Floodplains are one of the most productive and threatened ecosystems on earth. All the more, they play key roles in providing habitat, ecological productivity and maintaining biodiversity. Drone-based monitoring analysed with deep-machine learning provides an accurate, repeatable and efficient approach to evaluate vegetation changes in response to water management. We provide a spatial framework to explore the relationship between environmental water and vegetation – with a focus on the floodplain shrub Duma florulenta (tangled lignum) by integrating drone imagery, Convolutional Neural Networks (CNNs). We conducted drone surveys at 18 sites across three floodplains in spring 2023 and 2024, which are dominated by D. florulenta in the Mallee CMA (Catchment management Authority) in north-western Victoria, Australia. The imagery collected, were used as the inputs and outputs of the framework to develp a drone-based method to estimate a combined measure of percent cover and condition of D. florulenta by using two condition categories (high-quality: green stems and leafs, and low-quality: brown stems and leafless). Five different CNN methods were assessed and compared regarding accuracy and efficency (running time and computational resources).
The best performing methode are our study CNN model, it had an overall accuracy of 91.1% and a Kappa coefficient of 89.9%, with an accuracy for low-quality lignum of 94.5% and for high-quality lignum of 96.4%. The total percent cover of D. florulenta and cover of high-quality lignum was related to flooding history at a site and the use of environmental water with the more frequently flooded sites having greater percent cover.
The results provide scientific information for making adaptation strategies in water management. The study offers an accurate and repeatable method to estimate the influence of floodplain inundation and environmental water on D. florulenta and other wetland vegetation and attributes., which can be extended to other environmental monitoring programs.