The Roles of Spectral and Spatial Features in Enhancing Wetland Vegetation Classification Using Deep Learning

Problem StatementWetlands are complex and dynamic ecosystems that play an important role in biodiversity conservation, water purification, and climate regulation. The King Rails, a secretive marsh bird species of significant concern in Arkansas, depend on emergent marshes, characterized by shallow-water wetlands that sustain diverse plant communities of emergent, floating leaf, and aquatic submerged vegetation, as their primary home. The secretive behavior of marsh birds, their infrequent vocalizations, low visibility, and the challenging terrain of their habitat makes it difficult for researchers to evaluate their population dynamics and understand their particular ecological requirements. However, plant species and vegetation cover of semi-permanent emergent marsh are less studied because monitoring is time-consuming and expensive and often requires frequent surveys with high-precision data on the landscape. Often, traditional classification methods struggle with wetlands’ heterogeneous and spatially complex nature.ObjectiveThis study aims to (i) provide an accurate classification of wetland vegetation cover, particularly emergent marsh ecosystems of conservation concern that support diverse flora and fauna, which is essential for effective management and conservation planning efforts, and (ii) investigate the impact of integrating spectral and spatial information in deep learning-driven wetland vegetation classification, with augmentations applied at five rotation angles (0°, 15°, 30°, 45°, and 60°) for each input combination. The research will evaluate the performance of various deep-learning architectures in classifying vegetation in semi-permanent emergent wetlands using high-resolution NAIP data. We will achieve this by using remote sensing to classify and assess wetland vegetation cover in semi-permanent emergent marshes within the Arkansas Delta.MethodsTo address this problem, Deep learning algorithms, known for their proficiency in handling high-dimensional data, and National Agricultural Imagery Program (NAIP) data, which provides considerable benefits for observing vegetation characteristics at a landscape scale because of its high-resolution and multispectral imagery features, were used.These algorithms, combined Near-Infrared (NIR) bands, Normalized Difference Vegetation Index (NDVI), NAIP statistical bands, and digital elevation models, were utilized to create composite bands for generating additional input combinations to improve the reflectance of the pixels and the overall performance of the models. The composite bands were used to export training datasets for deep learning model training, with augmentation applied at five rotation degrees (0, 15, 30, 45, 60) for each combination. Randomly rotating the training imagery enhances rotation invariance, enabling the model to identify features irrespective of orientation, while mitigating overfitting by augmenting the effective dataset size. Three model types (U-Net, PSPNet, and DeepLabV3) and three backbone classifiers (ResNet18, ResNet34, and ResNet50) were used to train and configure the models.The entire model training process was carried out using the ArcGIS Pro software. The accuracy and performance of each model were further quantified through statistical analysis.ResultPreliminary findings show that wetland distribution and composition change dynamically because of various causes, including natural disturbances, land use changes, and agricultural practices. The Kappa coefficient and overall accuracy across all combinations range from 0.56 to 0.94, with values of 67.8% and 95.5%, respectively. Further statistical analysis indicates a significant correlation between the main effects of band combinations and rotation and their interactions on the models’ overall performance and processing time. Utilizing various input combinations significantly enhanced model performance, with F1 scores varying considerably depending on the specific combinations used. This highlights the importance of selecting appropriate input features to balance precision and recall, improving the model’s overall classification accuracy.ConclusionThese findings are essential for offering significant insights into assessing wetland birds' habitat value and energetic carrying capacity, including King Rail (Rallus elegans) and other relevant species. Applying deep learning to classify wetland vegetation cover shows considerable promise to enhance conservation initiatives by offering efficient and scalable monitoring methods. Deep learning models applied to NAIP data can facilitate precise spatial evaluations of vegetation cover and variety of resource management and conservation planning in wildlife management areas (WMA) and national wildlife reserves (NWR). Incorporating NDVI, DEM, and NAIP statistical data significantly improved classification accuracy, underscoring the significance of photosynthetically active vegetation and topographic information in capturing environmental variability, while a well-defined pixel reflectance aids in optimizing the model's precision in detecting and recalling pixels.