Christian Heipke

Reliable and safe navigation of autonomous systems is an active field of research, particularly in the context of complex urban environments. A key challenge is navigation integrity, defined as the trust in the correctness of the estimated system state; the transfer of this concept from aviation to collaborative multi-agent systems remains an open research challenge. Addressing this gap, this paper presents the methodological developments and key findings of the DFG-funded i.c.sens research project, which focused on navigation integrity in dynamic collaborative sensor networks. To account for key influencing factors and to structure the research activities accordingly, three central research directions are defined: Observations (sensor characteristics and uncertainties), Methodology (state estimation and integrity monitoring), and Map (dynamic, uncertainty-aware environmental representations). To evaluate the developed approaches and assess their performance under realistic conditions, extensive experiments were conducted. Based on experimental results obtained from publicly available datasets as well as datasets collected by the i.c.sens group, the paper summarizes key achievements of i.c.sens in integrity-aware state estimation under realistic urban conditions. These results show improved navigation performance through multi-agent information exchange, through rigorous uncertainty modeling and through the conception and usage of sophisticated map concepts. Based on these findings, we highlight remaining integrity challenges in dynamic sensor networks and point out promising directions for future research.

This paper presents GeoGraphJSON, a lightweight semantic data model that integrates spatial geometry with graph-based connectivity to support advanced spatial analysis and GeoAI applications. Conventional geospatial formats primarily focus on geometry and attributes, requiring relationships such as adjacency, containment, and connectivity to be derived dynamically through computationally intensive spatial operations. In contrast, GeoGraphJSON encodes these relationships explicitly as typed edges within a unified, interoperable structure. The proposed model introduces a hierarchical Unique Identifier (UID) system to ensure consistent lineage and cross-layer linkage, along with a structured edge framework to represent containment, adjacency, functional, and administrative relationships. A validation-driven methodology is developed to transform multi-layer geospatial datasets into graph-enhanced representations through data standardization, UID assignment, and rule-based edge generation. The framework is implemented using a large-scale urban dataset from Riyadh comprising approximately 10,859 nodes and 13,733 edges across administrative, transportation, and urban asset layers. Graph-based evaluation demonstrates realistic spatial patterns, including right-skewed degree distribution, strong network connectivity, and identifiable community structures. Comparative analysis shows that GeoGraphJSON effectively bridges the gap between geometry-centric GIS formats and graph-based models, enabling efficient querying and direct integration with graph analytics workflows. The results highlight the model’s scalability and potential for supporting urban analytics, digital twins, and AI-driven geospatial systems.

Remote sensing images (RSI) play a central role in the regular update of land cover datasets. A first step in this process is RSI semantic segmentation, which is mainly solved by deep learning today. Vision foundation models (VFM) have gained increasing importance in this context. Having been trained on large datasets, VFM for segmentation can yield good results on data from various domains without further training. We present a new method for using the VFM Segment Anything Model 2 (SAM 2) for multi-class semantic segmentation of Sentinel-2 images that does not require training data. Our method is based on a prompt engineering approach, using SAM 2 in its pre-trained form and generating different prompt types on the basis of existing topographic data. We also propose a post-processing step for merging the output of SAM 2 to obtain a multi-class label image. The results of our experiments show that our method achieves an overall accuracy (OA) of up to 93% at pixel-level using polygon mask prompts, while using point and box prompts leads to an OA of 81% and a mF1-score of 72%. Experiments with other Sentinel-2 composite images do not show significantly different results compared to R-G-B images. Incorporating data from different time steps for map updating shows good results, but remains inconclusive due to the small amount of change in the dataset.

Illegal waste disposal sites pose significant ecological and public-health risks yet remain difficult to track with traditional field inspections. We propose a multi-agent detection framework that fuses textural, spectral, and contextual cues from medium-resolution satellite imagery for this work. Three specialized agents - Waste-Pile, Road, and Industry detectors - are implemented as YOLO (You Only Look Once) convolutional models that generate partial hypotheses, which are then hierarchically aggregated through rule weights learned from expert-labelled samples. The system provides an interpretable set of object relations, allowing regulators to trace how individual cues contribute to the final decision. The method was validated on an independent test area near Taromske (Dnipropetrovsk region, Ukraine) and corroborated by ground surveys. Joint aggregation raised the posterior probability of the primary target cluster from 0.27 (single-detector confidence) to 0.91, while maintaining robustness to label noise and heterogeneous sensor characteristics. Compared with conventional CNN baselines, the proposed approach delivers three key advantages: explicit explainability of outputs, transferability to 10 m spatial resolution without extensive retraining, and seamless integration of heterogeneous evidence sources. The proposed framework can serve as a cost-effective backbone for regional and national waste-monitoring systems. Future work will focus on near-real-time processing of Sentinel-2 time series, incorporation of hyperspectral and thermal methane indicators to assess remediation stages, and extension of the array of features to other anthropogenic disturbances such as open-pit mining and construction debris.

Advancing the level of digitalization and automation in concrete manufacturing can substantially contribute to lowering CO2 emissions associated with the concrete production. This work introduces a new methodology for predicting the time-dependent properties of fresh concrete during mixing. For the prediction, a deep learning network is created which uses stereoscopic image sequences of the flowing material together with tabular data as input. Besides mix design parameters and process state data, like energy consumption, moisture and fresh concrete temperature, temporal information is included in the tabular data. The temporal information represents the time interval between image acquisition and the time for which the properties should be predicted. During training, this interval corresponds to the difference between the image acquisition and the time at which reference measurements are taken, allowing the network to implicitly learn the temporal evolution of the material properties, namely the slump flow diameter, yield stress, and plastic viscosity. Incorporating time-dependent prediction enables the forecasting of property changes throughout the mixing process, offering a valuable tool for real-time process control. This capability allows timely adjustments whenever deviations from the desired material behavior are detected. The experimental investigations presented in this paper demonstrate the feasibility of this method under realistic conditions.

3D multi-view, multi-object tracking (3D MV-MOT) makes use of multiple cameras to reduce the number of missed detections and to mitigate occlusions. Most current 3D MV-MOT methods suffer from information loss when associating 3D locations with 2D image features via a 3D-to-2D projection, as they use a discrete grid in 3D and sample image features only at the projected centers of each grid cell. Thus, all other feature information is lost. An additional information loss commonly arises during crossview aggregation when applying max or average pooling: these methods either overemphasize a single view or treat conflicting views, that depict different entities, e.g., due to occlusions, equally. In this work, we introduce two novel modules for 3D MV-MOT, employed to pedestrian tracking, that target these limitations: (i) VoxROI aggregates all image features that fall within the bounding box around a voxel’s projection into each respective image, instead of only sampling features at the projected voxel center. (ii) SimFuse aggregates per-view voxel features into one coherent feature representation per voxel, using similarity weights computed from re-identification (Re-ID) features. Subsequently, they are used to measure cross-view identity similarity. Views with higher Re-ID feature similarity receive larger weights, while inconsistent views are suppressed. Experimental results on the WildTrack dataset confirm our method’s effectiveness for multi-view pedestrian detection and tracking, reaching, and in particular in cross-view scenarios improving, the general state-of-the-art. The approach maintains strong performance across different camera configurations, demonstrating its generalization capability when training and testing on different camera setups.

Sustainable Development Goal (SDG) 11.3.1 monitors land use efficiency (LUE) through the relationship between the land consumption rate (LCR) and the population growth rate (PGR), yet current applications have largely remained retrospective and offer limited foresight for policy and planning. To address this gap, this study develops a scalable probabilistic deep learning (DL)-based framework that jointly forecasts built-up area (BU) and population (Pop) and propagates these forecasts into LCR, PGR, and their ratio (LCRPGR). The framework is based on a lightweight hybrid model composed of one-dimensional convolutional neural networks (Conv1D) and a recurrent network with Long Short-Term Memory (LSTM) units, trained using quantile loss to produce interval-bounded probabilistic forecasts. Ensemble aggregation was used to improve robustness, with a calibration stage included to ensure reliable coverage of prediction intervals. Applied to 8,478 urban centres using historical built-up area and population data from the Global Human Settlement Layer – Urban Centre Database (GHS-UCDB) 2025 edition, the framework generated probabilistic forecasts for the period 2025-2030, with prediction intervals shown to be well-calibrated on historical data. Results indicate that total built-up area is projected to expand from 108,808 km² in 2020 to 118,123 km² by 2030 (+8.6%), while population is expected to grow from 2.875 to 3.159 billion (+9.9%). At the aggregated level, LCRPGR values suggest a general trend toward efficiency, with PGR slightly outpacing LCR, though uncertainty intervals cross the efficiency threshold. While the implementation addressed the constraints of SDG 11.3.1 forecasting (short time series, limited covariates, and large-scale application), the framework remains flexible. It can be extended with denser time series, richer covariates, and spatiotemporal architectures. This study offers a practical basis for near-term LUE monitoring and risk-aware policy action as 2030 approaches by providing reproducible and uncertainty-aware forecasts.

Earth observation (EO) has become central to monitoring progress toward the Sustainable Development Goals (SDGs), particularly SDG Indicator 11.3.1, which assesses land-use efficiency (LUE) through the ratio of land consumption rate (LCR) to population growth rate (PGR). Current EO-based implementations remain predominantly retrospective and deterministic, relying on historical mappings of built-up area and population. However, decision-making for sustainable urban development increasingly requires forward-looking and uncertainty-aware information. This paper argues that the future of EO-enabled SDG monitoring lies in uncertainty-aware forecasting rather than deterministic retrospective assessment. We present a global EO-driven framework that integrates deep-learning-based forecasting of built-up area and population with Monte Carlo uncertainty propagation to derive probabilistic projections of SDG Indicator 11.3.1. Using multi-decadal EO-derived time series for 8,478 urban centres worldwide, we demonstrate how forecast uncertainty is structurally transmitted through SDG Indicator 11.3.1 and how deterministic forecasts can mask substantial uncertainty in future LUE classifications. The results highlight forecasting and uncertainty quantification as essential components of next-generation EO analytics for sustainable development.

Forest fires pose severe ecological and socio-economic threats, necessitating efficient tools for rapid damage assessment. This study presents a machine learning approach for detecting burnt forest areas in Ukraine using multispectral Sentinel-2 imagery. A new manually annotated dataset was developed for training semantic segmentation models, addressing the scarcity of open data for the region. The proposed convolutional neural network, based on an encoder–decoder architecture with Xception blocks, effectively captures spectral patterns associated with fire damage. Experiments conducted on Sentinel-2 Level-2A imagery of the Kinburn Peninsula (October 2022) demonstrate high detection performance, achieving an Intersection over Union (IoU) of 95%. The results confirm the model’s capability for accurate burnt-area mapping and highlight its potential for broader applications in regional fire monitoring and environmental management.