A Biological Sensor System Using Computer Vision for Water Quality Monitoring.

Fei Yuan,Yifan Huang,Xin Chen,En Cheng

DOI: https://doi.org/10.1109/ACCESS.2018.2876336

IF: 3.9

2018-01-01

IEEE Access

Abstract:Water pollution has seriously threatened our life, so an effective water quality monitoring mechanism is the most important part of water quality management. Most studies use biological monitoring methods to monitor water pollutants, such as pesticides, heavy metals, and organic pollutants. However, there are still many difficulties at present. Few methods consider the influence of illumination and complex background in the monitoring environment, and the characteristics parameters extracted in the systems are single. In addition, the results of using shallow neural networks for water quality classification are often not ideal. In order to solve the above problems, we design a water quality monitoring system combined with the computer image processing technology and use computer vision to analyze the fish behavior in real-time for monitoring the existence or not of water pollution. For the illumination problem, we use the no-reference quality assessment algorithm based on natural scene statistics for contrast distortion images to evaluate the video and configure the lighting conditions of the monitoring environment. White balance preprocessing is also performed to provide a great basis for moving target detection. Besides, we use background modeling to eliminate the influence of complex background on the moving target detection and the foreground is extracted using the saliency detection algorithm. In order to comprehensively analyze the influence of water quality on the fish behavior from the extracted foreground targets, multi-dimensional feature parameters are used to quantify the indicators, including movement velocity, rotation angle, spatial standard deviation, and body color which characterize the behavior changes of the fish. Finally, the classification model based on the long short-term memory neural network is used to classify the feature parameters data of the fish behavior in different water quality environments. In this paper, red zebra fish is used as the indicator organism and copper sulfate solution is used as the toxic pollutant to simulate the water pollution. Experiment results show that the classification accuracy rate of water quality using the proposed system can reach 100% at level 2 classification (93.33% at level 3 and 91% at level 4). Our system can achieve more accurate multi-level classification than the shallow neural network, such as RNN, and it is faster for real-time monitoring with a high reference for the water environment emergencies.