Cost-effective statewide wetland inventory update using weakly supervised deep learning: A case study in Minnesota, USA

ABSTRACT Wetlands provide important ecosystem services, including water purification, flood regulation, carbon storage, and habitat for diverse species. Despite their importance, North America continues to see significant loss of wetlands due to development, agriculture, and climate-related pressures. These ongoing declines threaten ecological integrity and reduce the capacity of wetlands to provide essential services. As a result, regular updates and monitoring are essential to protect these important ecosystems to support evidence-based management and meet the needs of evolving conservation policies. One cost-effective method for monitoring wetlands is the segmentation of satellite images. Automating the segmentation of remote sensing images to update land cover maps enhances the frequency of map production, enabling more timely and efficient monitoring. Deep learning models such as Convolutional Neural Networks (CNNs) have performed well for segmentation. Still, the need for large, densely annotated datasets has limited their adoption in remote sensing. Meeting this requirement poses substantial challenges for regular map updates because of the extensive number of labels, the complexity of annotations, and the significant time and financial resources required for field data collection campaigns. Therefore, this paper targets Minnesota's state-wide wetland monitoring by training deep CNNs with weak labels extracted from existing thematic products. Our approach obtains training samples from existing land-cover maps by applying change detection to identify stable pixels and then refining labels with objects produced by the Simple Non-Iterative Clustering (SNIC) algorithm. The resulting weakly labelled samples are used to train and evaluate U-Net++ and DeepLabV3+ architectures. The proposed method achieved robust performance in the study area, with an average F1-score of 91.3 % for U-Net++ across seven analyzed classes, compared to 90.6 % for DeepLabV3+ and 88.3 % for Random Forest (RF). Notably, U-Net++ outperformed the other models, indicating that dense skip connections effectively classify complex remote-sensing scenes such as wetlands. These results demonstrate the effectiveness of the proposed weak-label-driven deep learning workflow for large-scale wetland-inventory mapping in Minnesota, while remaining generalisable to other land-cover classification problems. Also, they support stronger wetland conservation by enabling clearer detection of wetland extent and condition. More precise maps also enhance carbon accounting and help identify priority areas for restoration. This scalable workflow provides timely, cost-efficient updates essential for informed management decisions.