Detection of tangled lignum (Duma florulenta) from UAVs using imagery and machine learning

This project sought to develop and test methods to estimate the cover and condition of tangled lignum (Duma florulenta) using high-resolution imagery captured with an unpiloted aerial vehicle (UAV). We collected standard RGB and LiDAR imagery using UAVs and developed and tested models and methods to analyse and process these data using machine learning convolutional neural networks (CNNs) and digital surface models (DSMs) at 18 (50 X 50m) plots in the Hattah Lakes, Wallpolla Island, and Lindsay Island in North-West Victoria. We also collected field-based data using the Lignum Condition Index (LCI) at each of the 18 sites to compare these remotely sensed approaches to the currently adopted field-based method. Four of the 18 sites had been managed with environmental water over the five years prior to the surveys in March 2023 and all 18 sites were inundated from a wide-scale natural flooding event in late 2022 and early 2023. This provided an opportunity to use these data to describe the cover and condition of tangled lignum in relation to recent and historic flooding history and environmental watering.

We developed a machine learning model using RGB imagery to classify tangled lignum within two condition classes (high quality and low quality) and other wetland feature classes (such as water, bareground and trees). The model had an overall accuracy of 0.868 and recognised the lignum high quality class with an accuracy of 90% and the lignum low quality class with an accuracy of 74%. Using the data obtained with the machine learning model, the sites managed with environmental water had the greatest cover of lignum high quality of the 18 sites. In contrast, sites which prior to the recent flooding event had not been inundated for 15 years had very low (< 10%) cover of tangled lignum.

The LiDAR imagery was processed into DSMs and used to estimate the percent cover of tangled lignum in different height classes within a plot. This method was found to be effective in estimating total cover and percent cover of tangled lignum within different height classes, providing estimates of the proportion of plants in different size classes. However, this approach is limited to sites where tangled lignum is the dominant mid-story species.

The imagery obtained using UAVs and processed using machine learning provides an effective and accurate way to estimate the cover and condition of tangled lignum and other wetland attributes. This approach combines a measure of tangled lignum condition (high and low) with an estimate of cover at a site. The currently adopted field-based approach the LCI, provides only an estimate of condition and we believe incorporating percent cover (with a measure of condition) has provided a more representative and complete picture of tangled lignum at a site.

Further model refinement will continue to improve the model outputs and accuracy of the machine learning CNN model. It is also recommended that further work is undertaken on the LiDAR imagery and DSM approach to continue to improve and test this approach. We recommend all 18 sites are resurveyed in early 2024 to test how these approaches detect changes in tangled lignum condition over time at sites with varying watering regimes and environmental watering histories.