Retrieving shallow stream bathymetry from UAV-assisted RGB imagery using a geospatial regression method

Abstract

Bathymetric mapping is a prerequisite procedure to conduct assessments of water quality, habitat and environmental flow for riverine ecosystems using hydraulic modelling. This study evaluates the capability of a geographically weighted regression (GWR) model, which can capture a spatially heterogeneous relationship between inputs and an output, to retrieve bathymetry of a shallow stream, of which water depth is less than about 1 m from simple RGB imagery. A field experiment was performed for measuring water depth and simultaneously for acquiring remotely-sensed data with RGB digital numbers (DN) using a digital camera mounted on an unmanned aerial vehicle (UAV). A 2D shallow water model, which was validated by comparison with the field-surveyed data, was used to simulate the water depth of unmeasured regions. Band ratios of ln(DNG/DNR) was selected as an optimal spectral input of bathymetric inversion models through the principal component analysis (PCA). Results showed that global inversion models based on multiple linear regression (MLR) and artificial neural network (ANN) resulted in large discrepancy between estimation and observation due to the spatially varying response of the PCA-selected band ratio to water depth over the experimental channel. In contrast, the GWR model successfully alleviated the biases of the conventional models as R2 increased to 0.85 from 0.60 by accurately modelling the effect of spatial heterogeneity, which arose from variable bottom types attributed to submerged vegetation, on the remote-sensing radiance-water depth relationship.