Wetlands are one of the most productive and threatened ecosystems on earth. The numerous ecosystem services supported by wetlands paired with their decline has prompted a need for regular and accurate data collection to inform management. As access to wetlands is often challenging or impossible, image collection via UAV (unmanned aerial vehicles) and processing them with deep learning techniques (convolutional neural networks, CNN) has become an important tool in ecological monitoring.
We conducted UAV surveys in a wetland dominated by Phragmites australis (common reed) in the Great Cumbung Swamp in Western NSW, Australia, seven times between October 2019 and November 2020. A total of nine (50x50m) plots were established along a hydrological gradient to determine the coverage of Phragmites australis. Data collection was undertaken to ensure capture of the temporal range of plant condition, all growth stages of Phragmites australis, different water levels as well as a range of other wetland attributes.
The UAV imagery were aligned into a single high-resolution image for each survey and plot. Using a subset of these imagery, which were sliced into 12 cm portions, we built, trained and tested a CNN model to identify and group wetland attributed Phragmites australis reeds, water, bareground, other vegetation and leaf litter. The model recognised and grouped these wetland features with a very high accuracy, e.g. Phragmites australis with a true positive rate of >98%.
Machine learning using remote sensing has been shown to be most effective when the target species has unique spectral and textural differentiation. This approach can be relevant to a range of wetland macrophytes with similar morphology and leaf structure (such as Schoenoplectus acutus and Typha spp).