Client: Department for Environment and Water, South Australia
Lake Hawdon North is a seasonally inundated wetland of 2475 hectares located approximately 15 km east of the coastal town of Robe in the south east of South Australia. Historically the area has been used for sheep grazing and mining dolomitic lime but now the South Australian Department for Environment and Water’s Healthy Coorong, Healthy Basin program, in partnership with the Limestone Coast Landscape Board, is investigating the potential to extend the period of Lake Hawdon North inundation, providing shorebird habitat for the entire period that migrating shorebirds are present in the region..
Gahnia filum sedgeland with shallow water mudflat (ideal shorebird habitat) in the background (Tania Rajic and Andrew Rettig in photo). Photo: Claire Harding.
To assist with the restoration investigations, Lynker Analytics developed an Artificial Intelligence (AI) based system to map in high detail land cover and plant communities within the Lake from high resolution natural colour imagery acquired in 2021 by Spatial Solutions.
Firstly, the image data was reduced into processing units (512x512 pixels) and 11 training classes were identified. These included a range of woody plants, shrublands, sedges, open water, mud flats and grass/pasture. A training dataset which captured the full diversity of the area and the radiometric properties of the imagery was then created using experts from the department as well as wetland ecology experts working on the restoration.
This training data was then used to model this wetland system, leveraging existing convolutional neural network classification architectures used by Lynker Analytics for land cover mapping. The ML model generated a polygon feature class depicting the extent and coverage of the eleven land cover categories including the plant communities.
To generate a final land cover map six inference cycles were run using Python and Tensorflow/Keras. The ML model generated detailed predictions with an output resolution that matched the input imagery. Output data was then vectorised and post-processed in Esri ArcGIS including filtering and topological checks. Finally, a human QA exercise was run to ensure all classes were representative and spatially correct. Finally the models were validated using a separate hold out set of ground truth data.
Orthophoto (top) and Land cover map (bottom), class schema with area shown below.
The distribution of land cover by area is shown below. One of the primary goals of the project was to map the distribution of Melaleuca halmaturorum (Swamp Paperbark) which is a native wetland species that has over time moved from the edges into the wetland basin, displacing other vegetation, since the water regime became drier following regional drainage in the 1950s.
Overall, the model had a mean F1-score of 0.93. Aquatic herbland dominaited by rushes and milfoil (Baumea arthrophylla (now Machaerina arthrophylla), Myriophyllum sp.), exotic pasture grass & native grass (Schoenus nitens) dry herbland validation points produced the highest F1-scores (0.98), while the deep water samples had the lowest F1-score of 0.81.
It was observed that additional classes such as algal blooms could also be mapped in future surveys. One observed limitation of the model was the overflow of overpredicted classes, Juncus krausii (Sea Rush) for example, was sometimes inferred at the edge of Gahnia filum (Cutting Grass), presumably due to their similar structure and habit. Apparent noise visible to the training models may also indicate under-story vegetation communities, which may warrant distinction in future studies.
Area per land cover class in study area
Overall, the machine learning based process enabled a mostly automated classification of land cover and areas of vegetation which will help with vegetation management and infrastructure decisions to regulate water levels.
Deliverables:
Natural Color nadir imagery, 0.10m GSD.
11 class land cover data set (polygon feature class)
Data Quality and Accuracy Report.
Melaleuca halmaturorum juvenile shrubland encroaching on areas of aquatic herbland. Photo: Claire Harding
Acknowledgements:
Lynker Analytics would like to acknowledge the contributions of Claire Harding, Ross Anderson and Andrew Rettig of the Department for Environment and Water, Tania Rajic of the Limestone Coast Landscape Board, Dr Kerri Muller of Kerri Muller NRM, Ben Taylor of Nature Glenelg Trust, and Brian Harvey of Spatial Solutions in the acquisition and quality assurance of the label class data used in the project.
The Healthy Coorong, Healthy Basin Program is jointly funded by the Australian Government and the Government of South Australia.