To improve planning for the long-term operation of the dredging lake, determining residual gravel quantities—the amount of gravel that can still be extracted at the same lake size (depth before surface)—is essential. Previously, these quantities were derived from terrestrial drilling located as close as possible to the deposit. Since this typically involves only one borehole per lake, no information is available on changes in deposit quality and thickness across the distance from the borehole over the lake. Variations of just a few meters in thickness or changes in quality can make a decisive difference for lake operations. Results from “Kiesdetektion” have shown that medium-frequency, compact seismic echosounders can penetrate fine sediments very well but not the gravel layer. As an alternative, this work package will test and evaluate the application of a UHRS sparker system in a dredging lake for the first time. The use of these powerful, low-frequency seismic systems has not been feasible in smaller, shallow water bodies with steep flanks or has not produced usable results.

In collaboration with Geo Marine Survey Systems, a prototype will be provided for this project (see letter of support). The newly developed single-layer multi-channel UHRS (ultra-high resolution seismic) system will be adapted to the conditions of a dredging lake (or reservoir). It will combine a smaller energy source (~1 kJ) with a specially adapted high-voltage cable and a custom-designed multi-channel streamer (for acoustic signal reception). This will enable use on small boats while minimizing the negative impact of side echoes (caused by steep slopes) via the multi-channel streamer. Target penetration depths are 20-40 m into the gravel bed with a vertical resolution of 15-30 cm. Due to the still considerable weight and dimensions of the equipment components, a technical modification of the limknow company boat will likely be necessary. This adaptation is also part of the feasibility and development efforts, as rapid and flexible sparker deployment is relevant for future applications.

Measurement campaigns are planned at the dredging lake in Freistett (FS) and one in Niederrimsingen (NR) to compare different deposits. Data from the Niederrimsingen lake will additionally be combined with sediment information and seismic data from “Kiesdetektion” to add value in evaluation and gain better understanding of layering. The sparker seismic data will be processed and visualized in collaboration with the Geo Marine Services team and prepared for data fusion. Starting from the nearshore borehole, acoustically detected layers will be traced across the lake and correlated with borehole layers. Since the drilling equipment used in “Kiesdetektion” cannot penetrate gravel layers further, validation of upper sediment layers for classification at softer locations will primarily use short cores and additionally the GraviProbe. For better validation, since submission of the sketch, another option has been considered. We would like to note that we are submitting a parallel application to the ESA BIC (https://www.esa-bic-bw.de/apply/) to fund development of a vibrocoring system. This would create significant added value for KiesVision. A vibrocorer can penetrate more compact layers via vibration and self-weight. The goal is to retrofit the KIT platform already used for a vibro-system and conduct a series of test boreholes. This could initially achieve core lengths of 3-6 m.

The aim is to optimize dredging operations for sediment removal through acoustic classification of upper sediments (0-2 m depth). Using a single-beam multi-frequency echosounder (EA440), the upper sediments will be acoustically categorized into gravel, sand, and fine material groups based on their physical properties. Digital maps or 3D models can be created to enable targeted extraction of specific sediments or adaptation of the extraction tool to the sediment type. The drone boat “Calypso” from the IPF, originating from the Kiesdetektion project, will be retrofitted as the platform for the multi-frequency echosounder. Key improvements include upgrading the GPS to a GPS-RTK system for centimeter-accurate positioning and technical adaptations for communication between the Calypso system and the EA440.

For classification, backscattered echoes at specific frequencies are decomposed into time slices, and the energy contained within them is calculated. The amplitude and energy distribution over time provide insights into the physical properties of the sediments. Three frequencies are expected to be used: 200 kHz for surface classification, and 38 kHz and 15 kHz for penetrating into the sediment to classify the upper sediment volume. Limknow GmbH contributes extensive expertise in seabed classification to the project [1][2]. Since wash sludges in gravel pits contain only minimal organic material, there are no gas (methane) inclusions that could interfere with the signals. Limknow will purchase the echosounder unit as an investment. Transducers/transmitters will initially be rented for the project duration or measurement campaigns to test the optimal frequency combination. Transducers differ not only by frequency (200-38-15 kHz) but also by available beam angles at the same frequency. Narrower beam angles result in larger and heavier transducers, so the weight and dimensions compatible with drone boat operation must be verified.

The previously tested methodology with parametric echosounders successfully detects sediment layers with sufficient spatial resolution for estimating sediment volumes in water bodies. However, it cannot be used for acoustic sediment classification, as parametric echosounders detect impedance boundaries in sediments but not physical changes within the sediment volume.

This work package does not primarily involve developing a new classification approach; instead, it focuses on an optimized (drone-compatible) campaign design, novel application to “pure” mineral sediments, and ultimately fusion with additional hydroacoustic data. To our knowledge, there are no scientific or commercial applications of sediment classification in gravel pits to date.

The added value arises not primarily from the classification approaches themselves, but rather from improved data processing and transfer to a new application for “optimized dredging.” This involves automated surveys before, during, and after sediment removal, with temporal resolution adapted to dredging operations. Parallel to this, classification will also track redeposited sediments during relocation, allowing inferences about water body ecological impacts and adjustments to operations. From a technical perspective, repeatedly surveying the same areas during sediment relocation is particularly promising. This opportunity is likely provided by accompanying the relocation project in the Freistett-Rheinau gravel pit.

Validation of the classification will use a free-fall penetrometer (GraviProbe) and extraction of short cores (up to 1.8 m length). Penetrometer measurements allow inferences about upper sediment layering, firmness, and grain size. Sediment cores will be analyzed for layering, wet bulk density, and grain size distribution. At least two classification campaigns are planned in the Niederrimsingen dredging lake and two during relocation in Freistett.

High-resolution data from the EA440 with shallow penetration depth are available from the work packages on seismic exploration (A) and seabed classification (B), alongside seismic data from the UHRS with lower resolution but much greater penetration depth. Additionally, seismic data from the SES 2000 Compact Sub-Bottom Profiler (Innomar) were already collected in the Kiesdetektion project at Niederrimsingen. These lie, in terms of frequency and penetration depth, between the EA440 and UHRS.

The goal of the data fusion work package is to integrate all datasets and fuse them into a common subsurface model with improved classification. The result will be a user-friendly GUI-based software solution in Python or Matlab that imports various data sources or pre-processed classification results, homogenizes them, and fuses them to create the best possible resolved subsurface model.

Insights and methodological studies from the Kiesdetektion project feed into the classification of seismic data. Generally, an approach is chosen where relevant features from seismic lines are first extracted using Convolutional Autoencoders (CAE). In addition to direct seismic signals, derived attributes such as the analytic signal and envelope—analogous to exploration seismics—are utilized. Based on the CAE features, seismic line sections are classified in a second step using unsupervised clustering algorithms. To date, k-means clustering has been used for this. Within the methodological development of the KiesVision project, the approach will be extended to other clustering algorithms (e.g., DBSCAN) and manifold learning techniques (e.g., UMAP and t-SNE). Classification results will be validated against existing and planned core sample data.

While a 1D-CAE has been used to date for ML-based classification, evaluating each seismic trace individually, KiesVision will extend this to a 2D-CAE. Accounting for adjacent traces in a 2D-CAE promises more robust classification, with slightly reduced horizontal resolution but improved vertical resolution. However, given the predominantly horizontal layering of sediments, horizontal resolution is of secondary importance compared to vertical. Additionally, the proven moving-window approach will be applied in both dimensions for the 2D-CAE to enable optimal localization of layer boundaries.

For the actual data fusion, two different approaches will be tested in the project. Feature-level fusion combines derived features or sediment body classes from individually processed datasets of different sensors at the CAE feature level. The added value of data fusion arises, on the one hand, from multi-frequency integration with increased classification prediction probability (Dempster-Shafer evidence theory) and, on the other, from continuous extension of classification through the sediment body. Fusion of datasets post-clustering is also conceivable. In contrast, data-level fusion merges data from different sensors at a very early stage (raw data). Specifically, seismic traces from different datasets are fused, allowing the CAE to derive features from a fused seismic dataset. This type of data fusion requires very precise georeferencing, including homogenized depth information for all datasets—especially for 2D-CAE. Multiple reflections and topographic effects must also be more strongly accounted for, which must be reflected in campaign design concepts (see 2.2). Prospective application of level fusion will be evaluated in the project.

Additionally, the extent to which penetration measurements from the GraviProbe, as well as sediment samples from short cores (Freistett-Rheinau and Niederrimsingen) and long cores (Niederrimsingen Kiesdetektion project), can be integrated or potentially fused will be investigated. The fused subsurface results will be fed into the MARPO block model.

SPE’s excavation control system currently aims to provide the dredge operator with all important information on the dredging process via suitable sensors. Clear yet comprehensive presentation of the information is the top priority, enabling optimal control and positioning of the excavation tool and floating unit. Specifically, positions of the dredger, excavation tool, predefined dredging fields—including track points and target/actual water bottom depths—are displayed in various sectional views or a 3D view from any angle.

Figure 2: Example of the MARPO model interface

The exploration data obtained and classified in this project will be fed into the MARPO system as a block model. To our knowledge, there is no comparable representation worldwide. Extensive modifications to the existing system are necessary to achieve this. The first step will be to evaluate the feasibility of data integration. MARPO currently works with rendering various boundary surfaces, which may prove unsuitable for complex geological structures. Thus, adaptation to volume body representation must be tested. Once the display issue is resolved, the detail accuracy of individual surfaces or bodies must be evaluated. Here, detail accuracy must be balanced against required computational power. Since the system is used not only for visualization but for actual excavation, real-time capability must be ensured. For reduced storage needs in representing the geological model, structured grids—where base geometry is predefined by regularity—are recommended. If system functionality with the newly integrated data proves viable, unstructured grids will be evaluated for potential use. This would allow more precise capture of areas with high spatial variability.

Furthermore, it will be necessary to assess whether simplifications of the provided geological subsurface are required to create a system-compatible model. The three-dimensional representation of the subsurface and various materials in the software enables highly targeted excavation. The operator/dredge operator can plan and execute extraction precisely according to material needs via the clear visualization, thereby avoiding time and energy loss, as well as extraction or deposition of unwanted materials over the target raw material. Unexpected material changes can lead to slope failures if the new sediment strength (cohesion forces) no longer matches the planned excavation angle. Knowledge of the subsurface allows anticipation and adjustment, reducing this risk. Precise deposit knowledge is essential to enable automated dredging and minimize risks.