Remote sensing retrieval of inland river water quality based on BP neural network
ZHANG Hongjian1， WANG Bing1， ZHOU Jian1， YU Yong1， KE Shuai1， HUANGFU Kuan2
(1.Xinyang Water Conservancy Survey and Design Institute， Xinyang，Henan 464000， China；2.School of Water Conservancy Engineering， Zhengzhou University， Zhengzhou 450000， China)
Abstract：At present， a large number of pollutants have been generated in industrial production， livestock and poultry husbandry， pesticide abuse， and human life， which seriously pollute rivers， lakes， and other bodies of water. Efficient， comprehensive， continuous and accurate detection of water pollutant concentrations in rivers and lakes can provide a reliable basis for water quality assessment and water pollution prevention. Traditional water pollutant concentration detection methods are limited by time and space， and can only detect fixed monitoring points at a specific time. However， remote sensing technology is able to overcome the limitations of traditional monitoring methods in terms of time and space， and continuously perform contaminant concentration inversion on water in a large scale and with high accuracy， which provides new technical means for the detection of water pollutants. In this paper， the Huaihe River in Xinyang City and its tributaries such as the Huanghe River， the Shihe River， the Bailu River， and the Zhugan River are studied. The landsat 8 remote sensing images are used to invert the water quality indicators COD and NH3-N. Inversion models of water contaminant concentration is constructed based on Traditional Linear Regression and BP Neural Network. In order to obtain high-precision inversion results， the correlation between COD and NH3-N are analyzed by the different bands and combinations of the bands of Landsat 8 images. Finally， a highly correlated band or band combination was selected for inversion， and the inversion results are compared with the measured pollutant concentration data of Xinyang Municipal Bureau of Hydrology. The experimental results are shown as follows. 1) The correlation coefficient of NH3-N with the Band2 data of Landsat 8 image is the highest， reaching 0.745. 2) The correlation between COD and Landsat 8 image in each single band is not very high， but after the combination of Band4 and Band6， the correlation coefficient reaches 0.852， and the correlation between total phosphorus and Band4 is as high as 0.873. 3) The inversion of COD and NH3-N based on BP neural network model is in good agreement with the measured data， and the relative errors were 14.35% and 29.30%， respectively. The research results demonstrate that the use of remote sensing images and the BP neural network model is able to accurately detect COD and NH3-N， which are the major pollution indicators of rivers and lakes.
2022, Vol. 56 
