Advanced AI-Supported Peatland Vegetation Mapping using Remote Sensing for Environmental Monitoring

Peatlands are complex ecosystems that provide a variety of ecosystem services, including the regulation of water. Peatlands also store huge amounts of carbon, thereby contributing to long-term climate protection. However, drainage and overuse are threatening these positive functions in many places. Consequently, numerous countries around the world have developed peatland conservation and restoration strategies. To support this process, monitoring approaches are needed that are suitable for areas that are difficult to access and for large-scale applications.

In previous studies, we conceptualized a methodology for the scalable monitoring of peatlands. A key component of this methodology is the automatic recognition of peatland plants. This methodology involves high-resolution drone images and metadata as input for an ecologically informed machine learning framework. We employ state-of-the-art deep learning segmentation architectures, such as DeepLabv3+ and OCRNet, which utilize a high-resolution network (HRNet) backbone. As a first step, we investigated the detectability of individual species, focusing on species for the initial class set for training the model.

Current project developments include expanding data-fusion strategies, model-architecture validation, and conceptualizing a new label strategy by introducing new vegetation class sets to address ecological issues and broaden applicability. We benchmarked multiple vegetation-classification architectures and optimized key hyperparameters via grid search to identify a robust domain-specific model. Auxiliary metadata (e.g., temperature sums, cloud cover) were integrated at different stages and early fusion (embedding metadata in the input data cube) techniques were compared with late-fusion approaches such as FiLM and feature weighting. Explainable AI was employed to identify the inputs that have the most significant impact on training and predictions. Vegetation indices (NDVI, EVI) were added as explicit input channels. As an additional target we evaluated plant dominance stand types instead of single species to better capture mixed stands. Furthermore, we expected dominance stands-based mapping to better support the integration into the GEST (Greenhouse-gas-Emission-Site-Type) approach and other applied peatland monitoring frameworks.

After classification, predicted vegetation/dominance patterns were combined with water-table maps. By using the GEST approach, spatially explicit peatland greenhouse gas emission estimates were derived and validated against a reference area. A Minimum Viable Product (MVP) combining vegetation maps, hydrological inference, and GEST-based emissions shall provide initial large-scale assessments of rewetting success and associated emission reductions. Further fields of application regarding the monitoring of ecosystem services and smart farming approaches for paludiculture will be investigated based on the results obtained.