts to rely on semi-automatic detections while ensuring human evaluation for interpreting and validating the results.
The proposed SNN architecture processes pairs of shot gather from baseline and monitor surveys in a cross-well configuration. It uses two identical neural networks with shared weights to encode the shot gathers into latent feature embeddings, which are then compared to identify similarities and detect changes. By transforming the data into a shared latent space, the model focuses on capturing relevant patterns while filtering out irrelevant variations, ensuring robust and accurate comparisons. When the SNN detects changes between the baseline and the monitor surveys, it highlights the regions where these changes occur. This approach is particularly effective for identifying subtle but important changes in seismic data, such as those caused by CO₂ migration, which alters the velocity and density of the subsurface. Even in noisy data, the SNN can detect these variations, thanks to its ability to learn features that are highly sensitive to small but meaningful changes. The SNN architecture is scalable and can be adaptable to various seismic monitoring tasks, requiring minimal preprocessing. The proposed framework harnesses the power of deep learning to provide insights into the dynamics of the storage complex, with a focus on identifying changes in time-lapse seismic data related to localized variations. The proposed migration detection tool offers a cost-effective and reliable solution to the modern challenges of gas storage monitoring. This study aims to enable operators to identify and address problems promptly, thereby minimising the impact of potential leakages.
How to cite: Pantaleo, G. and Pipan, M.: Assessment of a deep learning framework for time-lapse seismic monitoring, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-882, https://doi.org/10.5194/egusphere-egu25-882, 2025.
Seismic exploration heavily relies on the accurate processing of seismic data, as high-quality reconstructed data is essential for reliable imaging and interpretation. In recent years, data-driven approaches have shown great promise in seismic data processing. However, supervised learning methods require large amounts of labeled data, while generative models, such as GANs, often encounter issues like mode collapse and instability. On the other hand, generative diffusion models, leveraging principles from nonequilibrium thermodynamics and Markov processes, have emerged as powerful tools for capturing complex data distributions.
Despite these advantages, Denoising Diffusion Probabilistic Models (DDPM) purely generate data distributions from the latent space with the reliance on random noise, making it inadequate for seismic data reconstruction where the goal is to accurately recover missing traces. Thus, DDPM is lacks interpretability in seismic data restoration, and may disrupt the structured patterns crucial for interpolating seismic signals. Furthermore, we view the reverse process that starts from noise as unnecessary and inefficient for reconstruction task.
To address these challenges, we propose a novel Conditional Residual Diffusion Model (CRDM) that enhances both certainty and interpretability by incorporating residual diffusion and conditional constraints derived from observed seismic data (Fig.1). This approach better aligns with the inherent structure of seismic signals, enabling more accurate and interpretable reconstruction. The model is grounded in DDPM, with mathematical derivations for loss functions, conditional probability distributions, and reverse inference steps, ensuring both theoretical rigor and practical applicability.
Additionally, Our CRDM utilizes a shallow U-Net architecture featuring one down-sampling and one up-sampling layer integrated with Multi-Head Self-Attention (MHSA), which significantly enhances the model's efficiency and effectiveness. Experimental results (Fig.2) demonstrate that CRDM outperforms DDPM, denoising convolutional neural network (DnCNN), and fast projection onto convex sets (FPOCS), achieving a 15.1% improvement in reconstruction SNR and reducing computation time by 139 times compared to DDPM. Notably, CRDM achieves optimal results in a few diffusion steps, whereas DDPM typically requires thousands of steps.
The innovative approach generates data through residuals for determinism, while guiding the processing with noise for diversity. This not only enhances the interpretability and efficiency of seismic data reconstruction, but also positions the model as a promising tool for advancing data-driven seismic processing through flexible coefficient adjustment. Therefore, we believe this model has great potential for broader applications in geophysical data analysis, offering significant value in accurately depicting complex geological structures and providing more effective guidance for petroleum exploration.
Fig.1 The framework of CRDM. The model consists of two stage: (a) the training stage with forward diffusion process; (b) the sampling stage for seismic data reconstruction.
Fig.2 Reconstruction results and residuals of the 1994 BP seismic data with 50% irregular missing traces. (a) Complete data, reconstruction using (b) FPOCS (SNR=10.5dB), (c) DnCNN (SNR=15.6dB), (d) DDPM(SNR=18.4dB), (e) CRDM(SNR=21.2dB), and (f) observed seismic data with 50% missing traces. (g-j) display the residuals corresponding to reconstructions (b-e), respectively. The red box highlights a zoomed-in region, which is shown in detail in (k-t).
How to cite: Gong, X. and chen, S.: Enhancing Seismic Data Reconstruction with a Conditionally Constrained Residual Diffusion Model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16835, https://doi.org/10.5194/egusphere-egu25-16835, 2025.
Borehole images (BHI) are crucial for resource exploration, providing detailed fracture analysis at millimeter-scale resolution. However, their interpretation is typically carried out manually, a process that is time-consuming, costly, and subject to significant uncertainty due to interpreter bias and variability. Current state-of-the-art AI methods for automated or semi-automated fracture analysis of BHI often rely on field data for training, using manual interpretations as labels. This approach inherently embeds both aleatoric (data-related) and epistemic (manual) uncertainties, which may undermine the reliability and adaptability of these methods. This study proposes an alternative, synthetic data-driven approach to train a set of two deep neural networks (DNNs) connected in sequence. These DNNs are designed to replicate the primary cognitive tasks involved in manual interpretation: the segmentation of the BHI to identify potential edge zones and the tracing of sinusoids over these edges to approximate their best-fitting 2D representation. By utilizing synthetic data, we are able to systematically assess the sensitivity of both networks and explore various training strategies, including curriculum learning (CL) and self-attention mechanisms. Our proposed solution is designed for post-hoc human-machine collaboration, where the model supports but does not replace human expertise. This framework enables the possibility of a multi-level uncertainty assessment—at the human, machine, and human-machine interface levels—opening the door to new ways of understanding and quantifying the sources of uncertainty in BHI analysis. Additionally, the synthetic data-driven approach ensures the generalizability and scalability of the method, as demonstrated by its successful application to low-resolution logging-while-drilling (LWD) and high-resolution fullbore formation microimager (FMI) datasets from multiple global locations. By combining advanced AI techniques with geoscientific knowledge, this study outlines a potential pathway toward more robust, ethical, and sustainable fracture analysis workflows. Beyond the traditional benefits of reduced cost and time, the approach may provide a scientifically grounded framework for exploring the benefits of human-machine collaboration and uncertainty quantification in geoscience practices. If adopted, this framework could significantly advance the field of BHI analysis, offering new tools for resource exploration in hydrocarbon and geothermal applications.
How to cite: Molossi, A., Roncoroni, G., and Pipan, M.: NeuroFit: a robust and scalable synthetic data-driven deep learning solution for automated borehole image analysis at LWD and wireline resolution, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10472, https://doi.org/10.5194/egusphere-egu25-10472, 2025.
Advanced constitutive models, here represented by the hypoplastic clay model, are powerful tools which provide engineers with reliable responses in various practical applications. However, the model calibration is not an easy task. Calibration of these models can be addressed with several approaches, which are generally distinguished as stochastic or deterministic approaches. In general, these approaches extract information from the experimental data and the subsequent optimisation process finds the best combination of parameters to fit the desired constraints. . The deterministic approach was integrated and combined in development of the online automated calibration tool ExCalibre. This paper presents a Machine Learning approach for automated calibration of the Hypoplastic Clay model. By using pairs of input experimental data and calibrated results performed by ExCalibre as training data, a Deep Neural Networks (DNNs) model is constructed to recognise how the experimental data can be used to derive the asymptotic state parameters such as the slope and the interception of the Normal Compression Line (NCL), or the critical friction angle, and the optimised stiffness parameters. The training and testing data comprise of In-house protocols and User-upload data over 3 years of launching the ExCalibre, and synthetic data with small distortion to prevent overfitting. Finally, investigations on how the DNNs model recognises the asymptotic patterns, as well as its calibration results will be presented.
How to cite: Do, P. and Kadlicek, T.: Calibration of the Hypoplastic Clay model with a deep neutral network, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12065, https://doi.org/10.5194/egusphere-egu25-12065, 2025.
Joint inversion is an essential technique in potential field data processing. The current methodology largely relies on the geology model of anomalous bodies, especially for deep, complex structures. Inspired by the excellent nonlinear mapping capability of the image semantic segmentation model and the advantages of supervised learning, a regressive, end-to-end, encoder-decoder structural, convolutional neural network with a double-branch structure called PFInvNet(Potential Field Inversion Neural Network) is proposed for joint 3D inversion of physical properties from gravity and magnetic data. Its input is a four-channel dataset consisting of gravity and magnetic anomalies and their vertical gradients, and its output is a 3D matrix representing the spatial distribution of the remnant density and the magnetic susceptibility, which are predicted independently through the double-branch structure of the decoders and then concatenated in the final layer. For network training, a large amount of precisely labeled sample is exceedingly demanding; thus, forward modeling becomes a prerequisite approach. Two discretized forward modeling algorithms for gravity and magnetic anomalies of 3D homogeneous arbitrary-shaped bodies based on surface integrals are deduced and verified with analytic solutions of the sphere model. Furthermore, the neural network needs to learn from the anomalies generated by various forms of abnormal bodies with different physical properties. Therefore, different sizes and quantities of cuboids are randomly distributed in the model space to simulate different forms of abnormal bodies. The label represents the combined spatial distribution of remanent density and magnetic susceptibility for the cuboids, encompassing both spatial location information and physical properties information. With the help of the Marching Cubes(MC) algorithm, the surface of the cuboids can be easily extracted and divided into a triangular surface mesh. The surface mesh is then used to calculate the gravity and magnetic anomalies synchronously through the forward modeling algorithms. The anomalies are concatenated in the channel direction as a sample. A set of optimal network parameters has been determined, including the weight initialization method, the gradient calculation methods, the loss function, the training hyperparameters, the regularization method, and the normalization method. The PFInvNet is trained with 500 and 10000 pairs of samples and labels, respectively. The analysis and comparison of training results prove that PFInvNet has two crucial features: one is that the branch structure enables independent prediction of magnetic susceptibility and remanent density; the other is efficient anti-overfitting ability and efficient solution-finding ability .The prediction error of small samples is very close to that of large samples and is also not obviously enhanced by the noise-contaminated data , demonstrating the strong generalization and robustness of the network. Finally, the network is tested with magnetic and gravity anomalies of the Victoria Land Basin in the western Ross Sea through transfer learning and retraining, and definite 3D distributions of apparent remnant density and apparent magnetic susceptibility have been obtained and can be checked with geological evidences.
How to cite: Jiang, W., Zhao, Y., Gao, J., Ge, S., and Xie, Z.: 3D property inversion of gravity and magnetic data based on a double branch regressive CNN trained by synchronous forward modeling ：a case study of the Western Ross Sea, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1141, https://doi.org/10.5194/egusphere-egu25-1141, 2025.
The Frequency Domain Electromagnetic (FDEM) method is a cost-effective geophysical technique that simultaneously studies the electrical and magnetic properties of a medium, providing data as in-phase and out-of-phase components of the electromagnetic field. Although FDEM yields valuable insights, its results can be complex to interpret, and the two EM field components are normally only visually inspected to support findings from other techniques. This study aims to enhance FDEM data interpretation using an unsupervised learning technique. The proposed approach seeks to automate and expedite the interpretative phase. By applying the K-Means clustering algorithm, we divided the FDEM data into several clusters based on specific intervals of the in-phase and quadrature components, resulting in integrated maps of EM components. Combining these maps with geological and archaeological insights helped identifying areas of potential archaeological interest. This method was applied to the Torre Galli archaeological site in Calabria, Italy, known for its significance in Iron Age studies.
Based on comparisons with the findings of earlier excavations and results from a magnetic survey, the proposed procedure shows promise in improving the efficiency and accuracy of the FDEM method in identifying areas of archaeological interest. This suggests that automating the interpretation process could lead to a better cost management and time optimization in geophysical and archaeological studies.
How to cite: Capozzoli, A., Paoletti, V., Cella, F., La Manna, M., and Piegari, E.: Integrating FDEM data with K-Means clustering for improved archaeological site identification, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18082, https://doi.org/10.5194/egusphere-egu25-18082, 2025.
Hazelnuts are a significant crop, with global production exceeding 1.25 million tons by 2023 (INC, 2023). Quality is threatened by biotic agents, including insects causing the cimiciato defect. This defect, from insect bites during fruit growth, results in off-flavor, tissue alterations, and lipid oxidation (De Benedetta et al., 2023). Damage can be external or internal (hidden cimiciato). Industrial quality standards often exceed official regulations, making effective selection crucial. Traditional visual inspection is time-consuming and subjective. Non-destructive methods like NIR and NMR have potential but limited applicability. Deep Learning (DL) has revolutionized image classification, proving effective in agriculture, including disease and pest management (Mohanty et al., 2016; Dhaka et al., 2021; Meena et al., 2023). This study explores Deep Learning (DL) for automated detection of the cimiciato defect in hazelnuts using X-ray radiographs. Cimiciato, caused by insect feeding, degrades hazelnut quality, requiring product selection. Traditional methods are time-consuming and subjective. We propose a Convolutional Neural Network (CNN) model trained on X-ray images to classify hazelnuts as healthy or infected. Results demonstrate the model's effectiveness, offering a non-destructive, automated quality control solution.
Radiographs were acquired using a cone-beam micro-tomograph. Each hazelnut was positioned on a rotating stage. A CNN model was used for classification. CNNs effectively extract features from images. Convolutional layers apply filters to identify features; pooling layers reduce data dimensionality; fully connected layers combine features for classification. The Inception-ResNet-V2 architecture was chosen, combining Inception modules and residual connections (Szegedy et al., 2017). The model was trained (128 image batch size, 0.001 learning rate, 30 epochs), comparing SGD, ADAM, and RMSP optimizers. Images were pre-processed: resizing, pixel normalization, and data augmentation.
The test dataset evaluated the trained network. SGD, ADAM, and RMSP yielded similar results. Confusion matrices visualize performance. ADAM performed best, but all achieved good results, especially for cimiciato detection.
Keywords: Cimiciato defect, Hazelnut, Deep Learning, X-ray radiography, CNN
How to cite: Napolitano, A. G., Mele, G., Gargiulo, L., Giaccone, M., and Vitale, A.: Automatic Detection of Cimiciato Defect in Hazelnuts Using Deep Learning and X-ray Radiography, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12675, https://doi.org/10.5194/egusphere-egu25-12675, 2025.
Since 2014, ESA Sentinel missions have been producing an ever-growing amount of data for Earth Observation. This creates the opportunity to monitor changes in high temporal and spatial resolution, however, interpreting this huge quantity of data is challenging. In recent years, the rapid advancement and widespread adoption of applied Artificial Intelligence (AI) methods made it feasible to create deep learning models for specific Earth Observation applications. Combining Sentinel datasets with the appropriate amount of ground truth, robust pre-trained models can be created and applied to produce good-quality thematic maps for different years and large areas. Due to responsibilities related to the European Union’s Common Agriculture Policy (CAP) and the motivation for regional yield estimation, crop classification is one of the most frequently studied remote sensing problems these days. Several papers investigate the possible methods to construct robust and generally functioning models to map the spatial distribution of crops as accurately as possible.
In this study, we present the development of a modular, pre-trained deep learning model designed specifically for crop type mapping. The model is tailored to classify the most prevalent crops in Hungary, including winter and autumn cereals, corn, sunflower, alfalfa, rapeseed, grasslands, and other significant types. For pre-training, we leverage country-wide Sentinel-1 Synthetic Aperture Radar (SAR) data such as Sigma Naught or polarimetric descriptors from H-A-alpha decomposition, collected during the 2021–2024 time period. This dataset comprises annual time series of Sentinel-1 pixels at a spatial resolution of 20 meters. Our approach builds upon prior findings that Sentinel-1-based crop type classification performs comparably to methods using Sentinel-2 optical data. However, Sentinel-1 has the added advantage of producing consistent and regular time series, as it is unaffected by atmospheric conditions such as cloud cover.
The proposed model employs a hybrid methodology integrating self-supervised and fully supervised learning paradigms. This modular architecture allows for seamless integration of task-specific classifiers, which can be fine-tuned using supervised learning to address both current and future classification requirements effectively. The fully supervised component is supported by an extensive ground truth dataset covering over 60% of Hungary’s total land area. This proprietary dataset is derived from the national agricultural subsidy database, providing detailed and accurate annotations. The abundance and quality of labeled data enable the construction of robust, highly generalizable models, ensuring reliable performance across diverse classification tasks. This methodology offers great potential to advance national-scale operational tasks, such as early land cover prediction and annual crop type mapping.
How to cite: Richter-Cserey, M., Simon, M., Magyar-Santen, G., Pacskó, V., and Kristóf, D.: Exploring Pretraining Possibilities of Crop Classifiaction Models Using Large-Scale Sentinel-1 Datasets, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11754, https://doi.org/10.5194/egusphere-egu25-11754, 2025.
Meteorological data are vast and complex, and their rapid and accurate retrieval is essential for forecasting operations. However, traditional systems have struggled with limited search accuracy and inefficient processing speeds, hindering effective forecast support. To address these challenges, this study developed an AI-based system capable of performing speech recognition, URL search, extreme value detection, and local forecast error analysis. In speech recognition, the Whisper-large model achieved a character error rate (CER) of 3.19%, with GPU memory usage reduced by 15.7% and inference time by 38.18%, enabling real-time processing and scalability in multi-GPU environments. The URL search systems translated natural language inputs into SQL queries and URLs, achieving a Mean Reciprocal Rank (MRR) of 0.92, thereby enhancing data retrieval precision. The extreme value detection systems utilized GPT-4-based template augmentation to expand training data by approximately 111%, significantly improving detection performance and search accuracy. For local forecast error analysis, a prototype chatbot was implemented using prompt engineering and a Text-to-SQL model, allowing for the automated identification of inconsistencies in local forecasts and streamlining the analysis process. These systems have substantially enhanced operational workflows across meteorological tasks, facilitating rapid data retrieval through voice commands, precise responses to complex queries, and real-time analytical support. Future research will focus on further refining these technologies to tackle a wider range of meteorological challenges and integrate them into global forecasting systems for enhanced accuracy and reliability.
How to cite: Kim, B., Shin, H., Cho, A., Park, J., Lee, H., Park, C., Joe, J., Choo, J., and Seo, M.: AI-Enhanced Meteorological Data Retrieval Systems for Improved Forecast Operations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4404, https://doi.org/10.5194/egusphere-egu25-4404, 2025.
The combination of big data and AI/ML technologies shows tremendous promise within the domain of disaster management. So that benefits of AI/ML can be realized, and risks can be adverted, internationally agreed upon standards are an important mechanism. These can provide guidance on how to apply (and develop policy around) data collection and preprocessing, model training and evaluation, and operational implementation. In conjunction, they can cultivate interoperability and harmonization of AI-based systems. At the United Nations, the Global Initiative on Resilience to Natural Hazards through AI Solutions brings together experts from different disciplines (geosciences, disaster risk management, computer sciences) and sectors (government, research, NGO) to analyze use cases and lay the groundwork for such standards. Through proof-of-concept projects (e.g., HEU-funded MedEWSa), these standards can be further refined. Finally, through education and capacity building activities, the Global Initiative can help to democratize the responsible use of AI for this domain.
How to cite: Kuglitsch, M. and Xoplaki, E.: International standards for responsible AI in disaster management, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2478, https://doi.org/10.5194/egusphere-egu25-2478, 2025.
The big developments in Artificial Intelligence (AI) generated a mixture of fascination and apprehension, echoing historical responses to unknown phenomena. Like how medieval European societies interpreted certain natural events and instruments as magical, AI is currently perceived by many as a “black box,” blurring the lines between advanced technology and mystical force. This phenomenon fosters misconceptions and uncertainties about AI’s actual mechanisms, benefits, and risks. Consequently, it underscores the urgent need for science communicators and science museums to adopt an active role in enhancing the general public’s AI literacy and in debunking some of its enigmatic traits.
By drawing parallels with the Middle Ages—when objects such as mirrors and magnetite were often attributed supernatural capabilities—modern AI tools LLMs or image and video generators are frequently viewed as possessing a “magical” principle. The public’s limited grasp of how AI processes inputs and produces outputs further intensifies this impression. The lack of transparency (or “black box” effect) in deep learning algorithms, combined with the ambiguity of human language, has been shown to fuel both wonder and anxiety.
Science museums and communicators should have an active role by offering educational programs that demystify AI through different demographic and social activities as well as promoting the public debate. These initiatives could clarify AI’s underlying mathematical and computational principles, highlight practical examples of AI-driven applications, and examine ethical considerations surrounding its deployment. Public understanding of AI’s capabilities and limitations is crucial not only to temper undue fears but also to encourage informed engagement with emerging technologies.
How to cite: Azevedo Rodrigues, L.: Making Sense of AI: The Important Role of Education and Communication, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20509, https://doi.org/10.5194/egusphere-egu25-20509, 2025.
We present a novel approach to gravity forward modeling using conditional neural operators that establishes a forward generative model from the basin models and hyperparameters (reference basement depth, etc.) to gravity anomaly. Our methodology introduces an innovative adaptive embedding mechanism where scalar hyperparameters are first embedded into a 32-dimensional space and then adaptively expanded to match the dimensions of the basin depth model, enabling effective fusion with basin depth model data. Subsequently, Fourier Convolution Layers are employed to transform the fused data into gravity anomalies. The model demonstrates superior performance compared to existing convolutional neural networks on the test dataset, showcasing improved accuracy in capturing complex geological structures and their gravity responses. A key advantage of our architectural design is that it not only preserves the super-resolution capability of conventional neural operators but also enables controlled generation through different hyperparameters. This dual capability allows for both resolution-flexible modeling and parameter-controlled generation, while training on low-resolution data and producing high-resolution outputs, significantly reducing training data requirements and computational costs. The model's adaptive architecture effectively bridges the resolution gap between training and application scenarios, offering a practical solution for real-world geological surveys. Our results suggest that this approach could substantially improve the accessibility and applicability of gravity forward modeling in various geological settings, particularly in regions with limited high-resolution training data.
How to cite: Kang, R., Geng, M., Yang, Q., and Vega, F.: A Conditional Neural Operator Approach for Resolution-Flexible and Parameter-Controlled Gravity Forward Modelling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2242, https://doi.org/10.5194/egusphere-egu25-2242, 2025.
Stratigraphic correlation of well log data is a fundamental step in geosciences. It involves correlating stratigraphic units across multiple wells to build a comprehensive understanding of subsurface geology. Currently, stratigraphic correlation is predominantly performed “manually” by geoscientists. The process is labor-intensive and time-consuming, and interpretations may vary among interpreters due to differences in expertise, experience, and perspective.
Recent advancements in the application of the Dynamic Time Warping (DTW) algorithm have demonstrated its potential to automate and enhance the stratigraphic correlation of well logs. DTW can generate multiple correlation scenarios highlighting different interpretations of subsurface continuity. Thus, the aim of this work is to explore the potential of DTW as a supporting tool in the standard workflows of geoscientists and test it on well log data from IODP Expedition 381 from the Gulf of Corinth. We automatically correlate lithostratigraphic subunits within a 700 m thick stratigraphic unit across two wells, using Natural Gamma Ray (NGR) and Magnetic Susceptibility (MAGS) logs. We selected this dataset because it illustrates the evolution of geological interpretations over time. Between the first version of the IODP data interpretations and a second version published a few years later, significant differences in interpretation were proposed. These differences highlight the critical role of geological expertise in refining subsurface data interpretations and correlations.
The automatic correlations interpreted by DTW showed a minimal average absolute difference with the most recent and updated published correlation, making the human and the machine correlation almost identical. By applying DTW to this dataset, we demonstrate it would have been possible to identify discrepancies and challenges in the interpretations of subunits at the initial stages after data acquisition. This approach could have flagged potential issues even before the IODP data were made available on the public site. Such early identification highlights the potential of DTW as a valuable tool for providing immediate feedback and guiding more accurate stratigraphic interpretations faster.
While DTW significantly reduces the time required for the correlation phase, the time investment needed for data formatting upstream should not be underestimated. Future work on larger datasets will be crucial to better quantify and validate the overall time savings provided by DTW, as well as to optimize the preparatory steps to ensure efficiency in broader applications.
In conclusion, we show that DTW can offer innovative approaches to enhance geological investigations and speed up interpretations. More generally, we consider this work illustrates how data science methods can be leveraged to assist geologists in routine tasks, with our Corinth case study highlighting both the promises and current limitations of digital transformation in well correlations.
How to cite: Christ, A.-B., Hamieh, Z., Armitage, J., Divies, R., Rohais, S., Mattioni, L., and Bouziat, A.: Dynamic Time Warping algorithm: A geoscience aware AI for automatic interpretation in lithostratigraphy? Insights from an application to the Gulf of Corinth (Greece), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9590, https://doi.org/10.5194/egusphere-egu25-9590, 2025.
Geological discontinuities significantly influence rock mass behaviour. Understanding the origin, setting, and properties of discontinuities is of major relevance, especially in boreholes. Traditionally, manual interpretation of borehole logs is done by geologists, a process that is time-consuming, costly, and subject to variability based on the interpreter's expertise. Recent advancements in artificial intelligence have made it feasible to use machine learning models and automatically detect and differentiate various features in digital images. In this study, we employ a state-of-the-art semantic segmentation model to tackle domain-specific challenges, enabling the identification of discontinuity types (e.g., natural faults, fault zones) and rock mass behaviour features (e.g., breakouts, induced cracks). We applied the SegFormer semantic segmentation model, which integrates a hierarchically structured transformer encoder with a multilayer perceptron (MLP). The borehole data used in this study was collected from the Mont Terri underground rock laboratory. Specifically, we labelled several high-resolution optical logs from one borehole and divided the dataset into training and testing subsets. The borehole considered is an experimental borehole designed to investigate the spatial and temporal evolution of damage around an underground opening in faulted clay shale. Our strategy achieved robust and accurate segmentation results on borehole images. Following segmentation, post-processing techniques were employed to extract critical information such as the total length of induced cracks and the total area of breakouts, as well as their locations and frequencies. The experimental results demonstrate high performance, with the pixel accuracy of 96 % in under three minutes for a 10-meter borehole. Our study lays the groundwork for future research by introducing a powerful tool for extracting geological structures and demonstrating the potential of AI models in geological analysis. By reducing processing time and increasing consistency in the identification, mapping, and classification of geological features, our approach can reveal spatial and temporal patterns associated with the evolution of rock masses.
How to cite: Wang, R., Ziegler, M., Volpi, M., and Manconi, A.: Advances in the identification of geological discontinuities in boreholes with deep learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4298, https://doi.org/10.5194/egusphere-egu25-4298, 2025.
Accurate multivariate parametrization of subsurface properties is essential for subsurface characterization and inversion tasks. Deep generative models, such as variational autoencoders (VAEs) and generative adversarial networks (GANs), are known to efficiently parametrize complex facies patterns. Nonetheless, the inherent complexity of multivariate modeling poses significant limitations to their applicability when considering multiple subsurface properties simultaneously. Presently, diffusion models (DM) offer state-of-the-art performance and outperform GANs and VAEs in several tasks of image generation. In addition, training is much more stable compared to the training of GANs. In this work, we consider score-based DMs in multivariate geological modeling, specifically for the parametrization of categorical (facies) and continuous (acoustic impedance – IP) distributions, focusing on a synthetic scenario of sand channel bodies in a shale background. We benchmark modeling performance against results obtained by GAN and VAE networks previously proposed in literature for multivariate modeling. As for the GAN and VAE models, the DM was trained with a training dataset of 3000 samples, consisting of facies realizations and co-located IP geostatistical realizations. Overall, the trained DM shows significant improvements in modeling accuracy, for all evaluation metrics considered in this study, except for the sand-to-shale ratio, where the values are comparable to those of the GAN and VAE. In particular, the DM is 26% more accurate at reproducing the average (nonstationary) facies distribution and up to 90% more accurate at reproducing the IP marginal distributions for both sand and shale classes. Higher accuracy is also found in the reproduction of the facies-to-IP joint distribution, whereas the spatial IP distributions generated by the DM honour the two-point statistics of the training samples. The iterative generative process in DMs generally makes these networks more computationally demanding than VAEs and GANs. However, we demonstrate that with appropriate network design and training parametrization, the DM can generate realizations with significantly fewer sampling iterations while maintaining accuracy comparable to these benchmarking networks. Finally, since the proposed DM parametrizes the joint prior probability density function with a Gaussian latent space, it is straightforward to perform inversion. In addition to improved modeling accuracy, the mapping between the latent and image representations preserves a better topology than that of GANs, overcoming the well-known limitation of the latter for inference tasks, particularly for gradient-based inversion.
How to cite: Miele, R. and Linde, N.: Multivariate generative modelling of subsurface properties with diffusion models , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7000, https://doi.org/10.5194/egusphere-egu25-7000, 2025.
Artificial Intelligence (AI) is revolutionizing geosciences, aligning perfectly with the themes of session GI2.4 by enabling the analysis of complex, multidimensional datasets and delivering actionable insights at unprecedented scales. This study presents two innovative AI-driven frameworks addressing critical challenges in soil moisture prediction and geomagnetic disturbance forecasting. Both approaches leverage decentralized networks to achieve scalability, foster collaboration, and enhance model performance through continuous refinement by a distributed community of contributors.
For soil moisture prediction, our multi-stream base model integrates data from Sentinel-2 (high-resolution spectral imagery), SMAP L4 (volumetric water content), ERA5 (meteorological variables), and SRTM (elevation data). The model predicts surface and rootzone soil moisture with six-hour lead times, achieving RMSE values of 0.1087 m³/m³ and 0.1183 m³/m³, respectively, across diverse Köppen-Geiger climate zones. By utilizing a decentralized network, contributors perform inference on 100 km² global regions, generating predictions evaluated against SMAP data using Root Mean Square Error (RMSE) and R² metrics. This system ensures robust model performance while addressing the spatial and temporal gaps inherent in traditional observational networks. These advances have significant implications for agriculture, hydrology, and climate modeling, enabling better water resource management, crop planning, and drought mitigation strategies.
In geomagnetic disturbance forecasting, our GeoMagModel leverages Prophet, a time-series forecasting library, to predict the Disturbance Storm Time (Dst) index, a key indicator of geomagnetic activity. The model achieves an RMSE of 6.37 for December 2024 datasets, effectively capturing both trend shifts and weekly seasonality. The decentralized community enhances predictive accuracy by dynamically integrating historical and real-time Dst data, which is validated by benchmark predictions of the Kyoto World Data Center’s hourly records. This approach provides near-real-time forecasts critical for safeguarding power grids, satellite systems, and other infrastructure vulnerable to space weather events.
By integrating machine learning with decentralized computing and state-of-the-art data sources, these frameworks offer scalable solutions to longstanding challenges in geophysical monitoring. The decentralized network not only improves scalability but also incentivizes the geoscience community to refine baseline models, fostering innovation and enabling systems to outpace state-of-the-art benchmarks. The implications of this work extend beyond immediate applications, paving the way for hybrid models that combine AI-driven predictions with physical process-based simulations. This fusion has the potential to improve understanding and resilience in critical domains such as water resource planning, disaster mitigation, and space weather forecasting. By addressing the limitations of traditional observation systems and delivering actionable insights at scale, these AI-driven frameworks represent a paradigm shift in how we approach and solve complex geoscientific problems.
How to cite: Moraga, G., Hristopoulos, S., and Pearson Kramer, N.: Harnessing AI and Decentralized Networks for Next-Generation Geophysical Forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13315, https://doi.org/10.5194/egusphere-egu25-13315, 2025.
Artificial intelligence (AI) is transforming landscape archaeology by enabling the automated analysis of high-resolution datasets, such as airborne laser scanning (ALS). The Automatic Detection of Archaeological Features (ADAF) tool is an example of the potential of AI to streamline the identification of subtle surface features and demonstrate their value in uncovering and understanding archaeological landscapes. By improving the detection of archaeological sites, the ADAF plays a crucial role in the research, management and preservation of cultural heritage.
ADAF uses advanced AI models, including convolutional neural networks (CNNs) for semantic segmentation and object detection, to detect features in ALS datasets. The tool has been trained on a large archive of ALS data from Ireland and processes visualised inputs to detect patterns indicative of archaeological structures. The workflow integrates pre-processing with the Relief Visualisation Toolbox, inference with trained AI models and post-processing to refine the results to ensure reliable outputs with minimal false positives.
Designed with accessibility in mind, ADAF features an intuitive user interface that removes the barriers traditionally associated with AI-driven analyses. Users can process ALS data and export GIS-compatible results without the need for specialised knowledge, making the tool suitable for a wide audience. This approach democratises the use of AI in landscape archaeology and extends its utility to professionals and researchers in the field.
Tests with Irish ALS datasets have shown that ADAF is able to detect both known and previously unrecognised archaeological features in the landscape, while enhancing the spatial accuracy of identified sites. By automating complex data analysis, ADAF underlines the efficiency, precision and scalability of AI in landscape archaeology. In addition, the tool contributes to the preservation of cultural heritage by identifying sites that would otherwise remain undiscovered and enabling their preservation and integration into cultural heritage management strategies.
ADAF represents a significant advance in the application of AI in landscape archaeology, providing a powerful and accessible solution for surface feature recognition. Its development underlines the transformative potential of AI to revolutionise the study and interpretation of archaeological landscapes.
How to cite: Coz, N., Kokalj, Ž., Curran, S., Corns, A., Kocev, D., Kostovska, A., Davis, S., and O'Keeffe, J.: Advancing Landscape Archaeology with AI-driven insights from Airborne Laser Scanning data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9968, https://doi.org/10.5194/egusphere-egu25-9968, 2025.
Accurate estimation of Particle Size Distribution (PSD) in rock avalanche deposits is essential for understanding the fragmentation processes and spatial distribution characteristics during mass movement. However, traditional methods, such as physical sieving or visual field estimation, are time-consuming, labor-intensive, and impractical for large-scale field measurements. To address these limitations, this study presents an automated PSD estimation framework that combines UAV imagery and deep learning-based segmentation. A synthetic dataset was used to train the segmentation model, improving its robustness across different scenarios. Image resolution adjustments were applied to improve detection accuracy for small and overlapping particles. Additionally, Fourier analysis was utilized to reconstruct smooth and continuous particle contours, to effectively handle overlapping particles. The reconstructed 2D outlines were further used to estimate 3D particle volumes through the shape-volume model based on laboratory and literature data. Projection correction was applied to mitigate image distortions to ensure precise volume predictions. The proposed approach overcomes the limitations of traditional methods dealing with complex particle distributions in real field environments. The results demonstrate the effectiveness of the proposed method for large-scale particle detection and volume estimation, providing new insights into rock avalanche fragmentation dynamics.
Keywords: Rock avalanche; Particle size distribution (PSD); deep learning; UAV Imagery
How to cite: Lin, R., Jaboyedoff, M., Derron, M.-H., and Lu, T.: Automated Particle Size Distribution Estimation of Rock Avalanches using Deep Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12242, https://doi.org/10.5194/egusphere-egu25-12242, 2025.
Small satellite platforms are increasingly used for Earth observation due to their cost-effectiveness and flexibility. However, their limited payload size often results in reduced spatial resolution of captured images. In our work, we address this challenge by proposing an advanced multi-image super-resolution (MISR) approach tailored for small satellite applications.
It integrates:
- Sub-pixel image registration and on curvelet transform-based interpolation to preserve high-frequency details while reducing artifacts;
- A novel hybrid method called SP-MISR (Subpixel Multi-Image Super-Resolution), which leverages Convolutional Neural Networks (CNNs) for local detail analysis and Transformers for global spatial relationships.
Our experimental results demonstrate that this combined approach significantly improves image sharpness, preserves fine details, and reduces artifacts, outperforming traditional super-resolution techniques. Moreover, SP-MISR exhibits robustness in processing noisy and distorted images, making it particularly suitable for the constrained imaging systems of small satellites.
Future developments will focus on improving computational efficiency, reducing interpolation errors, and extending the method to multi-spectral imaging and interplanetary missions, by exploring explore pure deep learning techniques.
This work highlights the potential of integrating traditional and deep learning methodologies to enhance image quality, thus expanding the scientific and operational capabilities of small satellite missions.
How to cite: De Martino, C., Della Corte, V., Inno, L., Cozzolino, F., Ruggiero, G., Da Deppo, V., Zuppella, P., Moualla, L., and Venafra, S.: Advanced Super-Resolution Techniques for Optical Payloads in Earth Observation: Combining Traditional and Deep Learning Methods, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15949, https://doi.org/10.5194/egusphere-egu25-15949, 2025.
Soil carbonates are critical players in the global carbon cycle and have a profound influence on soil health and agricultural productivity. Their quantification is also central to carbon sequestration efforts, where accurate measurement of soil carbonates can inform strategies for reducing atmospheric CO2. However, conventional methods of carbonate analysis in soil---while effective---are often slow, costly, and labor-intensive (1).
In this study, we introduce Shifted Excitation Raman Difference Spectroscopy (SERDS) as a rapid, non-destructive alternative, further enhanced by advanced preprocessing techniques and machine learning algorithms. Specifically, we employ Asymmetric Least Squares (ALS) for background correction, Standard Normal Variate (SNV) for normalization, and Savitzky–Golay filtering for smoothing. Unlike conventional Raman spectroscopy, SERDS effectively eliminates background fluorescence and reduces overlapping peaks, resulting in clearer spectral signatures (2).
We employed Partial Least Squares Regression (PLSR) and eXtreme Gradient Boosting (XGBoost) to predict the inorganic carbon content from the carbonate vibrational modes in conventional Raman and SERDS spectra, benchmarked against total inorganic carbon (TIC) measurements from coulometric titration. Our results show that switching to dual-laser SERDS substantially boosted model performance. For PLSR, the coefficient of determination (R2) improved from 0.8 to 0.88 (an increase of about 10.5%), and the root-mean-square error (RMSE) declined from 0.29 to 0.22 (26% decrease). The XGBoost model exhibited an even greater increase, with R2 increasing from 0.63 to 0.93 (approximately 49% improvement) and RMSE dropping from 0.39 to 0.16 (59% reduction).
Figure 1: Left: All SERDS data of soil samples showing the main carbonate peak; Right: XGBoost model prediction of soil inorganic carbon using SERDS data.
These findings underscore the potential of SERDS to replace conventional methods for carbonate quantification, offering reduced cost, faster analysis, and essentially no sample preparation. Furthermore, by providing highly accurate carbonate measurements, this methodology can be pivotal for carbon sequestration assessments and large-scale soil management practices, helping to advance both environmental sustainability and agricultural productivity.
References:
1) Barra I, Haefele SM, Sakrabani R, Kebede F. Soil spectroscopy with the use of chemometrics, machine learning and pre-processing techniques in soil diagnosis: Recent advances–A review. TrAC Trends in Analytical Chemistry. 2021 Feb 1;135:116166.
2) Orlando A, Franceschini F, Muscas C, Pidkova S, Bartoli M, Rovere M, Tagliaferro A. A comprehensive review on Raman spectroscopy applications. Chemosensors. 2021 Sep 13;9(9):262.
How to cite: Poursorkh, Z., Solomatova, N., and Grant, E.: Quantifying Soil Inorganic Matter: Integrating Shifted Excitation Raman Difference Spectroscopy (SERDS) with Machine Learning for Enhanced Analysis of Carbonates, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16705, https://doi.org/10.5194/egusphere-egu25-16705, 2025.
Tomographic methods, such as electrical resistivity tomography (ERT), induced polarization (IP) and seismic refraction tomography (SRT) are often effective for detecting geophysical targets in disparate real-world scenarios. However, a final reconstruction expressed only in terms of individual geophysical parameters (resistivity, chargeability, P-wave velocity) leaves room for ambiguity in complex sites exhibiting several transitions between layers or zones having different geophysical properties. In such cases, the sensitivity of the geophysical parameters for the various methods can differ significantly, so that a univocal interpretation based only on a visual comparison of the different models is often ineffective. To overcome these limits, in this work we present a machine learning-based quantitative approach for the detection of geophysical targets associated with both geological and anthropogenic scenarios. We integrate two-dimensional ERT, IP and SRT tomographic data with a soft clustering analysis by the Fuzzy C-Means (FCM) to obtain a final combined section, where each pixel is characterized by a cluster index and an associated membership value. The membership function of the Fuzzy C-Means is a good estimator of the accuracy of the subsurface reconstruction, as it ranges from 0 to 1, with 1 reflecting a high reliability of the clustering analysis. We apply this method to two case studies, related to the detection of leachate accumulation areas in a municipal solid waste landfill and to the bedrock characterization in a site prone to instability. In both cases, we detect the cluster associated with the geophysical targets of interest and our final sections are validated by a good agreement with the available direct information (boreholes and wells). The accuracy of the reconstruction is consistently high across most areas (membership values > 0.75), even though it is reduced in areas where the resolution of geophysical data is lower. Therefore, this approach may be a valuable automatic tool for optimizing the cost-effectiveness of projects where new constructions or remediation interventions have to be planned.
How to cite: De Donno, G., Cercato, M., Melegari, D., Paoletti, V., Penta de Peppo, G., and Piegari, E.: Fuzzy clustering of electrical and seismic data for the detection of geophysical targets, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17957, https://doi.org/10.5194/egusphere-egu25-17957, 2025.
Posters virtual: Tue, 29 Apr, 14:00–15:45 | vPoster spot 4
EGU25-18933 | Posters virtual | VPS19
Deep Learning-based Spatial-Spectral Analysis for Peatland Degradation characterizationTue, 29 Apr, 14:00–15:45 (CEST) | vP4.3
The study explores using advanced deep learning (DL) techniques for spatial-spectral analysis to detect and map peatland degradation at a granular level. Peatlands, vital carbon sinks in global ecosystems, face degradation threats that demand precise and scalable monitoring solutions. Our method combines convolutional neural networks (CNNs), fully convolutional networks (FCNs), and 3D CNNs to examine complex spatial-spectral patterns in SAR, multispectral, and hyperspectral sensor data (e.g., Sentinel-1, Sentinel-2, PRISMA) over the temperate peatland study area of the Monte Bondone region (Latitude: 46°00’48.6” N, Longitude: 11°03'14.6” E), covering an area of 40 hectares as shown in the figures.
CNNs capture spatial relationships between precipitation, temperature, vegetation, soil, and moisture, offering a detailed view of peatland composition. Using multi-dimensional, gridded data from meteorological stations and remote sensing images, CNNs identify patterns affecting peatland health. Fully Convolutional Networks (FCNs) help with spectral unmixing, isolating land cover components at the pixel level, which aids in detecting vegetation degradation and understanding ecosystem changes.
3D CNNs incorporate temporal data to classify Peatland landscapes into different degradation states. The model identifies changes over time, distinguishing between healthy, partially degraded, and fully degraded regions. Deep clustering models also classify peatland areas into degradation states, revealing trends without labeled data.
This deep learning framework supports accurate degradation mapping through spatial-spectral feature extraction, providing precise, pixel-level information to aid ecosystem management and conservation. It helps monitor peatland health and assess environmental changes across diverse landscapes.
How to cite: Kaparthi, H. V. and Vitti, A.: Deep Learning-based Spatial-Spectral Analysis for Peatland Degradation characterization, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18933, https://doi.org/10.5194/egusphere-egu25-18933, 2025.
EGU25-13917 | ECS | Posters virtual | VPS19
Automated Mineral Grain Extraction for Geometallurgical Studies Using Segment Anything Model (SAM) and Core Scanning TechniquesTue, 29 Apr, 14:00–15:45 (CEST) | vP4.23
In mineral exploration and geometallurgical studies, accurately segmenting mineral grains from core scanning datasets may be used to predict metal recovery. This study introduces the application of the Segment Anything Model (SAM), a cutting-edge deep learning tool, to automate the segmentation and extraction of mineral grains from Laser-Induced Breakdown Spectroscopy (LIBS) and hyperspectral core scanning datasets. SAM demonstrates high efficiency and precision in identifying mineral grains, forming the foundation for downstream analyses, including the evaluation of mineral associations, grain size distribution, and other key geometallurgical metrics. Through case studies on pegmatite deposits, this research showcases the potential of SAM to address challenges posed by mineralogically complex ore. By enabling detailed mineralogical characterisation and advancing geometallurgical methods, SAM-based grain extraction presents a transformative tool for supporting sustainable and efficient mining practices.
How to cite: Cai, Y., Manton, R., and Williams, M.: Automated Mineral Grain Extraction for Geometallurgical Studies Using Segment Anything Model (SAM) and Core Scanning Techniques, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13917, https://doi.org/10.5194/egusphere-egu25-13917, 2025.
EGU25-16363 | ECS | Posters virtual | VPS19
Hybrid Machine Learning approach for Tropical Cyclones DetectionTue, 29 Apr, 14:00–15:45 (CEST) | vP4.24
Tropical Cyclones (TCs) are among the most impactful weather phenomena, with climate change intensifying their duration and strength, posing significant risks to ecosystems and human life. Accurate TC detection, encompassing localization and tracking of TC centers, has become a critical focus for the climate science community.
Traditional methods often rely on subjective threshold tuning and might require several input variables, thus making the tracking computationally expensive. We propose a cost-effective hybrid Machine Learning (ML) approach consisting in splitting the TC detection into two separate sub-tasks: localization and tracking. The TC task localization is fully data-driven: multiple Deep Neural Networks (DNNs) architectures have been explored to localize TC centers using a different set of input fields related to the cyclo-genesis, aiming also at reducing the number of input drivers required for detection. A neighborhood matching algorithm is then applied to join previously localized TC center estimates into potential trajectories over time.
We train the DNNs on 40 years of ERA5 reanalysis data and International Best Track Archive for Climate Stewardship (IBTrACS) records across the East and West North Pacific basins. The hybrid approach is then compared with four state-of-the-art deterministic trackers (namely OWZ, TRACK, CNRM and UZ), reporting comparable or even better results in terms of Probability of Detection and False Alarm Rate, additionally capturing the interannual variability and spatial distribution of TCs in the target domain.
The resulting hybrid ML model represents the core component of a Digital Twin (DT) application implemented in the context of the EU-funded interTwin project.
How to cite: Donno, D., Accarino, G., Elia, D., Scoccimarro, E., and Gualdi, S.: Hybrid Machine Learning approach for Tropical Cyclones Detection, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16363, https://doi.org/10.5194/egusphere-egu25-16363, 2025.
