Bathymetry from multispectral aerial images via convolutional neural networks
Recently, optical approaches were applied more often to derive the depth of waterbodies. In shallow areas, the depth can be deduced mainly by modelling the signal attenuation in different bands. In this approach, it is examined how well a convolutional neural network (CNN) is able to estimate water depths from multispectral aerial images. To train on the actually observed slanted water distances, the net is trained with the original images rather than the orthophoto. The trained CNN is showing a stand-ard deviation of 3 to 4 decimetres. It is able to recognise trends for varying depths and ground covers. Problems mainly occurred when facing sunglint or shaded areas.
CNN – convolutional neural network | multispectral aerial images | orthophoto | LiDAR
