Benchmarking Deep Learning for Wetland Mapping in Denmark Using Remote Sensing Data

Abstract

Wetlands are critical ecosystems providing numerous ecological services, yet they face significant threats from human activities and climate change. Therefore, accurate mapping and monitoring of wetlands are crucial for formulating effective conservation and restoration strategies. While remote sensing combined with deep learning (DL) offers a promising solution, inconsistencies in wetland classification systems—where different regions define wetland types based on their policy frameworks and conservation priorities—limit the applicability of these models. Such inconsistencies make it difficult to assess their limitations in different contexts. Notably, no study has yet leveraged DL for mapping wetlands within Denmark's unique wetland classification system, as defined by the Danish nature conservation framework. Therefore, this article presents a comprehensive benchmark analysis of several DL models for wetland classification in Denmark. We utilize publicly available high-resolution multispectral aerial imagery and digital elevation models (DEMs) and evaluate the performance of three well-established network architectures: Fully Convolutional Network, U-Net, and DeepLabV3. We also assess the impact of incorporating near-infrared and DEM data in addition to traditional optical imagery. The results show that DeepLabV3 model outperforms other models, particularly when additional data layers are included, achieving the highest overall accuracy and F-measure score. Our findings also reveal that while DL models can effectively classify certain wetlands, challenges remain in distinguishing wetland with ecological similarities and in handling noisy labels. This benchmark provides a foundation for future work aimed at improving DL methods for wetland mapping in Denmark.