Deep learning for wetland vegetation mapping in the southeastern United States: evaluating site-specific accuracy and classification challenges

Accurate wetland mapping is essential for monitoring ecosystem health and guiding conservation planning efforts. This study evaluated the effectiveness of deep learning architectures and high-resolution National Agriculture Imagery Program (NAIP) imagery for classifying diverse vegetation cover across five public land holdings in three states with significant seasonally flooded, emergent wetland habitats in the southeastern United States. We built and tested deep learning models in ArcGIS Pro using three architectural encoders (U-Net, PSPNet, and DeepLabV3) with different input combinations, including NAIP natural color (RGB), Near-Infrared bands (NIR), Normalized Difference Vegetation Index (NDVI), Digital Elevation Models (DEM), and Texture metrics derived from NAIP imagery using Gray Level Cooccurrence Matrix (GLCM) features, resulting in eight (8) input combinations and three (3) rotational augmentations (0°, 30°, and 60°). Results revealed differences in classification accuracy ranging from 71 to 90% across sites, with Kappa scores from 0.54 to 0.81. We also found differences in model performance under alternating conditions and in the impact of combinations on performance metrics and computational time. Classification of the forested and scrub-shrub classes was consistent with F1 scores above 0.90 at several sites. However, distinguishing open water, aquatic beds, and emergent marshes became more difficult for some model-training architectures. At Mingo National Wildlife Refuge, clear spectral signatures and high-quality training data yielded the highest accuracy, thereby improving classification performance. In contrast, classification accuracy was lower for Red Slough and Choctaw East Wildlife Management Areas. Across all scenarios, the U-Net and DeepLabV3 models consistently outperformed the PSPNet model, achieving mean F1 scores close to 0.91, whereas PSPNet underperformed, with several configurations falling below 0.80. These results explain that while deep learning methods are effective and adaptable across different wetland types, their success depends on the ecological diversity of these environments. Overall, this research highlights the potential of deep learning for wetland monitoring and its crucial role in supporting conservation and management efforts.