Fully-covered bathymetry of clear tufa lakes using UAV-acquired overlapping images and neural networks

•UAV-acquired overlapping images were applicable to obtaining water depth map of clear tufa lakes.•Single bathymetric methods caused depth anomalies in water areas with uniform texture or shadows.•Photogrammetric and spectral-based bathymetric methods were combined using neural network (NN).•NN-based bathymetric methods realized fully-covered and non-contact bathymetry of tufa lakes.•Shallow NN model achieved highest bathymetric accuracy while deep NN model added more details. Accurate and updated bathymetric data is of great significance for the management and protection of alpine tufa lakes. In recent years, unmanned aerial vehicle (UAV)-borne optical remote sensing has become a cost-effective technique for obtaining water depth of small and clear waters like tufa lakes. UAV-based bathymetry can be categorized into photogrammetric approach and spectrally derived approach. Photogrammetric bathymetry is contactless but invalid in water areas with uniform texture, while spectral-based bathymetry requires a large amount of in-situ depth measurements. In this paper, we combined the strengths of the two bathymetric methods to retrieve the depth of clear tufa lakes using neural networks. The surface elevation and orthoimage were first produced from UAV-acquired overlapping images, and then water color-depth tie points were sampled in the orthoimage and refraction-corrected bathymetric map. Next, the shallow and deep neural networks were separately used to train the regression models for predicting water depth. Lastly, the combined bathymetric methods were compared with the single ones in terms of effective spatial coverage and bathymetry accuracy. The results indicated that the combined methods were superior to single bathymetric methods in fully-covered bathymetry of clear tufa lakes. The shallow neural network-based model achieved the highest accuracy, with the coefficient of determination (R2) of 0.91 and the Root Mean Square Error (RMSE) of 1.12 m, whereas the deep neural network-based model increased the details of water depth distribution.