Phenological and Species-Level Classification of Aquatic Invasive Plants Using UAV Multispectral Imagery and Machine Learning

Monitoring aquatic invasive plant species (AIPs) and their phenological stages remains a challenge in complex freshwater environments. This study evaluates the potential of UAV-based multispectral imagery and machine learning for classifying six vegetation classes in a tropical urban reservoir composed of three phenological stages of Eichhornia crassipes and Brachiaria subquadripara, Pistia stratiotes, and Typha domingensis species distribution. UAV flights were conducted on three dates using the MicaSense RedEdge-Dual sensor. A two-step principal component analysis (PCA) was used to select spectral bands and derive Normalized Difference and Ratio indices, aiming to reduce redundancy and assess their usefulness in classification. Three classifiers—Linear Discriminant Analysis (LDA), Random Forest (RF), and Support Vector Machine with RBF kernel (SVM-RBF)— were tested using 5-fold cross-validation. RF and SVM-RBF achieved the highest accuracies, ranging from 0.71 to 0.84, while LDA presented the lowest accuracy, between 0.63 and 0.82. Including spectral indices yielded only marginal improvements and did not consistently enhance classification performance, particularly when using more robust algorithms like RF and SVM-RBF, indicating that the ten original spectral bands are adequate to capture the key spectral distinctions in most cases. Classification performance was more consistent for Brachiaria subquadripara and Pistia stratiotes, while considerable confusion was observed between Typha domingensis and the phenological stages of Eichhornia crassipes, likely due to spectral similarity. Overall, model selection had a higher influence on performance than feature augmentation. Future studies should explore spatial-textural features and sensor fusion to improve the generalization of AIP monitoring systems.