A RIVER MONITORING BASED ON CLASSIFICATION OF FLOW CONDITION IMAGES FOR TURBIDITY OF RIVER WATER BY DEEP LEARNING: INTENDED TO THE SAWATARI RIVER IN MITO, IBARAKI PREFECTURE

Abstract

 It is considered that the change of turbidity of river water is an element capable of catching flood risk such as the change of flow rate and water level. In this study, image classification by deep learning and river monitoring technique by cheap monitoring camera for turbidity of river water obtained from flow condition images were proposed. The image classification model by Deep Learning which set the classification item in proportion to turbidity condition of river water was made using flow condition images of the Sawatari River. The verification case by the two image classification models by DNN in which the class setting condition of the training data is different was set using the image in the summer in which the vegetation occupation area is big for the model preparation. The model obtained the best accuracy was fixed, and images of rainfall in summer to winter with different photographing dates were classified. As a result of the examination, the high classification accuracy was obtained for the image in the different time from the training data, and the possibility of developing as a technique which handled the image in the multitemporal was confirmed.