Sept. 3, 2025,
13:30 –
15:15
Room info:
Lecture Hall
Timetable
Sept. 3, 2025
Climate research often requires substantial technical expertise. This involves managing data standards, various file formats, software engineering, and high-performance computing. Translating scientific questions into code that can answer them demands significant effort. The question is, why? Data analysis platforms like Freva (Kadow et al. 2021, e.g., gems.dkrz.de ) aim to enhance user convenience, yet programming expertise is still required. In this context, we introduce a large language model setup and chat bot interface for different core e.g. based on GPT-4/ChatGPT or DeepSeek, which enables climate analysis without technical obstacles, including language barriers. Not yet, we are dealing with climate LLMs for this purpose. Dedicated natural language processing methodologies could bring this to a next level. This approach is tailored to the needs of the broader climate community, which deals with small and fast analysis to massive data sets from kilometer-scale modeling and requires a processing environment utilizing modern technologies, but addressing still society after all, such as those in the Earth Virtualization Engines (EVE - eve4climate.org ). Our interface runs on an High Performance Computer with access to PetaBytes of data - everything just a chat away.
Meeting room
Lecture Hall
The protection of critical underwater infrastructure, such as pipelines, data cables, or offshore energy assets, has become an emerging security challenge. Despite its growing importance, maritime infrastructure monitoring remains limited by high costs, insufficient coverage, and fragmented data processing workflows. The ARGUS project addresses these challenges by developing an AI-driven platform to support risk assessment and surveillance at sea.
At its core, ARGUS integrates satellite-based Synthetic Aperture Radar (SAR) imagery, AIS vessel tracking data, and spatial information on critical assets into a unified data management system. A key functionality is detecting so-called "ghost ships" – vessels that deliberately switch off their AIS transponders – using object detection techniques on SAR imagery.
At the same time, we are currently developing methods for underwater anomaly and change detection based on optical imagery. This work is still ongoing and focuses on identifying relevant structural or environmental changes in submerged infrastructure through automated image comparison and temporal analysis.
In this talk, we present the architecture and workflows of the ARGUS system, including our use of deep learning (YOLO-based object detection) in the maritime context. We share insights into the current capabilities and limitations of AI models for maritime surveillance, especially in the context of …
Meeting room
Lecture Hall
Current AI cannot function without data, yet this precious resource is often underappreciated. In the context
of machine learning, dealing with incomplete datasets are a widespread challenge. Large, consistent, and
error-free data sets are essential for an optimally trained neural network. Complete and well-structured in-
puts substantially contribute to both training, results and subsequent conclusions. As a result, using high-
quality data improves the performance and the ability of neural networks to generalize.
However, real-world datasets from field measurements can contain information leakage. Sensor failures,
maintenance issues or inconsistent data collection can cause invalid ('NaN', Not a Number) values to appear
in the neural network input matrices.
Imputation techniques are an important step in data processing for handling missing values. Estimating
'NaN' values or replacing them with plausible values directly affects the quality of the input data and thus
the effectiveness of the neural network.
In this contribution, we present a neural network-based regression model (ANN regression), that explains
the salt characteristics in the Elbe estuary. In this context, we focus on selecting appropriate imputation
strategies.
While traditional methods such as imputation by mean, median, or mode are simple and computationally
efficient, they sometimes fail to preserve the underlying data …
Meeting room
Lecture Hall
We present a comprehensive machine learning framework for predicting spatially distributed geographical data from point measurements. The framework takes as input a set of geographical features at a specified grid resolution (e.g., 5 arc-minute scale) and corresponding point measurements with their spatial coordinates and target values. The framework trains and evaluates multiple machine learning models, including both tree-based methods (Random Forest, XGBoost, CatBoost) and deep learning architectures (feed forward neural networks, TabPFN[1]), to identify the optimal predictive model for the given dataset.
The framework incorporates hyperparameter search(depth and width) for deep learning models and systematic parameter search for tree-based models (e.g., number of estimators). This ensures robust model selection and performance optimization across different geographical contexts and data characteristics. The framework outputs the best-performing model along with comprehensive performance metrics and uncertainty estimates.
As a non-trivial application, we demonstrate the framework's effectiveness in predicting total organic carbon (TOC) concentrations[2] and sedimentation rates in the ocean. This involves integrating features from both the sea surface and seafloor, encompassing a diverse array of oceanographic, geological, geographic, biological, and biogeochemical parameters. The framework successfully identifies the most suitable model architecture and hyperparameters for this complex spatial prediction task, providing both high accuracy and …
Meeting room
Lecture Hall
Urban-scale air quality data is crucial for exposure assessment and decision-making in cities. However, high-resolution Eulerian Chemistry Transport Models (CTMs) with street-scale resolutions (100 m x 100 m), while process-based and scenario-capable, are computationally expensive and require city-specific emission inventories, meteorological fields and boundary concentrations. In contrast, machine learning (ML) offers a scalable and efficient alternative to enhance spatial resolution using existing regional-scale (1 km - 10 km grid resolutions) reanalysis datasets.
We present a reproducible ML framework that downscales hourly NO 2 data from the CAMS Europe ensemble (~10 km resolution) to 100 × 100 m 2 resolution, using 11 years of data (2013–2023) for Hamburg. The framework integrates satellite-based and modelled inputs (CAMS, ERA5-Land), spatial predictors (CORINE, GHSL, OSM), and time indicators. Two ML approaches are employed: XGBoost for robust prediction and interpretability (via SHAP values), and Gaussian Processes for quantifying spatial and temporal uncertainty.
The downscaling is evaluated through random, time-based and leave-site-out validation approaches. Results demonstrate good reproduction of observed spatial and temporal NO 2 patterns, including traffic peaks and diurnal/seasonal trends. The trained models generate over 160 million hourly predictions for Hamburg with associated uncertainty fields. Although developed for Hamburg, the framework has been successfully …
Meeting room
Lecture Hall
This study presents an end-to-end deep learning framework, 4DVarNet, for reconstructing high-resolution spatiotemporal fields of suspended particulate matter (SPM) in the German Bight under realistic satellite data gaps. Using a two-phase approach, the network is first pretrained on gap-free numerical model outputs masked with synthetic cloud patterns, then fine-tuned against sparse CMEMS observations with an additional independent validation mask. The framework architecture embeds a trainable dynamical prior and a convolutional LSTM solver to iteratively minimize a cost function that balances data agreement with physical consistency. The framework is applied for one year data (2020) of real observations (CMEMS) and co-located model simulations, demonstrating robust performance under operational conditions. Reconstructions capture major spatial patterns with correlation R2 = 0.977 and 50% of errors within ± 0.2 mg/L, even when 27% of days lack any observations. Sensitivity experiments reveal that removing 60% of available data doubles RMSE and smooths fine-scale SPM spatial features. Moreover, increasing the assimilation window reduces edge discontinuities between the data-void area and the adjacent data-rich region, whereas degrades sub-daily variability. Extending 4DVarNet to higher temporal resolution (hourly) reconstruction will require incorporating tidal dynamics to account for SPM resuspension, enabling real-time sediment transport forecasting in coastal environments.