Deep Learning-Based Wetland Vegetation Classification: The Impact of Spectral and Spatial Features

Wetlands are intricate and dynamic ecosystems essential for biodiversity conservation, water purification, and climate regulation. Effective management and conservation planning for wetlands, particularly emergent marsh ecosystems of high conservation priority, require precise vegetation classification. However, traditional classification methods often struggle to address the spatial heterogeneity and complexity of wetlands. In addition, the high costs and time demands of repeated, accurate surveys have hindered efforts to monitor plant species and vegetation cover in semi-permanent emergent marshes. This study integrates deep learning models with high-resolution imagery from the National Agriculture Imagery Program (NAIP), leveraging spatial and spectral features to classify and evaluate wetland vegetation in semi-permanent emergent marshes within the Arkansas Delta. The proposed framework provides a robust and automated approach for detecting and assessing changes in vegetation cover. Results demonstrate that incorporating both spectral and spatial features significantly improves classification accuracy, particularly for vegetation categories with overlapping spectral signatures. The Kappa coefficient ranged from 0.56 to 0.94, while overall accuracy spanned 67.8<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> to 95.5% across various combinations. Statistical analyses further revealed significant interactions between band combinations, rotation effects, and model performance, influencing both accuracy and processing time. These findings offer valuable insights for wetland managers and stakeholders, particularly for estimating habitat quality for species such as the King Rail (Rallus elegans) and assessing energetic resources for wintering waterfowl. By applying deep learning to wetland vegetation classification, this study highlights the transformative potential of scalable and efficient monitoring tools, paving the way for innovative and data-driven conservation strategies.