CHRT is a density-based algorithm originally created to extract bathymetry from acoustic (multi-beam echo sounder or MBES) data. It has recently been adapted to airborne lidar data for shallow-water bathymetry. CHRT establishes a variably spaced grid of points across the data cloud area and develops one or more depth hypotheses for each grid point from its “neighbouring” pulse returns. If more than one hypothesis is identified, relatively unsophisticated disambiguation rules decide which hypothesis is most likely the true depth.
Because the disambiguation rules have been identified as a priority for future research, a machine learning approach has been explored. The result is a pulse return meta-data model that assigns a probability of being bathymetry – i.e., p(Bathy) – to each lidar pulse return. While it has been demonstrated that model accuracy is sufficient to improve CHRT, incorporation of model results into an operational workflow presents a number of challenges.
Principle among these is the seeming circularity that developing the ML model to improve the classification of lidar returns as Bathy/NotBathy requires an existing classification. This is addressed by adjusting CHRT hyperparameters to undertake an initial “highly certain” Bathy/NotBathy classification. From this an initial ML model is developed. Its p(Bathy) estimates are returned to CHRT which undertakes a second “less stringent” classification that incorporates the p(Bathy) estimates into the disambiguation rules. From this second CHRT classification, a second ML model is produced and used to re-estimate p(Bathy) which is again returned to CHRT for a third classification. This cycle continues until no additional Bathy points are identified.
It is expected that this workflow will further automate CHRT and lead to substantial cost-savings. In particular, given the accuracy of the ML models developed, the number of lidar pulse returns that need to be evaluated using time-consuming and tedious human examination should decrease substantially.
Kim Lowell is a Research Scientist at CCOM. His primary focus is the application of machine learning, deep learning, and other data analytics techniques to improve the accuracy of bathymetric charts. He has considerable experience in the analysis of geospatial information to address land management issues using GIS, spatial statistics, and optical, radar, and lidar imagery while also accounting for uncertainties inherent in those data.