Research On UAV-borne Hyperspectral Imagery For Retrieval Water Quality Parameters By Machine Learning Algorithms

With the extremely rapid economic growth and industrialization in China,it has led to the intensification of related sewage discharges,causing many secondary disaster problems related to water pollution such as health of residents.Inland water bodies are the main source of water for people’s daily life.Thus,monitoring the water quality is absolutely necessary things for the sustainable development of water bodies.Chl-a and SS are typical optical water quality parameters,which play an important role in water quality management.UAV-borne hyperspectral data is favored by researchers in various fields because of its high spectral resolution and high spatial resolution.Machine learning mod-els can quickly obtain the required information.The algorithmic structure of different machine learning models varies,and different extrinsic input data also have a large impact on the pre-dictive performance of the model.Therefore,in this study,we take Beigong Reservoir as the study area,combine nine machine learning models to quantitatively retrieval the water quality parameter concentrations by UAV-borne hyperspectral remote sensing imagery in Beigong res-ervoir,and obtain the spatial distribution map of water quality parameter concentrations in the study area.The main research conclusions are as follows:(1)The processing details of UAV-borne hyperspectral imagery and measured spectral data are described.In-situ radiometric calibration is accomplished by construct the linear rela-tionship between the pixel spectral curve of the sampling point in the UAV imagery and the measured spectral data.(2)Perform spectral ratio processing on the preprocessed hyperspectral remote sensing reflectance,which can remove the external environment effects.The prediction performance of the machine learning models by the processed spectral ratio characteristic variables and meas-ured water quality data were compared and analyzed.The traditional machine learning model has poor prediction accuracy,while the results of the ensemble machine learning model are better than the traditional machine learning model,and the overall performance is good.Com-pared to other models,the highest prediction accuracy in the inversion experiment is the CBR model.(3)Using water quality parameters Chl-a and SS,the performance of machine learning models in different water quality parameters was analyzed.The prediction accuracy of the CBR model in the Chl-a and SS experiments were(R~2=0.96,MAE=0.47,RMSE=0.53mg/m~3)and(R~2=0.94,MAE=1.11,RMSE=1.20mg/L),respectively.CBR model have highest precision performance compared to others.(4)Using the trained CBR model combined with the pre-processed UAV-borne hyperspec-tral imagery to obtain the concentration spatial distribution map of Chl-a and SS.By comparing with the investigated water quality results,it demonstrates that the inversion results generated by the CBR model combined with the UAV-borne hyperspectral images have high accuracy,which can be used for water quality monitoring.