Automatic collection and additional learning of teacher data in water-level measurement using neural network

Abstract

As our previous work, we have reported a method of measuring the river water level from camera image by detecting water border positions based on the deep learning technology, targeting non-installation sites of water gauges.

For this time, we have considered an additional method which automatically collects training data from images taken by cameras installed by river sides and lets neural networks learn the classification between water images and non-water images. We also have estimated the effectiveness of our new method using actual river image data set. As the results, the true positive ratio of classification reached from 6 to 37 points improvement.