Machine learning approaches for delineation of bankfull stream channel dimensions from LiDAR data
Catchment urbanisation has profound effects on the physical form and functioning of stream channels, with far reaching economic, ecological and social implications for our cities and suburbs. In this overarching project, being undertaken in partnership with Melbourne Water (mwrpp.org), the aim is to develop statistical models for the expected extent and severity of stream channel change relative to the level of catchment urbanization across the Greater Melbourne region. The outcomes will assist stream managers to plan for geomorphic change, and to develop new ways of protecting streams from catchment urbanization.
A key input to this model is stream channel dimension data, which is challenging to gather across broad scales despite good coverage of LiDAR topographic data. Existing datasets are incomplete, low-quality and non-reproducible, and current methods to extend and improve them are labour-intensive, severely limiting our modelling. The team proposes a collaboration with MDAP to explore machine learning methods to identify bankfull channel extents from LiDAR digital elevation models.
If a machine-learning method can perform comparably to current methods, the potential research impact is considerable, both on this project and on river research globally. This collaboration may to lead to improved coverage and quality of channel dimension data, and hence improved models of pressure-response relationships in stream geomorphology (both here in Melbourne and worldwide). Ultimately, these advances are expected to lead to better stream protection, management and planning.
Garber J, Thompson K M, Burns M J, Kunapo J, Zhang G Z, Russell K (2024) Artificial Intelligence and Objective-Function Methods Can Identify Bankfull River Channel Extents. Water Resource Research 60(1). doi: 10.1029/2023WR035269
Figure caption: Bar plot of Dice coefficient for each imageset, aggregated for the entire test set (a.), and aggregated by reach (b). Note that the imageset is delineated by symbols denoting the three layers as defined in Table 3. (a) shows the Dice coefficients aggregated for the entire image, and aggregated within only a 50m buffer of the river centerline. (b) shows Dice coefficients aggregated within a 50 m buffer of the river centerline for each reach's test set. Reach properties are given in Table 2. Reaches 3 and 4 which are the furthest downstream are predicted with highest accuracy, and image-normalized elevations with a minimum relief of 2 m (DZL-2_0_0) yield the best results overall.