Live Oral Presentation in person as part of National Virtual Conference AFSS Conference 2020

Estimating the cover of Phragmites australis and other wetland features following environmental water using unmanned aerial vehicles and neural networks (#6)

William Higgisson 1 , Adrian Cobb 2 , Alica Tschierschke 1 , Fiona Dyer 1
  1. Univeristy of Canberra, Belconnen, Australian Capital Territory, Australia
  2. Private, Company, Canberra

Phragmites australis (common reed) is a widely distributed aquatic perennial grass of ecological and economic value. Declines and dieback in Phragmites australis have been observed in Europe and Australia, while the species has experienced range expansions in North America. The economic and ecological implications of both declines and range expansions has prompted a growing need for regular and accurate data collection to inform management and implement adaptive management strategies effectively.

Understanding the spatial arrangement of plants such as Phragmites australis and other wetland features in response to management actions (such as environmental flows) is an important part of managing freshwater systems such as wetlands. The use of unmanned aerial vehicles (UAV) provide high resolution and detailed imagery. Computational deep learning techniques are transforming the way in which these remotely sensed imagery and data can be used and are having an increasing role in remote sensing.

This study describes a novel image analysis technique using UAV and machine learning known as Convolutional Neural Networks for mapping the cover and extent of features of an ephemeral Phragmites australis reedbed in the Great Cumbung Swamp, on the Lachlan River, NSW. We trained our model to recognise five wetland features (Phragmites australis, bareground, leaf-litter, water and other vegetation) using point cloud imagery of five 50 X 50 m plots. The model was then validated using imagery which was previously unseen. The validation process demonstrated an overall high accuracy (0.90) and high precision (0.92) and could correctly identify Phagmites australis to 98% accuracy.

Using the model, we compared nine sites across three environmental watering frequencies in late January 2020. Sites which received two environmental waters in the past 12 months had the greatest cover of Phragmites australis, followed by sites which received one. Sites which had not received environmental water in the past 12 months had significantly lower cover of Phragmites australis.