Drone-based aerial surveillance for detecting water hyacinth (Pontederia crassipes) using deep learning models

Riner AE, Glueckert JS, Leary J, Sperry BP, Macdonald B, Elrahman A-A. 2026. Drone-based aerial surveillance for detecting water hyacinth (Pontederia crassipes) using deep learning models. Lake Reserv Manage. XXX–XXX.Water hyacinth is a rapidly growing, floating aquatic invasive plant that threatens biodiversity, impedes navigation, and increases flood risk across the southeastern United States. Traditional water hyacinth management involves frequent monitoring and herbicide application from airboats; however, escaped plants lead to reinfestation, requiring continuous efforts. Surveys using unoccupied aerial systems (UAS) can potentially provide increased efficiency for large-scale monitoring, and deep learning can potentially enhance plant detection in aerial imagery. In this study, we evaluated 3 deep learning models (UPerNet, Unet, and DeepLabv3) for water hyacinth detection using red, green, and blue (RGB) imagery. To assess model transferability, we evaluated trained models under 3 conditions: (1) on datasets from the same site and time period as the training images, (2) on datasets from the same site but a different season than the training images, and (3) on datasets from a different site but the same time period as the training images. All models exhibited higher recall than precision, indicating strong plant detection capabilities, but frequently misclassified other species as water hyacinth. DeepLabv3 outperformed other models on images similar to the training data (69% accuracy), Unet outperformed other models on in a different lake (62%) and in a different season (57%) than the training set. UPerNet exhibited the lowest accuracy across all conditions. These findings highlight the potential of UAS and deep learning for water hyacinth surveillance, offering a scalable approach to invasive species management.