Mapping Wetland Plant Communities Using Unmanned Aerial Vehicle Hyperspectral Imagery by Comparing Object/Pixel-Based Classifications Combining Multiple Machine-Learning Algorithms

Abstract

Understanding the spatial patterns of plant communities is important for sustainable wetland ecosystem management and biodiversity conservation. With the rapid development of unmanned aerial vehicle (UAV) technology, UAV-borne hyperspectral data with high spatial resolution have become ideal for accurate classification of wetland plant communities. In this article, four dominant plant communities (Phragmites australis, Typha orientalis, Suaeda glauca, and Scirpus triqueter) and two unvegetated cover types (water and bare land) in the Momoge Ramsar wetland site were classified. This was achieved using UAV hyperspectral images and three object- and pixel-based machine-learning classification algorithms [random forest (RF), convolutional neural network (CNN), and support vector machine (SVM)]. First, spectral derivative analysis, logarithmic analysis, and continuum removal analysis identified the wavelength at which the greatest difference in reflectance occurs. Second, dimensionality reduction of hyperspectral images was conducted using principal component analysis. Subsequently, an optimal feature combination for community mapping was formed based on data transformation (spectral features, vegetation indices, and principal components). Image objects were obtained by segmenting the optimum object feature subsets. Finally, distribution maps of communities were produced by using three machine-learning classification algorithms. Our results reveal that object-based image analysis outperforms pixel-based methods, with overall accuracies (OAs) of 80.29–87.75%; RF has the highest OA of 87.75% (Kappa = 0.864), followed consecutively by CNN (OA = 83.31%, Kappa = 0.829) and SVM (OA = 80.29%, Kappa = 0.813). Phragmites australis dominates the plant community (55.9%) at the study area, followed by Typha orientalis (16.2%), Suaeda glauca (16.2%), and Scirpus triqueter (4.6%). The results highlight the importance of spectral transformation features in red-edge regions. The mapping results will help establish basic information for subsequent studies involving habitat suitability assessment at this study site.

Keywords

• Community classification
• hyperspectral remote sensing
• machine learning
• unmanned aerial vehicle (UAV)
• wetland