This session explores both practical applications and conceptual frameworks of generative, agentic, and hybrid AI (inclusive of advanced ML/DL models) in the geosciences, with a strong emphasis on human-centred design. Topics include AI-assisted data analysis and modelling, knowledge discovery and curation, decision support, science communication, and AI-enhanced workflows guided by domain expertise. We particularly encourage contributions that demonstrate how human insight and causal reasoning complement AI.
The session also addresses ethical and societal aspects of AI deployment in geosciences, including transparency, bias mitigation, accountability, sustainability, and trustworthiness. The overarching goal is to highlight AI as a tool to empower and extend human expertise—keeping scientists and domain experts at the centre of innovation. Contributions from research, industry, and policy communities are all welcome.
AI-driven weather and climate prediction has become highly visible in recent years, and most Earth scientists are now familiar with learned forecast models. Far fewer are aware that the same advances in artificial intelligence are producing general-purpose systems that can autonomously review literature, write and debug code, design experiments, and carry out extended research tasks with minimal supervision. These capabilities may ultimately have a greater impact on everyday scientific practice than any single prediction model.
AI-based forecasting represents only a narrow entry point into a broader transformation driven by hybrid intelligence, in which domain-specific Earth system models are combined with general AI systems such as large language and multimodal models and autonomous agents. In practice, this hybrid intelligence already spans simulation, data assimilation, downscaling, and analysis, while general AI systems increasingly handle coding, synthesis, and workflow orchestration. Together, these systems function less as isolated tools and more as adaptive research partners. Drawing on examples from NVIDIA’s Earth-2 research program and related international efforts, this talk examines how this shift reconfigures the human role toward problem formulation, validation, interpretation, and ethical governance, and highlights practical AI-assisted workflows already reshaping research productivity. Framing AI for environmental prediction within this wider context invites a broader discussion of how hybrid intelligence should be integrated thoughtfully into future Earth system science.
How to cite: Hall, D.: Beyond the Forecast: Hybrid Intelligence as a Force Multiplier for Earth Science, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5926, https://doi.org/10.5194/egusphere-egu26-5926, 2026.
Data generated in the field of geoscience has unique properties that can be characterized by high complexity, sparsity, and site-specific variability. Owing to the unique characteristics, applying general artificial intelligence frameworks and achieving model generalization in geoscience still remains a challenging problem. In this work, we introduce the KIGAM GeoAI Platform, an integrated AI environment designed to bridge the gap between advancing AI technology and practical geoscience research. The platform supports the entire research workflow through a user-friendly, web-based interface, systematically covering essential stages: data uploading, preprocessing, model development, testing, validation, and the final deployment of analytical applications. By providing a centralized online environment for collaborative research, the platform aims to reduce technical entry barriers for geoscientists who may not be AI experts, while establishing a robust foundation for data-driven cooperation. We plan to continue improving and scaling this platform to ensure it remains a stable, accessible, and high-performance tool for both domestic and international geoscience communities. Through these efforts, the KIGAM GeoAI Platform is expected to accelerate digital transformation and foster a more integrated global research ecosystem in the field of geoscience.
How to cite: Kwon, J.: Innovate with Ease: Introducing the KIGAM GeoAI Platform, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2290, https://doi.org/10.5194/egusphere-egu26-2290, 2026.
Climate change, land-use change, and increasing socio-economic pressures are reshaping water and environmental systems, while the volume and heterogeneity of available data—from in situ observations to reanalysis products, remote sensing, and citizen-generated sources—continue to grow. Machine learning (ML) has become an important component of hydro-environmental modelling for forecasting, classification, and pattern discovery. However, in practice, many ML applications remain highly case-specific and dependent on implicit expert decisions related to problem formulation, predictor selection, validation design, and interpretation, which are rarely made explicit or transferable across regions and users.
This contribution presents a human-in-the-loop hybrid intelligence framework that integrates ML workflows with Large Language Models (LLMs) to support structured reasoning during environmental model development and evaluation. Rather than using LLMs for automated optimisation or model selection, the framework positions them as a guidance and scaffolding layer that helps make modelling assumptions, choices, and limitations explicit and traceable, while retaining expert control over all final decisions.
Methodologically, the framework combines (i) hands-on ML pipelines, ranging from baseline statistical models to more advanced learning algorithms for forecasting and classification, and (ii) an LLM-based guidance layer that structures expert reasoning through prompts, checklists, and decision logs. This guidance supports key stages of the modelling process, including the definition of modelling objectives, assessment of data quality, selection of environmentally meaningful predictors, and the design of validation strategies. Particular emphasis is placed on encouraging validation schemes that account for temporal dependence and spatial heterogeneity, such as blocked or spatial cross-validation, rather than default random data splits.
The framework is currently being developed and iteratively evaluated through expert-led case studies using real hydro-environmental datasets, rather than through formal classroom deployment. Initial applications focus on groundwater level analysis and hydro-environmental forecasting problems in Greece, including collaborative work in Crete, where the framework has been used to structure modelling choices and interpret model behaviour under non-stationary conditions. Additional exploratory applications using existing datasets have been used to stress-test the transferability of the workflow across contrasting environmental settings. Ongoing extensions include the application of the framework within coastal erosion modelling activities currently being developed in Colombia.
The LLM layer supports explicit reasoning about why a model performs well or poorly under specific conditions, how assumptions propagate into uncertainty, and where data-driven learning diverges from physical expectations. This reflective use of hybrid intelligence helps expose failure modes and modelling sensitivities that are often hidden in automated pipelines.
Results from the expert-led evaluations indicate that the proposed framework improves the transparency and reproducibility of modelling decisions, facilitates comparison across case studies, and supports more consistent interpretation of ML results across regions and scales. At the same time, the approach lowers the entry barrier for non-specialists without removing expert oversight or domain judgement.
The framework is being developed within the context of the Erasmus+ AI-LEARN project (Project reference: 2025-1-NL01-KA220-HED-000355215), where it serves as a methodological backbone for future training and capacity-building activities in water and environmental intelligence.
How to cite: Corzo P, G. A., Varouchakis, E., Kamińska-Chuchmała, A., Agioutanti, R., and Dominguez, V.: Integrating Machine Learning and Large Language Models for Next-Generation Water & Environmental Intelligence, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18439, 2026.
We present SeaScope, an explainable AI agent that accelerates interaction with complex Earth Observation (EO) workflows. Users express analytical questions in natural language, which are transformed into transparent, executable EO analyses. By combining generative AI, vision–language models, and Retrieval-Augmented Generation (RAG), SeaScope links scientific literature, satellite data descriptions, and validated analysis methods to automatically generate, execute, and explain EO workflows. For example, a query such as “Detect vessel activity and possible oil spills in May 2025” triggers dataset selection, code generation, cloud execution, and map outputs with traceable reasoning.
SeaScope is designed as a geoscience-specific AI agent that supports both rapid decision-making and accelerated research. Non-technical users can obtain EO-based insights in time-critical situations without continuous involvement of expert programmers, while researchers benefit from faster hypothesis testing, automated pipeline generation, and reproducible workflows. Human expertise remains central: users inspect retrieved sources, review generated code, and validate analytical steps, ensuring scientific control and accountability. This setup combines domain knowledge with AI-driven scalability, addressing challenges such as sensor-specific scripts and fragmented tools.
As a pilot use case, SeaScope is applied to maritime EO in the Mediterranean region, supporting environmental monitoring and marine activity analysis using satellite data. Beyond the application, the project delivers research insights on generative and vision-based AI for EO, including lessons learned from benchmarking LLMs for code generation, evaluating vision-language models for image understanding, and comparing different RAG and knowledge ingestion strategies. The findings highlight practical trade-offs in accuracy, robustness, explainability, and user validation in real-world workflows.
How to cite: Sekas, C., Philippopoulos, K., Agathangelidis, I., Cartalis, C., Neophytides, S., and Mavrovouniotis, M.: Agentic AI for Earth-Observation-Driven Maritime Monitoring - the SeaScope Project, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13454, https://doi.org/10.5194/egusphere-egu26-13454, 2026.
Typically, to start working on a remote sensing–based application, various analyses and insights are needed from domain experts. A significant amount of time and effort goes into preprocessing, structuring, and analyzing the data, which can be a repetitive task, especially when a multi-sensor approach is involved. This often takes away time that could otherwise be invested in innovation or research. To address this, training an LLM to understand and process the context of remote sensing tasks can improve efficiency and reduce human-induced errors.
In this work, we develop an AI agent that can reason and think like a remote sensing expert. This agent uses a RAG-based foundational model (FM) and is equipped with various image processing tools to complete a task. We use gpt-4.1-mini as the FM and the Agno framework to deploy the agent. The knowledge base provided to this agent is specially curated with relevant research articles, books, and remote sensing methodologies. This knowledge base helps the model break down a problem into logical steps that can be performed using the tools available within the agent.
These tools can download data, process it, and provide relevant statistics and visualizations. The user can prompt the agent to download multi-sensor (optical and SAR) data, perform time-series analysis for forest monitoring, and identify deforestation hotspots. The agent can fetch data from Google Earth Engine (GEE), plan processing workflows, dynamically generate Python code, and complete the prompted tasks. This approach highlights the feasibility of integrating LLMs with domain-specific knowledge bases and geospatial processing tools to create autonomous, context-aware systems. Figure 1 depicts the overall workflow of the proposed agentic system, illustrating the interaction between the user, the knowledge base, the foundational model, and the integrated processing tools. The framework is directly usable for operational forest monitoring applications and can be further fine-tuned and extended to support a broader range of environmental monitoring and geospatial analytics use cases.
Figure 1: Workflow of the Agentic AI system
How to cite: Jain, A. and Sabir, A.: Development of a Context-Aware AI Agent for Forest Applications Using Multi-Sensor Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10390, https://doi.org/10.5194/egusphere-egu26-10390, 2026.
Earth system science relies on the integration of knowledge from many branches of geoscience, including climate dynamics, hydrology, ecology, land use and biogeochemical cycles. However, the scientific literature informing these domains has become vast and increasingly difficult to navigate due to its rapid development and disciplinary spread. This complexity makes it difficult to maintain an integrated overview of relevant findings and to identify scientific connections in a systematic manner. Recent advances in generative artificial intelligence (AI) and large language models (LLMs) provide opportunities to support these tasks, particularly when combined with retrieval methods and transparent source attribution.
Here we propose a retrieval-augmented AI platform designed to assist scientific knowledge integration in Earth system science. The platform is conceived as a living system, built on a continuously expanding and updateable knowledge base that aggregates scholarly literature from major scientific databases. User queries initiate targeted retrieval of relevant documents followed by the generation of concise, source-linked summaries using locally hosted open-weighted LLMs. By explicitly grounding outputs in retrieved literature, the platform alleviates the need for manual screening and limits hallucination risks that currently constrain the use of general-purpose LLMs in geoscientific research.
Evaluation of the initial prototype demonstrates that domain-specific retrieval-augmented generation systems can provide reliable, traceable synthesis of Earth system knowledge and help address the growing gap between accelerating publication rates and the need for timely, verifiable scientific assessment.
How to cite: Kart Tokmak, Ö., Caesar, L., and Sakschewski, B.: A Living AI Platform for the Earth System Science, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3316, https://doi.org/10.5194/egusphere-egu26-3316, 2026.
The assessment of environmental and resource performance of energy transition technologies relies on quantitative information scattered across heterogeneous sources, including scientific articles, patents, and industrial reports, such as ESG (Environmental, Social, and Governance) disclosures. These documents contain key data for Life Cycle Inventory (LCI) and Material Flow Analysis (MFA), such as material and energy intensities, water consumption, mining production volumes, emissions, and technological descriptors. However, this information is predominantly embedded in unstructured PDF documents optimized for human reading, making large-scale, traceable data aggregation difficult and costly when performed manually.
This work presents an automated and modular methodology designed to extract and contextualize quantitative LCI and MFA data from three major categories of technical documentation. The approach combines large-scale document collection, relevance screening, and multimodal artificial intelligence within a reproducible and auditable workflow.
- Scientific Articles
Peer-reviewed articles are collected through automated scraping workflows based on structured search outputs. Documents are screened for LCI/MFA relevance using domain-specific keywords, methodological markers, and quantitative signal density. Relevant articles are then processed using a multimodal AI-based extraction core in which each page is analyzed through a combined text and image input. This enables robust extraction of numerical values from tables, text and figures while preserving contextual information such as units, methodological assumptions, and source location.
- Patents
Patent documents contain information about future trends on technologies and metal uses. Patents are collected via dedicated scraping pipelines and processed separately from scientific articles. The workflow focuses on extracting and structuring patent metadata, including publication year, country, and technology class, in order to characterize technological activity related to energy transition technologies. While quantitative LCI/MFA extraction from patents is not yet systematically performed, the pipeline enables descriptive statistical analyses of patent dynamics, including temporal trends and geographical patterns of technological development.
- Mining technical and ESG Reports
Official mining companies reports, with a specific focus on ESG ones, are processed through a screening module acting as a gatekeeper. The screening relies on sequential text parsing and, when necessary, geometric reconstruction of tables to identify reports containing sufficiently granular and structured quantitative information. Following human validation of the screening results, selected reports are analyzed using a IA-multimodal vision–language model combining page images and extracted text, enabling structured extraction of industrial metrics with associated context and traceability.
This automated methodology addresses one of the core challenges of data collection and significantly improves the granularity, consistency, and verifiability of LCI datasets and MFA inputs. The application of methodology is illustrated through examples related to battery and hydrogen technologies based on scientific articles and patents, and through case studies on copper and nickel production with a focus on mining based on industrial report. Although applied for LCA and MFA, the approach can also support the extraction of other types of data and indicators relevant to environmental and resource analyses. The tool provides automated and reliable support for researchers aiming to extract comprehensive foundational data from heterogeneous sources.
How to cite: Bejjit, C. E., Monfort, D., Muller, S., Lai, F., Beylot, A., and Hennioui, D.: Mining and raw materials sector: Automated Data Extraction and Contextualization for Life Cycle Inventory (LCI) and Material Flow Analysis (MFA) Across Scientific Articles, Patents and Mining companies Reports, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6501, https://doi.org/10.5194/egusphere-egu26-6501, 2026.
Multi-agent GIS systems are increasingly emerging as a general paradigm for complex geospatial tasks. However, many existing approaches rely on text-only large language models (LLMs) as the primary reasoning substrate. In the absence of explicit geometric constraints and verifiable evidence, spatial relations are often indirectly represented through linguistic statistical correlations. This makes LLMs prone to inconsistency when interpreting and inferring topological, directional, and distance relations in geospatial data, and leads to error accumulation across multi-step tool invocations and long-horizon decision-making, ultimately degrading the accuracy and efficiency of task reasoning and execution. In this work, we propose VisCritic-GIS, a multi-agent framework for geospatial task reasoning and execution driven by visualized evidence review. VisCritic-GIS introduces a Visualization Generation Agent and a Visualization Critic Agent into conventional multi-agent GIS pipelines. The generation agent renders key spatial data and intermediate results into 2D maps, explicitly externalizing spatial relations in a visual form. The critic agent leverages multimodal LLMs to read and critically review these map-based evidence, producing textual feedback on spatial relations, anomalous results, and reasoning deviations, which constrains and drives iterative refinement of other agents’ reasoning trajectories and toolchain configurations. We build evaluation protocols over representative remote sensing and geospatial tasks, and systematically demonstrate that VisCritic-GIS improves task accuracy, execution efficiency, and interpretability. Overall, our framework provides a mechanism for shifting geospatial reasoning from “text-only probabilistic completion” toward “visually grounded, verifiable inference,” thereby strengthening the robustness of spatial relation understanding in multi-agent GIS systems.
How to cite: Lan, Q., Hu, L., Wu, S., and Du, Z.: VisCritic-GIS: A Visualization-Critic–Empowered Framework for Multi-Agent Geospatial Task Reasoning and Execution, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10473, https://doi.org/10.5194/egusphere-egu26-10473, 2026.
Urban public spaces are highly dynamic systems where traffic patterns, pedestrian flows, and human activities vary strongly across temporal scales. Capturing these dynamics at high temporal resolution remains challenging, particularly using low-cost and reproducible observation methods. In this study, we present an automated workflow for continuous urban activity monitoring based on publicly available webcam imagery and deep learning–based object detection.
A public webcam overlooking Augustusplatz, a central urban square in Leipzig (Germany), is continuously accessed, and still frames are extracted from the video stream at one-minute intervals. Each frame is processed using the YOLO11 object detection model to identify and count relevant object classes, including passenger vehicles and pedestrians. The detection results are converted into structured JSON records and enriched with metadata such as timestamp and geographic location. All data are stored in an InfluxDB time-series database and visualized and statistically analyzed using Grafana.
This setup enables near-real-time and long-term analysis of urban activity patterns across multiple temporal scales. Distinct signatures of recurring and episodic events can be identified, including daily commuting cycles, evening rush hours, road closures, public celebrations, and large seasonal events such as Christmas markets. The minute-scale resolution allows for detailed investigation of short-term dynamics, while continuous operation over longer periods enables comparative and trend analyses.
The presented approach demonstrates how publicly available visual data and open-source tools can be combined into a scalable and transferable framework for urban monitoring. Potential applications include event detection, urban mobility analysis, validation of traffic models, assessment of public space usage, and integration with other environmental or socio-economic datasets. The method provides a cost-efficient complement to traditional urban sensing infrastructures and offers new opportunities for data-driven urban and environmental research.
How to cite: Oesen, B., Wagner, R., and Goblirsch, T.: High-Temporal-Resolution Urban Activity Monitoring Using Public Webcams and Deep Learning: A Case Study from Leipzig, Germany, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17514, 2026.
Generative AI is now woven into the daily study practices of geoscience students, often more deeply than educators acknowledge. This study examines how bachelor students in Earth Sciences (GEOL1008, NTNU) and master students in Engineering Geology (TGB4200, NTNU) use AI tools to understand literature, analyse data, synthesise research findings, and prepare written and oral assignments. The analysis draws on two structured surveys designed to map the extent and character of AI use in both cohorts.
Preliminary results indicate that AI has become the default support tool. Students turn to it to decode complex concepts, troubleshoot coding tasks, analyze data, structure reports, and polish presentations. Many see little distinction between traditional digital tools and generative AI, and the boundary between personal work and AI-augmented work is increasingly blurred. At the same time, students express uncertainty and worry about ethical expectations, disclosure practices, and the legitimacy of relying heavily on AI in academic work.
These trends have immediate consequences for assessment. Home exams, reports, and pre-prepared presentations no longer reliably reveal individual understanding, since nearly all students now use AI during preparation. Emerging evidence from portfolio-based courses suggests grade inflation and reduced differentiation between students, not because learning outcomes have improved, but because AI elevates the baseline quality of submitted work. In practice, written in-person exams and oral examinations remain among the few ways to assess unassisted reasoning.
The findings underscore a need to rethink teaching and assessment in geoscience education. AI is not a future challenge but a present reality, and universities must adapt if they aim to evaluate what students actually know rather than what their tools can produce.
How to cite: Fredin, O.: Geoscience Education in the Age of Generative AI: What Do Students Actually Learn?, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16861, 2026.
Climate Service Recipes (HACID-CSR) is an agentic system designed to assist providers of climate services in developing their advice for a wide range of clients. HACID-CSR guides providers by navigating the large and ever-increasing corpus of knowledge as well as an area without established standards and with limited access to scientific experts. It automatically generates detailed workflows (or “recipes”) by leveraging both a large language model’s internal reasoning and contextual knowledge from a domain knowledge graph for climate services (CS-DKG). The CS-DKG is an expert-curated ontology of climate service concepts with mapped relationships between climate variables, emission scenarios, indices, hazards, sectors, and key datasets (CORDEX, CMIP5, UKCP18), built as part of the Horizon Europe-funded HACID project (Hybrid Human Artificial Collective Intelligence in Open-Ended Decision Making).
The HACID-CSR architecture consists of a memory-enabled supervisor agent orchestrating multiple specialised agents. A planning agent first proposes an initial workflow outline, and a preliminary recipe agent uses only the LLM’s knowledge to draft answers to key workflow steps. The system then engages a knowledge graph retrieval sequence: a class selection agent identifies relevant classes in the CS-DKG, an instance selection agent finds specific instances (entries) highly relevant to the query within those classes following a two-stage selection process, i.e. semantic similarity based pre-selection and LLM-enabled refined selection, and a subgraph extraction agent retrieves the corresponding subgraph of related knowledge entities. Next, a recipe generation agent creates each step of the workflow by combining the LLM’s reasoning with the retrieved graph context using graph retrieval-augmented generation (GraphRAG). Finally, a recipe refinement agent compares the preliminary LLM-only solution with the knowledge-enhanced solution and refines the output, yielding a diverse and context-aware workflow.
By using this multi-agent approach, HACID-CSR increases the diversity of solutions and fills the knowledge gap between climate information and domain specific applications, helping experts to identify suitable methodologies and datasets. The resulting workflows are more traceable and transparent, improving user trust compared to answers from a general-purpose chatbot. We have also developed a bespoke automatic evaluation method to complement human expert validation of the generated recipes. We highlight the potential of the HACID-CSR approach for multi-hazard climate service design, and discuss remaining challenges and opportunities for further refinement of this agentic LLM-based system.
How to cite: Abele, A., Xie, H., Biswas, A., Dong, H., Fung, F., and Williams, H.: Climate Service Recipes: automatic multi-hazard climate information workflow generation using agentic Large Language Models (LLMs) and knowledge graphs, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2819, https://doi.org/10.5194/egusphere-egu26-2819, 2026.
Large Language Models (LLMs) have emerged as powerful tools for text and data processing, with potential extending far beyond conversational interfaces. We demonstrate that integrating LLMs into agentic workflows enables automated climate and oceanographic data analysis while minimizing hallucinations through strict reliance on real data sources.
ClimSight combines LLMs with climate model data to deliver localized climate insights for decision-making. Specialized agents consult external databases, extract variables from climate models, generate Python scripts for post-processing, and validate outputs through visual analysis. The workflow iteratively corrects errors until reliable results are achieved.
PANGAEA GPT enhances accessibility to the PANGAEA data repository through a supervisor agent that interprets queries, delegates tasks to domain-specific subagents, and coordinates data extraction, statistical analysis, and visualization of oceanographic and atmospheric datasets.
Both systems leverage automatic Python execution and image analysis for quality control. By constraining outputs to verifiable data sources and implementing multi-agent verification, we demonstrate that LLMs can play a significant role in geoscientific data pipelines and automated research workflows.
How to cite: Kuznetsov, I., Pantiukhin, D., Grassi, J., Shapkin, B., Jung, T., and Koldunov, N.: Integrating Large Language Models into Climate and Geoscientific Data Workflows, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13612, https://doi.org/10.5194/egusphere-egu26-13612, 2026.
The scale and heterogeneity of modern geospatial datasets, coupled with expanding suites of statistical and dynamical models, produce analysis outputs that are increasingly difficult to navigate and synthesise. We present a practical case study on using a large language model (LLM)-assisted coding tool (GitHub Copilot with GPT-5 mini within Visual Studio Code) to accelerate the development of a lightweight, HTML-based platform that visualises results from pre-calculated climate indicators.
Our starting point was a dataset comprising more than 130 climate indicators derived from gridded observations spanning over 60 years. These indicators originate from multiple meteorological variable groups (e.g., temperature, precipitation) and are aggregated at several temporal resolutions (e.g., annual, seasonal). Downstream analyses include spatiotemporal statistics, extreme value analyses and statistical significance testing, yielding hundreds of figures that are difficult to navigate and analyse. To make these outputs tractable, we prompted Copilot to generate a simple web application for visualisation and analysis purposes. The pre-generated plots from the climate indicator workflow were displayed there in an organised way, allowing for quick filtering through all indicators and different temporal resolutions, comparing different plots next to each other and using a subpage to concisely display aggregated group plots.
The platform is embedded and deployed via a GitLab CI pipeline, ensuring reproducible updates and immediate web accessibility for collaborators and users, thereby enabling rapid and easy access to vasts amount of output results. Our process of prompting a LLM to generate a visualisation platform offers a convenient and transferable workflow to aid geospatial data analysis.
How to cite: Lehner, S. and Schlögl, M.: Using copilot for the rapid generation of a visualisation platform to aid geospatial analyses, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20992, 2026.
This study explores the integration of EVE (Earth Virtual Expert), a Large Language Model specialized in Earth Observation (EO) and Earth Sciences, developed under ESA’s Φ-lab #AI4EO initiative in collaboration with Pi School. The primary objective is to enable EVE to connect ESA’s EO platforms and data clusters, creating an integrated ecosystem for the community. This approach leverages agentic capabilities, allowing EVE to dynamically interact with EO tools, databases, and APIs to reason and act autonomously.
To demonstrate this concept, we present a use case where EVE operates within an agentic framework to interact with the EO Dashboard, a joint initiative by ESA, NASA, and JAXA that provides global indicators and narratives derived from multi-mission EO data. Using the MCP protocol, this work enables dynamic connectivity between EVE and the Dashboard, allowing the model to interpret and summarize narratives, extend insights with additional context, and facilitate advanced information retrieval across datasets and stories. In addition, the study considers potential directions for agentic behaviors, assessing early-stage possibilities and limitations for features such as autonomous task chaining. These capabilities enable EVE to perform multi-step reasoning, for example, by interpreting quantitative trends in dashboard indicators such as air quality changes, greenhouse gas concentrations, or land cover dynamics. This links EVE to underlying datasets and enables the generation of scientifically grounded responses. This proof-of-concept demonstrates EVE’s potential to foster interoperability and accelerate Earth system science by improving knowledge accessibility and enabling more effective use of EO data resources.
How to cite: Gmelich Meijling, E., D'Ercole, R., Anghelea, A., Cocchiara, C. M., and Longepe, N.: Building Connected Earth Observation Ecosystems with Agentic AI using EVE, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20206, 2026.
Reservoir modelling in heterogeneous carbonate systems is often constrained by sparse well control and labor-intensive interpretation, which increases uncertainty when extrapolating between wells. We present an enhanced Pix2Geomodel.v2 workflow that reframes facies and petrophysical modelling as paired image-to-image translation. Facies and petrophysical properties are exported from a reference reservoir model, converted into paired 2D training images, and used to train a Pix2Pix-style conditional generative adversarial network (cGAN). The architecture couples a U-Net generator with a PatchGAN discriminator, enabling the model to learn spatial relationships directly from examples. To reduce data requirements while retaining geological heterogeneity, the workflow operates on a streamlined grid of 54 vertical layers and targets complex facies distributions. Preliminary results show stable training and predictions that reproduce the main geological patterns of the reference data. In facies-to-property translation, the network learns meaningful mappings to porosity, permeability, and volume of shale.
How to cite: Al-Fakih, A., Hanafy, S., Saraih, N., Koeshidayatullah, A., and Kaka, S.: Data-efficient enhanced Pix2Geomodel.v2 for complex facies settings, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2220, https://doi.org/10.5194/egusphere-egu26-2220, 2026.
Capturing cross-property correlations while preserving spatial continuity is essential for reliable reservoir characterization, especially in heterogeneous reservoirs where facies architecture controls petrophysical variability. In this study, we evaluate Pix2Geomodel.v2 as a bidirectional image-to-image translation framework that learns mappings between facies and petrophysical properties using paired 2D slices exported from a reference reservoir model. To reduce data demands while maintaining geological complexity, the workflow operates on a streamlined grid of 54 vertical layers, enabling efficient training and rapid experimentation without removing key stratigraphic and facies patterns. The approach is based on a conditional generative adversarial learning strategy. A U-Net generator is trained to synthesize target facies or property maps from input images, while a PatchGAN discriminator encourages locally realistic textures and geologically plausible transitions. The paired-slice formulation allows the model to learn both large-scale structural organization and fine-scale heterogeneity directly from examples. We investigate two complementary directions: (i) facies-to-property translation, where facies maps are used to predict continuous property fields such as porosity and permeability, and (ii) property-to-facies translation, where petrophysical images are used to reconstruct discrete facies distributions. Beyond conventional forward mapping, the reverse translation experiments are particularly informative because they test whether the model captures meaningful cross-property dependencies rather than superficial patterns. The reconstructed facies maps recover coherent large-scale facies trends and geologically consistent connectivity, indicating that the learned representation encodes relationships between depositional architecture and petrophysical response. Spatial realism is further examined using experimental variograms, providing a continuity-based check that generated outputs qualitatively align with the reference model in terms of spatial correlation structure. Overall, the results suggest a data-efficient route to robust forward and reverse translations that can support faster reservoir model prototyping, property population guided by facies, and consistency checking between facies and petrophysical interpretations.
How to cite: Kaka, S., Al-Fakih, A., Saraih, N., Koeshidayatullah, A., and Hanafy, S.: Bidirectional translation + spatial continuity validation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2222, https://doi.org/10.5194/egusphere-egu26-2222, 2026.
Seismic interpretation of mass transport deposits (MTDs) relies heavily on expert knowledge and conceptual reasoning yet remains difficult to formalize and scale. While recent artificial intelligence (AI) methods have shown strong capabilities in seismic pattern recognition, most approaches operate as black boxes and remain poorly aligned with the interpretative frameworks used by geoscientists, limiting transparency and trust.
This study proposes a geoscience-aware hybrid intelligence framework that integrates expert knowledge graphs (KGs) with large language models (LLMs) to support interpretable seismic interpretation of MTDs. The approach builds upon the conceptual methodology of Le Bouteiller et al. (2019), which organizes MTD interpretation through causal relationships linking environmental controls, mass transport properties, and observable seismic descriptors across trigger, transport, and post-deposition phases.
The KG provides a structured reference for interpretation that constrains vocabulary, causal direction, and temporal logic. Our workflow reads scientific papers, identifies relevant descriptors and processes, checks them with LLMs, and evaluates how well they support interpretation. In this setup, seismic descriptors give different levels of support (weak to strong) for geological processes, like how experts reason under uncertainty.
Preliminary results show that ~68% of expert defined concepts are recovered in the inferred graph, with a semantic validation score of 0.73, indicating good conceptual alignment. However, descriptor matching based on textual similarity remains difficult, with average scores around 0.41. This gap highlights the difference between semantic agreement (conceptually correct) and textual agreement (exact wording), mainly due to synonymy and variable phrasing in the literature. We plan to address this by using domain-specific LLMs and ontology-based synonym expansion to improve semantic matching in future iterations
How to cite: Talbi, F. B., Armitage, J., Charléty, J., Rabaute, A., Bouziat, A., Vittaut, J.-N., and Leroy, S.: Geoscience-Aware AI for Interpretable Seismic Interpretation of Mass Transport Deposits Using Knowledge Graphs and Large Language Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6459, https://doi.org/10.5194/egusphere-egu26-6459, 2026.
Electromagnetic methods are among the most widely used techniques in the geophysical exploration industry due to their efficiency and non-invasive nature. However, their data processing workflows are highly time-consuming and strongly dependent on expert intervention. With the rapid and broad success of deep learning, applying deep learning techniques to electromagnetic methods to overcome the limitations of traditional approaches has become an active area of research. The effectiveness of deep learning methods, however, largely depends on the quality of the dataset, which directly influences model performance and generalization capability. Existing applications typically rely on self-constructed datasets composed of randomly generated one-dimensional models or structurally simple three-dimensional models, which fail to capture the complexity of realistic geological environments. Moreover, the absence of a unified and publicly available three-dimensional geoelectrical model repository has further constrained the development of deep learning for three-dimensional electromagnetic exploration. To address these challenges, we introduce OpenEM, a large-scale, multi-structural three-dimensional geoelectrical model repository that incorporates a wide range of geologically plausible subsurface structures.
OpenEM comprises nine categories of geoelectrical models, encompassing a wide spectrum of subsurface structures ranging from simple to complex. These include models of homogeneous half-spaces with embedded anomalous bodies, as well as configurations featuring flat stratigraphy, curved stratigraphy, planar faults, curved faults, and their variants containing anomalous bodies. The resistivity values span from 1 to 2000 Ω·m, with the number of layers ranging from three to seven. In models containing anomalous bodies, the number of anomalies varies from one to five, and both regular and irregular geometries are considered to enhance dataset diversity and realistic representativeness. In addition, OpenEM is accompanied by a three-dimensional model generator that enables fully controllable model construction, allowing users to customize structural configurations, including resistivity magnitudes, fault geometries and locations, as well as the size, shape, and placement of anomalous bodies.
OpenEM provides a unified, comprehensive, and large-scale dataset for common electromagnetic exploration systems, thereby promoting the application of deep learning methods in electromagnetic prospecting.
How to cite: Wang, S., Wang, X., Deng, F., and Jiang, P.: OpenEM: Large-scale multi-structure 3D dataset for electromagnetic methods, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8976, https://doi.org/10.5194/egusphere-egu26-8976, 2026.
Artificial intelligence (AI) tools increasingly enhance the efficiency and consistency of seismic interpretation, particularly in structurally complex areas or areas where data quality is reduced by acquisition limitations. As a result, interpretations can become difficult and time-consuming, especially in the context of structural interpretation and fault tracking. To evaluate the performance of AI-based fault detection, we applied Geoplat AI software to a 3D seismic volume from the Drmno Basin, located at the southeastern margin of the Pannonian SuperBasin in Serbia.
A conventional structural interpretation was first performed by mapping the major fault systems, then minor fault systems, generating fault sticks and polygons for all visible faults and developing a structural model to illustrate the basin's opening and evolution. Subsequently, AI-based workflows were applied in order to enhance the quality of the seismic data. This involved removing noise, restoring reflections, highlighting fault zones, and applying smoothing filters. The final step was the utilization of a fault tracking tool that segments the seismic data, recognizes fault zones, traces them, identifies structural patterns, and calculates a probability field. The AI-derived fault interpretation was then compared with the manual interpretation.
The results indicate that the Drmno basin was developed under an extensional tectonic regime during the Early Miocene, which formed a large Morava detachment fault and opened accommodation of the basin. The basin itself has complex architecture in the syn-rift phase, with many synthetic and few antithetic faults, oriented from the east to the west. During the stage of the rift climax, the dominant fault systems remained consistent, with most syn-rift structures continuing to accommodate the subsidence formed by the Morava detachment. The shift in the tectonic conditions in the post-rift stage leads to the formation of systems of parallel faults in the younger sediments, adjusting strike-slip movements in a compressional tectonic field. The younger structures are dominantly oriented in the north-south direction, or reactivated older fault structures.
The AI tool effectively interpreted fault systems in the younger geological units, benefiting from higher data quality, and clearly indicated younger fault systems with a high level of certainty. However, in the lower part of the seismic cube, the basement structures remain unclear or unrecognized. Reactivated fault surfaces and a significant fault zone are evident in the interpretation. In areas with low-quality seismic data, the AI tool struggled to trace faults accurately, resulting in geologically inconsistent fault patterns.
Overall, the AI-based 3D fault tracking tool proved effective in resolving the main structural framework of the basin. The dominant fault directions are clearly identifiable, and the main geological structures have been mapped with reasonable precision. The AI-supported interpretation successfully captures the main structural trends and provides a solid basis for evaluating the tectonic evolution. This case study demonstrates the potential of AI to support structural interpretation and tectonic analysis of complex sedimentary basins.
How to cite: Ninić, A., Radivojević, D., and Đurić, D.: The application of artificial intelligence in fault tracking on 3D seismic data – A case study from Drmno Basin (SE Serbia), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2380, https://doi.org/10.5194/egusphere-egu26-2380, 2026.
This study investigates the domain adaptation of the Vision-Language Model (VLM) for road damage assessment, focusing on a fine-tuning strategy optimized for resource-constrained engineering environments. Unlike conventional object detection models that operate within fixed label spaces, VLMs provide superior semantic understanding and generalization in complex scenarios. To facilitate practical deployment, this research systematically analyzes key variables of Parameter-Efficient Fine-Tuning (PEFT) to mitigate the high computational demands inherent in large-scale VLMs.
In the experimental phase, hyperparameter tuning was conducted using the Low-Rank Adaptation (LoRA) technique. The primary variables included LoRA ranks (16, 32, 64, and 96), training data scale, and image resolutions (1,024ⅹ28ⅹ28 vs. 1,536ⅹ28ⅹ28). A comprehensive dataset of 26,796 images comprising six damage categories and negative samples was established, utilizing a 7n sampling strategy (n=500, 750, 1,000) to address class imbalance. The impact of data volume was evaluated by augmenting the 7,000-sample set (corresponding to n=1,000) to match the full dataset size of 26,796, with zero-shot inference serving as the performance baseline.
Experimental results demonstrated substantial improvements over zero-shot inference, indicating that performance positively correlates with increased data scale with augmentation and higher image resolution, while lower LoRA ranks (16, 32) proved most effective for this domain. Furthermore, the introduction of specialized ad-hoc metrics, MmAP and MF1, verified a stable trade-off between False Positives and False Negatives. Notably, to minimize safety-critical False Negatives, a prompt engineering-based 'Double Check' mechanism and multi-turn interactions were utilized. This approach successfully leveraged the model’s inherent reasoning capabilities to refine damage identification through iterative feedback.
Acknowledgements This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(RS-2025-25437298)
How to cite: Kim, D. and Youn, H.: Optimizing Vision-Language Model for Robust Road Damage Assessment via Parameter-Efficient Fine-Tuning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3674, https://doi.org/10.5194/egusphere-egu26-3674, 2026.
Large Language Models (LLMs), a class of contemporary artificial intelligence systems, are increasingly used in scientific practice to support research workflows, accelerate discovery, and automate routine administrative tasks. This contribution identifies and analyzes three underexplored aspects of LLM adoption in scientific research. The first aspect concerns the uneven adoption of LLMs among scientists and the inconsistent application of established best practices. The second examines how LLMs can be employed to improve the robustness and reproducibility of scientific practices. The third addresses institutional strategies by which large scientific organizations—such as universities and research networks—can reduce dependence on commercial technology providers while increasing trust in LLM-based systems.
The findings found in our contribution are the partly summarization of István Bozsó’s experiences serving in the role of an “AI ambassador” at the Institute of Earth Phyisics and Space Science (EPSS) of the Hungarian Research Network (HUN-REN).
In our experience many scientists are still skeptical of using LLMs in any capacity or lack the time to invest in learning these technologies. These barriers are primarily sociotechnical rather than purely technical in nature, and they require, on one hand materials that teach best-practices and show motivating examples for using LLMs, on the other hand services provided by research organisations.
Recent advances in open-weight LLMs enable self-hosting within institutional computing infrastructures, which means research institutes can run these models on their own hardware and thereby ensuring that sensitive data and research materials remain within the organization’s controlled digital environment. This also ensures that the LLM usage stays independent of Large Technology corporations and builds trust with colleagues.
Regarding motivating examples, we wish to focus on two areas which can be addressed with the help of LLMs. One area is scientific communication. LLMs can easily generate materials (primarily text, sound and video) which can be used to inform the wider public on new scientific discoveries and push back against misinformation and disinformation campaigns. The involment of scientists is paramount in the review and finalization of such materials to ensure they represent accurate scientific information.
The other area is scientific programming. Many scientists are not trained as professional software engineers and often lack the time and background to apply software development best practices. In many cases this results in software artifacts that are fragile, difficult to reproduce, and challenging to maintain and usually only work on the machine of the researcher who developed the package. LLMs can help out in these situations by suggesting and even implementing best practices and giving programming advice to the researcher during the development of the scientific code.
The common theme in these examples is that the LLM is not meant to replace the scientist but enhance their capabilities with the goal of increasing the robustness, transparency, and sustainability of the scientific research process.
How to cite: Bozsó, I., Horváth, A., and Kuslits, L.: Bridging the gap between scientists and large language models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17176, 2026.
Compound drought and heat extremes (CDHEs) exert impacts that exceed the sum of their individual components. With global warming amplifying the associated risks, CDHEs have become a critical threat to agricultural production. Thus, identifying and monitoring CDHEs in cropland systems is key for food security. As CDHEs formation and evolution are shaped by climatic factors, hydrological cycles, and ecosystem feedbacks, their fine- and large-scale identification in agricultural areas poses substantial challenges.
Our study reviews existing methods for identifying CDHEs, including combined threshold approaches, comprehensive index methods, traditional machine learning techniques, and improved mechanistic modeling. We summarize the current limitations of these methods as follows: (1) Combined threshold and comprehensive index methods often focus on a single aspect of CDHEs, failing to systematically describe the complex processes of compound events. (2) While traditional machine learning methods attempt to integrate characteristics of the hazard-bearing body, disaster-causing factors, and hazard-inducing environment to establish complex nonlinear relationships between multiple elements and compound event indices, their "black-box" nature lacks mechanistic interpretability. Furthermore, these methods rely heavily on large volumes of high-quality samples to achieve satisfactory accuracy. (3) Improved mechanistic models, typically based on classical agricultural process models such as APSIM and AquaCrop, introduce CDHE impact modules to address the oversimplification of these effects in original models. Nevertheless, these mechanistic models require extensive input parameters, and their calibration processes depend on substantial amounts of measured data. Additionally, the computational resources needed for simulations are considerable, making the cost of analyzing CDHEs over large farmland areas under various future climate scenarios prohibitive for individual researchers.
To address these challenges, this study highlights the potential of physics-informed neural network models for identifying compound events and proposes future research directions regarding mechanistic constraints, neural network architecture design, and experimental plans: (1) Farmland CDHEs are essentially phenomena of water and heat imbalance within the soil-crop-atmosphere (SCA) system. Utilizing the Richards equation and the Penman-Monteith formula can characterize this process by constraining the water and heat environmental factors at the two key interfaces: root-soil and leaf-atmosphere. (2) Solar-induced chlorophyll fluorescence (SIF), a byproduct of vegetation photosynthesis closely related to GPP, responds rapidly to physiological damage caused by stress. Utilizing multi-band SIF data can provide a detailed depiction of crop physiological responses to stress from the perspective of the hazard-affected body. (3) Automated design of model architectures incorporating mechanistic information for farmland compound events can be achieved through distillation learning. (4) Future work should integrate ground-based water and heat control experiments with site-specific hyperspectral SIF observation data. Through continuous combinatorial experimental design, this approach can lead to the development of accurate and efficient physics-informed neural networks. Coupled with large-scale satellite and reanalysis data products, this framework aims to enable the large-area identification of farmland CDHEs under future climate scenarios.
How to cite: Xia, H., Wu, J., Zhou, L., and Du, R.: Towards a Mechanism-Informed Intelligent Framework for Identification of Compound Drought-Heat Extremes in Croplands, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8579, https://doi.org/10.5194/egusphere-egu26-8579, 2026.
Nitrate contamination of groundwater is often addressed as a diffuse agricultural management problem, yet monitoring in Denmark depicts that exceedance risk at abstraction and observation wells is spatially structured and closely linked to surrounding land-use composition and configuration. This suggests a land-use policy opportunity: if landscape fractions and fragmentation patterns help drive nitrate vulnerability, interventions could be spatially targeted and tailored rather than uniformly applied. In this study, we present a scenario-based planning framework for policy appraisal, enabling regulators, municipalities, and water utilities to test alternative policy packages and targeting rules and quantify their expected effects on groundwater nitrate hotspot risk. The system operates a predictive pipeline that relates nitrate outcomes to land-use fractions and landscape configuration metrics computed within configurable protection zones. Model outputs are formulated as a binary hotspot classification (hotspot vs. non-hotspot) based on exceedance of a drinking-water nitrate threshold, producing vulnerability maps to prioritize locations for intervention and prevention. The core functionality is a “what-if” engine built on an AI-based ensemble that generates a baseline nitrate-risk probability map and re-predicts risk under user-defined scenarios. Scenario levers are organized into two policy bundles: (i) land-use policy and management, implemented as controlled reallocations among land-cover fractions (e.g., reducing large contiguous cropland blocks, increasing wetland/riparian woodland cover, restricting impervious expansion) while enforcing feasibility constraints; and (ii) agricultural management, implemented as proportional reductions or caps on nitrogen surplus and fertilizer inputs. For each scenario, the system outputs an updated probability map and a different map relative to baseline, supporting spatial prioritization, instrument design, and transparent justification of differential targeting. By combining ex-ante scenario testing with ex-post monitoring of hotspot transitions after implementation, the framework supports adaptive groundwater governance and moves from risk mapping toward operational, spatially explicit nitrate-reduction policy design.
How to cite: Naghibi, A., Ahmadi, K., and Berndtsson, R.: Policy-oriented Land-Use and Agricultural Management Scenarios for Groundwater Nitrate Hotspot Mitigation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4407, https://doi.org/10.5194/egusphere-egu26-4407, 2026.
The increasing availability of satellite remote sensing data has made automatic land cover change detection a persistent research focus. However, real-world applications show that single AI models struggle to cope with the combined challenges of spatial-temporal complexity, feature diversity, and evolving engineering requirements. Consequently, the accuracy of automatically extracted land cover changes is often compromised, making the results insufficient for direct engineering application. Guided by practical application need, this paper focuses on how to utilize satellite remote sensing data, various knowledge and AI technologies to improve the accuracy and efficiency of automatic land cover extraction. This research focuses on the key technologies involved in the complete land cover monitoring process. Central to this study is the proposal of a progressive intelligent change detection technology for satellite remote sensing, characterized by a “identify all, discriminate precisely, refine extraction” workflow. Specifically, the “identify all” step extracts all potential change patches using models such as generic binary change detection. Building on these results, the “discriminate precisely” step filters out patches that are not of current interest. Finally, the “refine extraction” step employs models like semantic segmentation to further screen the results and enhance overall accuracy. An application demonstration in Shanxi Province, China, for new PV facilities, buildings, and roads demonstrated a recall rate of 89.3% for automatic extraction. The high-quality outputs confirm the practical applicability of the results. Consequently, this research affirms the technology as both a valuable and transferable solution for land cover monitoring.
How to cite: You, S., Du, L., He, Y., and Ye, F.: Research on Key Technologies for High-Precision Land Cover Change Monitoring Using Satellite Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19954, 2026.
The Yellow River Basin (YRB) serves as a critical repository of literature for understanding human-earth systems, yet existing automated metadata-level review methods suffer from deep semantic loss and deficiencies in spatial representation: They neither capture fine-grained logic chains from full texts nor possess the capability to extract the spatial and hierarchical attributes of geographic entities. However, rapid developments in Large Language Models (LLMs) provide a technological opportunity for the automated extraction of full-text knowledge. To this end, this study proposes the Geo-Knowledge Infused Reasoning Framework (GK-IRF), coupling full-text semantics with multi-level spatial indexing. Methodologically, we first construct an ontology-based full-text parsing mechanism based on 8,493 YRB-related papers (2015-2024), utilizing LLMs to accurately extract structured semantic triplets. Simultaneously, we introduce an adaptive multi-level GeoHash indexing model to map textual toponyms into hierarchically nested grid sets, reconstructing the spatial coverage and multi-scale associations of geographic entities. Validations against a manually annotated dataset indicate that GK-IRF achieves an F1-score comparable to human performance in full-granularity semantic extraction; furthermore, the Spatial Coverage Accuracy of the multi-level grids for the YRB substantially outperforms traditional geocoding methods, effectively resolving the challenge of multi-scale coverage representation.
How to cite: Wu, S. and Wang, H.: Coupling Full-Text Semantics with Multi-Level Spatial Indexing: A Knowledge Representation Framework for Yellow River Basin Literature, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6201, https://doi.org/10.5194/egusphere-egu26-6201, 2026.
Posters virtual: Mon, 4 May, 14:00–18:00 | vPoster spot A
EGU26-13942 | ECS | Posters virtual | VPS31
From Generative Sampling to Urban Typology: A PRTS-AI Supported Framework for Multi-Decadal Urban LULC Mapping and Cross-City Transferability AnalysisMon, 04 May, 16:15–18:00 (CEST) vPoster Discussions
High-quality, temporally consistent training samples are the cornerstone of accurate long-term urban Land Use/Land Cover (LULC) mapping. However, traditional sample generation relies heavily on labor-intensive manual interpretation and often lacks reproducibility. To address this, we developed PRTS-AI (Primary Regulated Time-series Sampling), an open-source system that integrates OpenStreetMap (OSM) data extraction, Large Language Model (LLM)-driven semantic classification, and LandTrendr-based temporal filtering into an automated workflow. By leveraging generative AI (e.g., DeepSeek/ChatGPT/Gemini) to interpret polygon attributes and using POI-based consistency checks, the system significantly reduces manual workload while ensuring semantic accuracy.
The PRTS-AI system integrates multi-source spatial and temporal data into a streamlined workflow, including:
(1) extraction of OpenStreetMap (OSM) features for user-defined study areas;
(2) semantic classification of polygon features using large language models;
(3) detection and filtering of change pixels using the LandTrendr time-series algorithm;
(4) recommendation of city-specific sampling parameters based on a six-dimensional urban typology framework.
This system enables reproducible multi-temporal sample generation, spatial heterogeneity validation, and fine-scale classification support across diverse urban settings. Furthermore, this system can operate in parallel with the usual land cover sample selection and subsequent classification processes.
We applied PRTS-AI to map the urban evolution of diverse cities in Liaoning and Shandong provinces, China, from 2000 to 2020. The framework achieved an overall mapping accuracy of ~80%, with residential categories reaching 90%. Beyond mapping, we utilized the fine-grained Local Climate Zone (LCZ) metrics generated by the system to investigate the transferability of samples. Through Principal Component Analysis (PCA) of residential morphologies, we quantitatively identified that cities cluster into distinct typologies driven by macro-factors (e.g., coastal vs. resource-based industrial cities) rather than administrative hierarchies. These findings challenge the assumption of universal sample transferability, suggesting that sample migration is most effective within specific urban typologies. Consequently, PRTS-AI incorporates a typology-based parameter recommendation module to guide city-specific sampling. This study presents a scalable, AI-empowered solution for urban mapping and offers new insights into the spatiotemporal heterogeneity of urban forms.
However, limited sample transferability may still be achieved between cities with similar characteristics, based on a preliminary six-dimensional classification framework.
PRTS-AI provides a lightweight, reproducible, and extensible solution for urban LULC research, supporting both academic investigations and practical urban planning applications.
How to cite: Tian, T., Yu, L., Chen, B., and Gong, P.: From Generative Sampling to Urban Typology: A PRTS-AI Supported Framework for Multi-Decadal Urban LULC Mapping and Cross-City Transferability Analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13942, https://doi.org/10.5194/egusphere-egu26-13942, 2026.
EGU26-7766 | Posters virtual | VPS31
Automated Taxonomic Identification of Calcareous Nannofossils from Microscopic Imagery Using Convolutional Neural NetworksMon, 04 May, 16:15–18:00 (CEST) vPoster Discussions
Calcareous nannofossils represent a key proxy for biostratigraphy and paleoenvironmental reconstructions, due to their high abundance, widespread distribution and rapid evolutionary turnover. However, conventional taxonomic identification under optical or electron microscopy remains time-consuming and strongly dependent on expert interpretation, especially when working with large datasets and heterogeneous assemblages. This limitation is critical for high-resolution stratigraphic studies in complex sedimentary settings where reworking, redeposition and tectonic transport may generate mixed-age associations.
This poster focuses on qualitative and quantitative investigations of Quaternary calcareous nannofossils based on microscopic analyses and the development of an automated taxonomic identification workflow. We propose a deep learning approach using a convolutional neural network (CNN) trained on curated image catalogues of nannofossil taxa, aiming to achieve end-to-end classification of microfossil imagery. The targeted temporal interval spans approximately on the last 25,000 years (since the LGM – Last Glacial Maximum), focused on samples from the NW Black Sea cores.
Beyond accelerating routine identifications, automated classification has the potential to provide more objective and reproducible taxonomic assignments, enabling consistent quantitative counting and supporting multidisciplinary analyses linking nannofossil variability to paleoenvironmental controls such as salinity, nutrient input and temperature. The proposed workflow represents a step toward scalable microfossil taxonomy, supporting robust stratigraphic correlations and palaeoceanographic interpretations in Quaternary successions.
Keywords: nannofossils, neural networks, image recognition
How to cite: Cudalbu, C., Cudalbu, B., and Melinte - Dobrinescu, M.: Automated Taxonomic Identification of Calcareous Nannofossils from Microscopic Imagery Using Convolutional Neural Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7766, https://doi.org/10.5194/egusphere-egu26-7766, 2026.
