Bathymetry estimation of a shallow proglacial lake through UAV imagery and a geospatial regression method

Proglacial areas are one of the most rapidly changing ecosystems due to glacier and permafrost degradation. To better understand these environments and their dynamics, bathymetric mapping is a necessary step in hydraulic modelling. This is essential for assessing water quality, sediment and pollutant movement, and evaluating habitats. This thesis aims to evaluate the effectiveness 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 proglacial lake, of which water depth is less than about 1 m, from RGB and multispectral imagery. The case study is a system of proglacial channelized streams joining in a shallow lake originating from the melting of the Rutor alpine glacier, in Valle d’Aosta. Field experiments were carried out during summer 2021 and 2023 for GNSS positioning along different sections of the streams and simultaneously for acquiring photogrammetric data with digital numbers (DN) using an uncrewed aerial vehicle (UAV). The digital cameras mounted on the UAV were RGB and multispectral sensors respectively for 2021’s and 2023’s survey. Water depth along the surveyed sections was retrieved from measured points and the Digital Elevation Model (DEM) generated from the UAV imagery. After obtaining the orthomosaic of the area, band ratios of spectral bands were computed and, through the principal component analysis (PCA), were selected as an optimal input of the GWR model. For 2021’s model inputs, ln(DNB/DNR) and ln(DNG/DNB) were selected, while for 2023’s one, ln(DNG/DNNIR) and ln(DNR/DNG). Results showed that the GWR models based on a single band ratio input led to a discrepancy between estimation and observation, especially for 2021’s dataset. In contrast, considering both PCA-selected band ratios, the GWR models showed improved performances, as R2 increased from 0.47 to 0.77 for the 2021 dataset, and from 0.92 to 0.96 for the 2023 dataset. Moreover, the 2023 bathymetry reconstruction was more accurate than the 2021 reconstruction, likely due to the larger water depth dataset and the use of multispectral UAV acquisition. Multispectral data could more effectively model the effect of spatial heterogeneous bottom types, which arose from variable bottom types caused by submerged vegetation and sediment.