Conference Agenda

This project, a collaboration between the Landesamt für Geobasisinformation Sachsen (GeoSN) and M.O.S.S. Computer Grafik Systeme GmbH, pioneers the development of a Digital Twin Forest prototype for Saxony. The initiative leverages high-resolution aerial orthophotos (DOP) and advanced AI methods to generate detailed, current forest information. The core methodology centers on the “DeepTrees” workflow, a convolutional neural network (CNN)–based approach developed by the Helmholtz Center for Environmental Research (UFZ). This workflow processes DOP imagery at 10–20 cm resolution to segment individual tree crowns and extract key forest indicators, including crown area, crown radius, and tree density.

The process unfolds in three main stages: (1) preprocessing and model adaptation using transfer learning, (2) inference and postprocessing for accurate tree segmentation, and (3) integration into GeoSN’s data infrastructure. This integration utilizes OGC-compliant services and moGI-based data management, enabling automated processing, configuration, and visualization.

Results from the prototype confirm the feasibility of precise, large-scale tree crown segmentation from aerial imagery. The system also demonstrates the potential to derive temporal and structural forest information from recurring DOP datasets. These outputs can be directly incorporated into operational geospatial systems, supporting climate adaptation, forest management, and policy-making.

In conclusion, the Project showcases how explainable, interoperable AI workflows can strengthen national geodata infrastructures and serve as a model for future federated, AI-driven digital forest twins across Germany.

This study presents a scalable framework for mapping aboveground biomass and moisture content in peatlands using intensive field sampling, multi-sensor satellite imagery, and advanced machine learning. Field data collected from diverse bog and fen sites in Western Newfoundland are integrated with Sentinel-1/-2 synthetic aperture radar and optical data, complemented by 3 m PlanetScope imagery for site-level detail. Ensemble learning models, particularly XGBoost, yield high biomass mapping accuracy, with regional maps revealing major biogeographical gradients and fine-scale site mosaics. Feature importance analysis highlights the role of red-edge and SAR bands in prediction. The results demonstrate that free satellite archives and machine learning can overcome limitations of costly airborne campaigns, supporting operational carbon monitoring and ecological management in northern peatlands. This approach establishes a foundation for wide-area wetland monitoring and future expansion using emerging remote sensing technologies.

Surface soil moisture (SM) plays a significant role in environmental and hydrological processes, particularly runoff and evapotranspiration. Within forest ecosystems, changes in SM can lead to significant ecological impacts, including paludification and greater susceptibility to forest fires. Microwave remote sensing facilitates large-scale monitoring of SM. Moreover, machine learning (ML) have demonstrated strong potential for capturing the nonlinear relationships between SM and satellite data. In general, supervised ML techniques achieve higher success rates when trained on larger ground measurements. However, obtaining extensive ground measurements of SM over vast areas such as forests is challenging, expensive, and time-consuming. To address this limitation, this study proposes a self-supervised method based on pre-task learning to estimate SM over forested areas using multisensor data. The core idea of the self-supervised approach is to leverage the knowledge gained during pre-task learning from multisensor data and transfer it to the SM estimation task, thereby improving the model’s generalization ability to SM estimation. The self-supervised learning method achieved an overall coefficient of determination (r²) of 0.74 and an RMSE of 0.04 m³/m³ on the testing dataset By focusing on each forest site, the model obtained r² = 0.75 with RMSE = 0.04 m³/m³ at Millbrook, r² = 0.63 with RMSE = 0.04 m³/m³ at Massachusetts, and r² = 0.74 with RMSE = 0.03 m³/m³ at Saskatchewan. The results highlight the potential of multisensor data for SM estimation in forested areas. Our method, which utilizes self-training on the input data, reduces dependence on ground SM measurements and enhances generalization capability.

Land cover data need to be updated regularly. Typically, remote sensing images (RSI) play a central role in this process. A first step is RSI semantic segmentation. Today, this task is mainly solved by deep learning. Especially 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. The different prompt types are generated 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 mask prompts. Experiments with other Sentinel-2 3-channel composite images do not show significantly different results compared to R-G-B images. Incorporating data from different time steps, intended to be used for map updating, shows good results. But the small amount of changed areas indicate limitations. In general, the proposed method is suitable for further

research into semantic segmentation tasks with little or no training data, as well as for the process of updating databases.