Benchmarking deep learning models for wetland mapping in Denmark using high-resolution earth observation data

Wetlands are essential ecosystems providing critical ecological services, yet they face significant threats from human activities and climate change. Monitoring and mapping these areas accurately is fundamental to formulating effective conservation and restoration strategies. Remote sensing, combined with advanced deep learning techniques, offers a scalable and efficient solution for wetland classification and monitoring. However, the application of these technologies is often constrained by regional variations in wetland classification systems and the challenges of distinguishing ecologically similar wetland types. Notably, no study has yet leveraged deep learning for mapping wetlands within Denmark's unique wetland classification system, as defined by the Danish Nature Conservation framework.

This study presents a comprehensive benchmark analysis of three state-of-the-art deep learning models—Fully Convolutional Network (FCN), U-Net, and DeepLabV3—for wetland segmentation using high-resolution Earth observation data. We utilize the publicly available multispectral aerial imagery (RGB and NIR) and digital elevation models (DEM) to classify Denmark’s wetland areas, such as bogs, freshwater meadows, and salt marshes. By evaluating multiple input configurations, this study investigates the impact of integrating additional spectral and elevation data on the segmentation performance.

The results demonstrate that the DeepLabV3 model outperforms other architectures, achieving the highest accuracy and F-measure when leveraging the combined RGB, NIR, and DEM data. Despite these advancements, challenges remain, particularly in distinguishing ecologically similar wetland types (e.g., freshwater meadows and bogs) and addressing issues of label noise in ground truth datasets. This study highlights potential solutions, such as the inclusion of Synthetic Aperture Radar (SAR) data for temporal analysis and the adoption of noise-robust training and contrastive learning methods to enhance model robustness.

This benchmark not only establishes a foundation for improving deep learning methodologies for wetland mapping in Denmark but also contribute to global efforts aimed at developing innovative, scalable solutions for wetland conservation and restoration.