Recent advances in artificial intelligence (AI) techniques have enabled the processing and analysis of vast datasets, such as archives of satellite observations. In the geosciences, remote sensing has transformed the way in which the atmosphere and surface are observed. Traditionally, substantial funding is directed towards the development of new satellites to improve observation accuracy. Nowadays, novel methods based on AI could become a complementary approach to further enhance the resolution of observations. Therefore, we developed a new, state-of-the-art super-resolution methodology.

Satellites commonly measure electromagnetic radiation, reflected or emitted by the earth's surface and atmosphere, in different parts of the spectrum. Many instruments capture both panchromatic (PAN) and low-resolution multi-spectral (LRMS) images. While PAN typically covers a broad spectral range, LRMS focuses on details in narrow bands within that range. Pansharpening is the task of fusing the spatial details of PAN with the spectral richness of LRMS, to obtain high-resolution multi-spectral (HRMS) images. This has proven to be valuable in many areas of the geosciences, leading to new capabilities such as detecting small-sized marine plastic litter and identifying buried archaeological remains. Although HRMS images are not directly captured by the satellite, they can provide enhanced visual clarity, uncover intricate patterns and allow for more accurate and detailed analyses.

Technically, pansharpening is closely related to the single image super-resolution task, where attention-based models have achieved excellent results. In our study a new Cross Band Transformer (CBT) for pansharpening was developed, incorporating and adapting successful features of vision transformer architectures. Information sharing between the panchromatic and multi-spectral input streams was enabled through two novel components: the Shifted Cross-Band Attention Block and the Overlapping Cross-Band Attention Block, implementing mechanisms for shifted and overlapping cross-attention. Each block led to a more accurate fusion of panchromatic and multi-spectral data. For evaluation, CBT was also compared to seven competitive benchmark methods, including MDCUN, PanFormer and ArbRPN. Our model produced state-of-the-art results on the widely used GaoFen-2 and WorldView-3 pansharpening datasets. Based on peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) scores of the generated images, CBT outperformed all benchmark methods. Our AI method can be integrated in existing remote sensing pipelines, as CBT converts actual observations into a high-resolution equivalent for use in downstream tasks. A PyTorch implementation of CBT is available at https://github.com/VisionVoyagerX/CBT.

Furthermore, we developed the Sev2Mod dataset, available at https://zenodo.org/record/8360458. Unlike conventional benchmark datasets, Sev2Mod acquired input and target pairs from two different satellite instruments: (i) SEVIRI onboard the Meteosat Second Generation (MSG) satellite in geostationary orbit and (ii) MODIS onboard the Terra satellite in polar, sun-synchronous orbit. SEVIRI measures a fixed field of view quasi-continuously, while MODIS passes only twice a day but observes at a much higher spatial resolution. Our study investigated image generation at the spatial resolution of MODIS, while preserving SEVIRI's high temporal resolution. Since Sev2Mod is better aligned with actual situations one may encounter in applications of pansharpening methods (e.g., noise, bias, approximate temporal matching), it provides a solid foundation to design robust pansharpening models for real-world applications.

How to cite: Wijnands, J. S., Ntantis, N., Meirink, J. F., and Dibenedetto, D.: Super-resolution for satellite imagery: uncovering details using a new Cross Band Transformer architecture, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-340, https://doi.org/10.5194/egusphere-egu24-340, 2024.

In this study, we propose a method for monitoring the surface area of agricultural reservoirs in South Korea using Sentinel-1 synthetic aperture radar (SAR) images and deep learning models. This approach includes verifying the correlation between water surface area and water level, using data from both the monitored water surface area and real-time water level gauges. Leveraging the Google Earth Engine (GEE) platform, we constructed datasets for seven reservoirs, each with capacities of 700,000 tonnes, 900,000 tonnes, and 1.5 million tonnes, covering the period from 2017 to 2021. The model training was conducted on 1,283 images from four reservoirs, applying shuffling and 5-fold cross-validation techniques. The models' detection results were evaluated based on mean Intersection over Union (mIoU). Utilizing the highest-performing model, we analyzed the correlation between surface area and water level changes from 2017 to 2021. By integrating the water surface area data calculated by the model with real-time reservoir water level information from RAWRIS (Rural Agricultural Water Resource Information System), we confirmed the correlation between changes in water surface area and water levels from 2017 to 2021. This study illustrates that monitoring of water surface areas by satellite can be effectively utilized for tracking status changes in agricultural reservoirs in South Korea.

How to cite: Choi, S. and Lee, Y.: Waterbody Detection of Korean Reservoirs from Sentinel-1 Images and the Analysis of its Relationship with Water Level: A Deep Learning Approach, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16211, https://doi.org/10.5194/egusphere-egu24-16211, 2024.

This work explores the potential of content-based image retrieval to enable efficient search through vast amounts of satellite data. Images can be identified across multiple semantic concepts without needing specific annotations. We propose to use Geospatial Foundation Models (GeoFM), for remote sensing image retrieval and evaluated the models on two datasets. The GeoFM named Prithvi uses six bands and outperforms other RGB-based models by achieving a mean Average Precision of 61% on ForestNet-4 and 98% on BigEarthNet-19. The results demonstrate that the model efficiently encodes multi-spectral data and generalizes without requiring further fine-tuning. Additionally, this work evaluates three compression methods: i) binary embeddings, ii) trivial hashing, and iii) locality-sensitive hashing. Compression with binarized embeddings isthe best option for balancing retrieval speed and accuracy. It matches the latency of much shorter hash codes while maintaining the same accuracy as floating-point embeddings.

How to cite: Brunschwiler, T., Blumenstiel, B., Moor, V., and Kienzler, R.: Geospatial Foundation Models for Efficient Retrieval of Remote Sensing Images, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18300, https://doi.org/10.5194/egusphere-egu24-18300, 2024.

Underground mining activities cause ground deformations that threaten the stability of surface infrastructure and ecosystems, posing risks to both the environment and human populations. In response to these threats, this study focuses on developing a method for forecasting and modelling ground deformations, which are a consequence of underground mining activities, using advanced artificial intelligence (AI) techniques. The primary goal is to create a model utilising data from Differential Interferometric Synthetic Aperture Radar (DInSAR) and specialised mining data, enabling precise monitoring and forecasting of future changes which will support an adequate upgrade of the decision-making procedure in the mining industry.

The study employed two categories of neural networks: Convolutional Neural Networks (CNN) and Feedforward Neural Networks (FNN). In the application FNN, a detailed analysis was conducted on a per-pixel basis across the entire dataset. Each pixel, representing a specific point on the terrain, was analysed with its associated feature vector. This vector comprised multiple attributes derived from the mining data and DInSAR images, effectively capturing the local characteristics of each point, such as its relative position, historical deformation patterns, and proximity to mining activities. For the CNN method, the study focused on exploring the impact of different kernel sizes on model performance. Kernels in CNNs are small matrices used to process data across the image, essential for extracting and learning features crucial for understanding and predicting ground deformations. Varying kernel sizes allow the network to capture different aspects of the data. Considered features included the distance from the centre of the subsidence basin and the mining face at different time intervals. In the context of forecasting, the use of high-quality data is crucial. Unfortunately, some DInSAR images exhibited noise, due in part to a lack of stable coherence and adverse atmospheric effects. A key aspect of the study was therefore the creation and testing of a classifier for the suitability of DInSAR images for forecasting purposes. The analyses showed that the developed classifier achieved an accuracy of 83%. The training data for the prediction study came from the Budryk-Knothe method. The network was tasked with reproducing the operation of this method while simultaneously predicting six days ahead. The models were evaluated based on the mean squared error (MSE) in the areas of the subsidence basin. The test set consisted of specially prepared and trimmed DInSAR images. The FNN-based solution achieved the best results. For this network, satisfactory accuracy was achieved in determining the direction of settlement, with an MSE of 0.12, corresponding to a percentage error of approximately 10% (5 cm for a subsidence of 50 cm).

The  results from the study highlight the significant potential of integrating AI techniques with advanced geodetic methods, opening new possibilities in monitoring the impact of mining on the environment. Future work may focus on further optimization of AI algorithms to increase forecasting accuracy over longer periods and in various geological and operational conditions.

How to cite: Balak, P., Tymków, P., and Bogusławski, P.: Ground Deformation Forecasting and Modeling in Mining Areas Using Artificial Intelligence Techniques, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19507, https://doi.org/10.5194/egusphere-egu24-19507, 2024.

The present study classifies echoes from meteorological and biological targets using dual-polarization Doppler weather radar data from the Next Generation Weather Radar (NEXRAD).  Preliminary results are presented using six key variables namely, Base reflectivity, Base velocity, Spectrum width, Differential reflectivity, Correlation Coefficient, and Differential Phase. A threshold-based filtering methodology was implemented for biological scatterers and heavy precipitation events. To automate the classification, machine learning algorithms were implemented. Multiple machine learning algorithms were implemented and fine-tuned for the highest classification accuracy. Through the integration of machine learning techniques with dual-polarization Doppler weather radar data, this research endeavors to contribute to the development of robust models capable of distinguishing multiple types of physical scatterers from each other.

How to cite: Agrawal, A., Pahuja, S., Neelatt, A., and Indu, J.: Classification of different physical scatterers in weather radar data using machine learning techniques, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-22089, https://doi.org/10.5194/egusphere-egu24-22089, 2024.

Minor crops are crucial for food security, especially due to their resilience against climate change related challenges (Renard & Tilman, 2019) . Consequently, accurate crop mapping is essential for monitoring policies seeking to incentivize minor crop production. However, the class imbalance problem in machine learning introduces a bias against these crops, leading to unfair classifications. This research aims to explore how this bias is mitigated through two main class imbalance correction approaches: sample balancing methods and cost-sensitive learning. Apart from investigating how these methods address the typical class imbalance problem, where there are simply less labelled samples of a specific class, we investigate how these methods can be used to address another level of bias, that created by omitted sensitive attributes. These are attributes such as parcel size, which are not explicitly considered by the classifier, yet significantly impact accuracy and contribute to the unfairness of the classification, as evidenced by notably lower accuracy for smaller parcels. By integrating these attributes into the class imbalance correction methods, we assess the potential for enhancing fairness. This approach is vital, as it corrects performance biases affecting specific sub-groups, which are not necessarily class dependent, thus addressing a critical but overlooked dimension of fairness in classification.

Utilizing the BreizhCrops dataset, we create sub-sampled datasets that represent a variety of class imbalance problems. This enables us to conduct an across-the-board comparison of the selected class imbalance correction techniques, providing insights that may help streamline future research looking to employ these techniques. For the classifier architecture, we select the transformer encoder, chosen for its greater performance among deep learning methods tested on the BreizhCrops dataset (Rußwurm et al., 2020).

This research contributes to the broader understanding of class imbalance correction in classification tasks, particularly for crop mapping, though the methods can also be applied in other GeoAI contexts. By evaluating sample balancing and cost-sensitive learning in varied contexts, we provide insights into optimizing classification tasks for fairness. Our work contributes to the development of responsible AI practices by offering valuable insights on how fairness can be enhanced across GeoAI applications.

Renard, D., & Tilman, D. (2019). National food production stabilized by crop diversity. Nature, 571(7764), 257–260. https://doi.org/10.1038/s41586-019-1316-y

Rußwurm, M., Pelletier, C., Zollner, M., Lefèvre, S., & Körner, M. (2020). BreizhCrops: A Time Series Dataset for Crop Type Mapping (arXiv:1905.11893). arXiv. https://doi.org/10.48550/arXiv.1905.11893

How to cite: Gorbunov, I., Gevaert, C., and Belgiu, M.: Optimizing crop type mapping for fairness , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19021, https://doi.org/10.5194/egusphere-egu24-19021, 2024.

Accurately inferring realistic subsurface structures according to different types of geophysical observations and interpretations remains a significant challenge as the morphology controls the subsurface flow and transport behavior. Hard data and soft data are most commonly used in subsurface characterization. Among them, hard data are observations obtained from subsurface measurements. Soft data are remotely sensed geophysical information and interpretations. The deep learning technique is able to automatically learn the nonlinear relationships and extract implicit features of prior samples. In this work, we propose a deep-learning-based multiple-condition fusion network (MCF-Net) to accurately characterize subsurface structures according to both hard and soft data. Two encoders are used to extract the deep implicit features of the hard and soft data in multiple-level representations. To effectively integrate the multiple-level representations of multiple data, multiple-condition fusion blocks are introduced to extract the representing characteristics. These blocks are also equipped for the reconstruction of heterogeneous structures. MCF-Net can accurately estimate subsurface structures using a variety of available observations. Experimental results prove the utility of MCF-Net in many fields of Earth science.

How to cite: Chen, Q. and Cui, Z.: Characterizing subsurface structures from hard and soft data with multiple-condition fusion neural network, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15426, https://doi.org/10.5194/egusphere-egu24-15426, 2024.

This study introduces a novel approach for estimating lateral river-aquifer exchanges by employing Physics Informed Neural Networks (PINNs). The methodology compares the predictive capabilities of neural networks with the physics-based modeling provided by MODFLOW's Horizontal Flow Barrier (HBF) package, implemented through FloPy. As a foundation, the HBF package in MODFLOW establishes a baseline model, serving as a benchmark for performance comparison.

The integrated model leverages observed data and the fundamental principles of hydrogeology, enabling a robust estimation of lateral exchanges. The synergy of PINNs and MODFLOW HBF enhances the model's adaptability to diverse hydrogeological conditions, providing accurate predictions of intricate river-aquifer interactions. The comparative analysis with the MODFLOW HBF package underscores the efficacy of the proposed approach, offering insights for improved water resource management and environmental decision-making.

How to cite: Bajpai, M., Mehulkumar Rajkumar, L., Mishra, S., and Gaur, S.: Estimation of Lateral River Aquifer Exchanges with Physics Informed Neural Networks, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10105, https://doi.org/10.5194/egusphere-egu24-10105, 2024.

Subduction plate boundary faults and splay faults in accretionary wedges are capable of generating some of the largest earthquakes and tsunamis on Earth. Owing to the complexity of geological structures and the inherent ambiguity in geophysical data, comprehensively characterizing potential thrust fault systems presents a considerable challenge. Current automated fault detection methods, primarily targeting normal faults, show limited efficacy in complex fault systems of subduction zones. Treating the task of fault detection as a binary image segmentation issue, we propose a supervised end-to-end fully convolutional neural network (U-Net) to automatically and accurately delineate thrust faults from seismic data. To circumvent the labour-intensive and potentially subjective manual labelling process required for model training, we have designed a workflow to efficiently auto-generate more than 10000 training pairs comprising both 2D synthetic seismic images and their corresponding labelled images of the thrust faults simulated in the seismic images. Each synthetic seismic image includes randomly undulating stratigraphic strata and faults with dip angles between 5 and 40 degrees, aiming to simulate realistic and varied geological structures and thrust fault features in subduction zone, which equipped the U-Net model to achieve a 91% accuracy rate in fault detection within the test dataset. The example from the Hikurangi subduction zone, New Zealand demonstrates that the U-Net trained by only synthetic data is superior to conventional automatic methods, such as unsupervised methods or supervised methods trained by normal faults, in delineating more than 70% thrust faults from seismic images. To enhance the U-Net model's adaptation to specific regional fault characteristics and reduce the interference from noise, we incorporated a select set of real 2D seismic images and manually interpreted fault labels into the transfer learning process, which significantly improved its prediction accuracy and make the results clearer. From the comprehensive 2D characterizations based on the U-Net model, we can further extract 3D thrust fault systems and quantitatively measure their geometric parameters.

How to cite: Zheng, W., Bell, R., John, C. M., and Guasch, L.: Deep Learning for Detecting Thrust Faults in Subduction Zones, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5905, https://doi.org/10.5194/egusphere-egu24-5905, 2024.

The integrity of well logging data is paramount in geophysical explorations for accurate subsurface analysis, notably in the North Sea Dutch region known for its extensive hydrocarbon exploration. Addressing the common challenge of missing data in well logs, our study introduces an AI-driven methodology employing generative models. These models utilize machine learning to analyze existing data patterns and generate realistic imputations for missing values. The approach has shown to not only enhance the quality of geological interpretations but also to streamline the workflow in hydrocarbon exploration. This integration of AI signifies a substantial move towards more precise and efficient geoscience data analysis. A qualitative comparison using Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE) was conducted to evaluate the results. The PCA comparison demonstrates the synthetic data’s alignment with real data in principal component space, effectively capturing the variance. The t-SNE analysis further validates the model's fidelity, with the synthetic data exhibiting clustering behaviors analogous to real data. Together, these results showcase the transformative potential of machine learning in geosciences, providing a robust framework for enhancing data reliability in geophysical studies.

How to cite: Al-Fakih, A., Koeshidayatullah, A., and Kaka, S.: Enhancing Geoscience Analysis: AI-Driven Imputation of Missing Data in Well Logging Using Generative Models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10627, https://doi.org/10.5194/egusphere-egu24-10627, 2024.

Water's impact on the physicochemical attributes of mantle rocks makes it a pivotal factor in mantle evolution. Mid-ocean ridge basalts (MORBs) are essential for analyzing the upper mantle's composition, yet many global MORB samples lack direct water content assessment. The common method, using a correlation between H₂O and trace elements like Ce to estimate MORB water contents, often presume a constant H₂O/Ce ratio. Sometimes this assumption is unreliable due to the heterogeneity in H₂O/Ce ratios, even within short ridge segments. For addressing this gap, we utilize compositional data from 1,467 global MORB glasses with measured water contents to develop a Random Forest Regression model. This machine learning-based model can predict water concentrations of MORB glasses based on major and trace element data, without the need for a fixed H₂O/trace element ratio. Our model accurately recovers water contents of MORB glasses, showing comparable precision to traditional analytical methods. Applying this model to 1,931 MORB glass samples has significantly expanded the global MORB water content database, revealing the widespread presence of high-water MORBs. Importantly, this innovative approach enables the exploration of water content in MORBs from regions previously without such data, like the Chile Ridge and the Pacific-Antarctic Ridge. Moreover, it allows us to deduce variations of water contents of MORB sources by applying the model to transform fault samples, thereby offering novel insights into the dynamics of the mantle.

How to cite: Zhou, J., Liu, J., Xia, Q., Su, C., Kuritani, T., and Hanski, E.: Recover the Water Content of Mid-Ocean Ridge Basalts by a Machine Learning Method, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7440, https://doi.org/10.5194/egusphere-egu24-7440, 2024.

Determining pressures and temperatures of magmas is crucial for addressing diverse challenges in petrology, geodynamics, and volcanology. However, inherent inaccuracies, especially in barometry, have limited the effectiveness of existing models in unravelling the architecture of crustal igneous systems. In this presentation, I will introduce a novel machine learning model, calibrated using an extensive experimental database, to create regression models for extracting P-T conditions of magmas. Calculations are conducted by considering melt chemistry and the coexisting mineral assemblage as input variables.
Our approach is versatile, applicable across a wide range of compositions from basalt to rhyolite, covering pressures from 0.2 to 15 kbar and temperatures ranging from 675 to 1400°C. Testing and optimization demonstrate that the model can recover pressures with a root-mean-square error of 1.1-1.3 kbar and temperature estimates with errors as low as 21°C. This indicates that melt chemistry-mineral assemblage pairs reliably capture magmatic variables across a broader spectrum of conditions than previously thought. We propose that this reliability arises from the relatively low thermodynamic variance in natural magma compositions, despite the presence of numerous oxide components.
Applying our model to two cases with well-constrained geophysics - Mount St. Helens volcano (USA) and the Askja caldera in Iceland - we analyse dacite whole-rocks from Mount St. Helens, erupted between 1980-1986. These rocks, inferred to represent liquids extracted from a complex mineral mush, yield melt extraction source pressures that align remarkably well with geophysical constraints. For Askja caldera, our model allows to assign basaltic and rhyolitic magma chemistries to distinct seismic wave speed anomalies, highlighting the potential of our model to bridge the gap between petrology and geophysics. Our model, named MagMaTaB, is accessible through a user-friendly web application.

How to cite: Weber, G. and Blundy, J.: MagMaTaB: A machine learning-based model for magmatic liquid thermobarometry, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-11007, https://doi.org/10.5194/egusphere-egu24-11007, 2024.

Estimating the water content of mantle-derived magma using clinopyroxene (cpx) phenocrysts serves as a valuable constraint on the water budget in deep Earth. Intricate magma processes and the high hydrogen diffusion rate necessitate careful evaluations of whether the water content in cpx preserves its original state. Machine learning (ML) has been utilized to develop a classifier for judging hydrogen diffusion in cpx. Never- theless, the opaqueness and complexity of most ML models hinder a clear understanding of their classification principles. To elucidate the mechanistic basis of the ML model, the Shapley theory is integrated to determine the contributions of major elements of cpx as features in a linear additive manner. This study achieves superior classification performance using an extreme gradient boosting model and innovatively presents a quantitative evaluation of feature importance at the sample level for each observation. The results indicate that Na plays a predominant role in the diffusion process surpassing other major elements and its associated hydrogen can easily diffuse out of cpx. Our model also identifies various hydrogen association modes in different elemental com- positions and puts constraints on the properties of incorporated hydrogen with non-lattice forming elements in cpx. The findings demonstrate that the application of explainable ML methods in mineralogy holds significant potential for advancing the comprehension of geological phenomena.

How to cite: Li, A., Wu, S., Chen, H., Du, Z., and Xia, Q.: Explainable machine learning to uncover hydrogen diffusion mechanism in clinopyroxene, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19015, https://doi.org/10.5194/egusphere-egu24-19015, 2024.

Machine learning has significantly advanced geochemistry research, but its implementation can be arduous and time-consuming. In response to this challenge, we introduce Geochemistry π, an open-source automated machine learning Python framework. With Geochemistry π, geochemists can effortlessly process tabulated data and execute machine learning algorithms by selecting preferred options. This streamlined process operates in a user-friendly question-and-answer format, eliminating the need for coding expertise. Following automatic or manual parameter adjustment, Geochemistry π furnishes users with comprehensive performance metrics and predictive outcomes for their machine learning models. Leveraging the scikit-learn library, Geochemistry π has developed a tailored automated workflow encompassing classification, regression, dimensionality reduction, and clustering algorithms. The framework’s extensibility and portability are enhanced through a modular pipeline architecture, segregating data handling from algorithm application. Geochemistry π’s Auto Machine Learning module integrates Cost-Frugal Optimization and Blended Search Strategy hyperparameter search methods from the A Fast and Lightweight Auto Machine Learning Library. Additionally, model parameter optimization is expedited using the Ray distributed computing framework. Efficient machine learning lifecycle management is facilitated through integration with the MLflow library, allowing users to compare multiple trained models at various scales and manage generated data and visualizations. To enhance accessibility, Geochemistry π separates front-end and back-end frameworks, culminating in a user-friendly web portal. This portal not only showcases the machine learning model but also presents the data science workflow, making it accessible to both researchers and developers. In summary, Geochemistry π offers a robust Python framework that empowers users and developers to significantly enhance their data mining efficiency, with options for both online and offline operation.

How to cite: Zhao, J., ZhangZhou, J., He, C., and Lyu, Y. and the ZJU Earth Data Group: Geochemistry π: Automated Machine Learning Python Framework for Tabular Data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5002, https://doi.org/10.5194/egusphere-egu24-5002, 2024.

Groundwater is a natural water source crucial in sustaining ecosystems and meeting various human needs. Groundwater is often contaminated due to the various anthropogenic and non-anthropogenic activities. Nitrate is the most abundant pollutant of Groundwater, which may be exogenic and anthropogenic. We studied nitrate ion concentration in Groundwater from dug well data. The Karnataka state's nitrate ion concentration varies from 0 to 1696 mg/l, which is higher in most places than the admissible limit of 45 mg/l as per the World Health Organisation (WHO). The correlation of various parameters, such as pH, electrical conductivity (EC), fluoride, chloride, etc., was studied with nitrate, and maximum correlation was found with chloride and EC.  Our prediction concentration of nitrate ion using Different Machine Learning (ML) algorithms, including Regression, Random Forest (RF), Support Vector Regression (SVR) and Decision Tree (DT) models using the input parameters as pH, EC chloride, and fluoride.  The result showcased that the best model is Support Vector Regression (SVR) with an R² value of 0.93 and a Mean Square Error (MSE) value of 0.02 for the region. The region's nitrate pollution might be forecast using the SVR model for better estimation.

How to cite: Bansal, H., Devarakonda, V., and Dixit, M.: Nitrate contamination prediction in Groundwater data in Karnataka, India, using Machine Learning (ML) Techniques, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-14857, https://doi.org/10.5194/egusphere-egu24-14857, 2024.

Underwater seismic events such as earthquakes are known to produce not only seismic waves but also hydro-acoustic waves. Indeed, seismic waves arriving at the ocean bottom convert into acoustic waves in the water column. Other events, such as hot lava-seawater interactions or icequakes, also generate water-born acoustic signals. Monitoring these different signals with moored hydrophones proved to be useful and very efficient thanks to the little attenuation of acoustic waves propagating in the Sound Fixing and Ranging (SOFAR) channel. This led to the deployment of wide-range moored hydrophone networks to monitor active seafloor-spreading ridges in the world ocean. However, analyzing year-round data recordings from several stations is a cumbersome, user-dependent and most importantly time-consuming task. Despite some efforts to develop automatic detection algorithms, the community still lacks efficient and available off-the-shelf tools, as well as open datasets and benchmarks against which they could be compared objectively. To address this problem, we are glad to make publicly available three partially-annotated hydroacoustics datasets consisting of recordings from Atlantic and Indian oceans, with a total of ~60,000 hours. We propose a benchmark of models on a first task of binary classification, and an original convolutional neural network (CNN) model called TiSSNet showing promising results. To maximize the reliability of the evaluations, two datasets have been carefully and exhaustively annotated to serve as evaluation datasets. The getting started codes have also been made available on GitHub. We wish the datasets and benchmarks will be used as references upon which the state-of-the-art could be developed in a collaborative way. In the future, the best model, used as an automatic or semi-automatic detection framework, will be applied to larger datasets, and combined with multi-stations association and trilateration techniques to output nearly complete catalogs of geophonic events (source type and location, with signal characteristics).

How to cite: Raumer, P.-Y., Bazin, S., Royer, J.-Y., Dorian, C., and Vijay Ingale, V.: Hydroacoustic geophony automatic detection: an open benchmark dataset with an open model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5256, https://doi.org/10.5194/egusphere-egu24-5256, 2024.

One of the main objectives of the European project GSEU (Geological Service for Europe) is to, starting from the EGDI (European Geological Data Infrastructure), develop a system for exchanging knowledge and skills in the geological and geothematic fields, as well as strengthening the development of data and information standards and harmonization frameworks.

To achieve this objective, it was decided to evolve EDGI towards a Geospatial Knowledge Infrastructure (EGDI-KI), for which the conceptual model and the first prototype under development are presented here, which were developed starting from the document developed by UN-GGIM in 2021 for Geospatial Knowledge Infrastructure.

Essential elements of the EGDI-KI are shown in figure 1 and are: The Knowledge Hub forms part of the EGDI structure and serves as a gateway to facilitate access to other components. It enhances the accessibility and usability of the following components:

• Data Hub: This component focuses on data exchange, access technologies, and supports data science, data engineering, and data warehouse endpoints. The Knowledge Hub helps streamline access to the data hub, making it easier for users to leverage data resources.
• Applications: The Applications component encompasses WebGIS or thematic portals designed to share information, data, and enable big data analysis. The Knowledge Hub contributes to the seamless integration and utilization of these applications, ensuring efficient access to information and facilitating analysis.
• Collaboration Tools: Collaboration Tools within EGDI enable the sharing of documents, models, and methods among users. The Knowledge Hub complements this by providing a platform for organizing and accessing these shared resources, fostering collaboration and knowledge exchange.
• Educational Facilities: EGDI includes educational facilities that support end-users and thematic domains in sharing and transferring knowledge. The Knowledge Hub plays a role in facilitating access to these educational resources, making them readily available to users seeking to enhance their understanding of relevant topics.
• Expertise & Networking Hub is thematic expert and physical Infrastructure catalogue as well as the possible research and Industry community interactions.

Finally, the knowledge infrastructure platform is the portal to query and navigate all the knowledge resources available in the Knowledge Hub.

The Knowledge Hub plays an important role in the European Geological Data Infrastructure (EGDI) by ensuring that the wealth of knowledge and expertise available within the system is not fragmented and disconnected. Instead, it enables the organisation and accessibility of this knowledge using a semantic Knowledge engine.

By leveraging the semantic Knowledge engine, the Knowledge Hub facilitates the integration and structuring of diverse pieces of information within the EGDI system. It allows for the establishment of meaningful connections and relationships between different data sources, ensuring a coherent and organized presentation of knowledge.

Through the Knowledge Hub, users can efficiently navigate and explore the EGDI system, accessing relevant information in a structured and interconnected manner. It enhances the overall usability and effectiveness of the system, enabling users to leverage the collective knowledge and expertise within the EGDI framework.

How to cite: Cipolloni, C., Šinigoj, J., Schiegl, M., Martin, Á. P., Tulstrup, J., and Gruijters, S.: The EGDI Knowledge Infrastructure – the next step of Geological data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20777, https://doi.org/10.5194/egusphere-egu24-20777, 2024.

Despite the increased availability of UAV / drone imagery in Low- to upper  Middle-Income countries and the demonstrated potential of deep learning to support the interpretation of these images for sustainable development purposes, practical operations in these countries are constrained by the need for sufficient labeled data-sets which are often difficult to obtain (especially for tropical forest). This makes it difficult to train suitable networks and assess whether the model is performing well. One such example is the use of drones to monitor the Atlantic Forest in Sao Paulo, Brazil. Here, members of the Sao Paulo Municipal Green and Environment Secretariat (Secretaria do Verde e do Meio Ambiente - SVMA) are starting to use drones to  identify some native and invasive species in their forests. Deep learning will quickly speed up this process, but there is little training data available. This refers to the so called ‘small-data problem’ commonly found in DL for remote sensing applications [1]. A workflow was designed to support this application through a novel zero-shot learning technique and explainable AI methods. A pre-trained tree-crown detection model ‘DeepForest’ [2] is used to identify individual tree crowns in the UAV imagery. The detected tree-crowns are further classified using a Siamese network architecture using zero-shot learning – the model is trained on relevant data-sets but not exposed to species found in the test data-set. A Siamese network architecture is motivated by the need for explainability in DL models – the results will be used for making administrative decision for forest management. A more intricate DL model (such as image segmentation) could be more accurate but at the cost of transparency/explainability. In particular, we apply a variation of the ‘What I Know’ (WIK) explainability method [3] which provides examples from the training set along with the test sample increasing transparency and understanding of the model results.

[1] Safonova, Anastasiia, et al. "Ten deep learning techniques to address small data problems with remote sensing." International Journal of Applied Earth Observation and Geoinformation 125 (2023): 103569.

[2] Weinstein, Ben G., et al. "DeepForest: A Python package for RGB deep learning tree crown delineation." Methods in Ecology and Evolution 11.12 (2020): 1743-1751.

[3] Ishikawa, Shin-nosuke, et al. "Example-based explainable AI and its application for remote sensing image classification." International Journal of Applied Earth Observation and Geoinformation 118 (2023): 103215.

How to cite: Chandramouli, P., Gevaert, C. M., Nattino, F., Ku, O., Aguiar Pedro, A., do Prado Oliveira, P., Hortal Pereira Barreto, E., and de Ovileira, F.: XAI for small-data problems in remote sensing: monitoring Atlantic forests with UAVs, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9192, https://doi.org/10.5194/egusphere-egu24-9192, 2024.

In Korea, when a disaster occurs, numerous news related to the disaster are reported very quickly through various media. These news contain useful disaster-related information that disaster researchers need, such as the causes of disasters, problems in the process of disaster occurrence, and improvement measures suggested by related experts. However, finding articles containing the disaster-related information we need from the numerous news reports in the media is not easy and takes a long time. Accordingly, in this study, the R-Scanner model using text mining technology was developed to extract disaster and safety information desired by users from large-scale news data, which is 'unstructured big data'. Here, R stands for Risk. The developed model was constructed based on natural language processing systems for Korean and English and was developed to perform Sentence Segmentation, Tokenization, and Morphological Analysis using text as analysis data. In the Morphological Analysis process, the model was developed to perform Entity Recognition, Semantic Role Labeling, and Semantic Chunking. Additionally, the model was developed to extract articles containing the desired information from news big data reported through the media when the user inputs keywords related to the desired information, and the extracted articles can be downloaded in Excel format. To verify the performance of the developed model, we applied it to landslides that resulted in 14 deaths due to torrential rains in Korea in 2023. Problems and improvement measures in the landslide occurrence process were set with the desired information, and keywords were set to extract each information. About 200 keywords related to problems were set, such as 'procrastination', 'defenseless', 'ignored', 'sloppy', and 'careless', and about dozens of keywords such as ‘suggested’, ‘should be prepared’, and ‘necessary’ were set as keywords related to improvement measures. As a result of applying the model, a total of 364 articles related to problems and improvement measures were extracted from 30 media news 15,911,665 articles, and as a result of grouping the extracted problems and improvement measures into similar contents, 24 problems and 22 improvement measures were finally derived. As a result of the review of related experts on the problems and improvement measures derived, it was confirmed that the contents were quite meaningful. The problems and improvement measures derived in this way were used as basic data for the establishment of government measures to prevent landslides. In the future, the developed model is expected to be used not only to establish the government's countermeasures for disaster, but also to monitor real-time disaster and safety issues, and furthermore to detect disaster risks at an early stage.

How to cite: Choi, S., Kim, D. W., Shin, E. H., and Kim, Y. J.: Development and Application of a Model to Extract Disaster and Safety-related Information from News Big Data reported in the Media using Text Mining, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6817, https://doi.org/10.5194/egusphere-egu24-6817, 2024.