Based On The Quantitative Remote Sensing Monitoring Of The Weihe River Water Quality Study Of Integrated Learning

With the rapid development of industry and agriculture, environmental pollution is becoming increasingly serious, and especially, water pollution is one of the severe environmental problems the whole world is currently facing.The improved technology of space remote sensing makes it convenient and economical to monitor the water quality of surface water dynamically and continuously in a large scale.Previously,studies of monitoring water quality using remote sensing images mainly focused on large area waters such as bays, lakes, reservoirs et cetera, but now, with the improved resolution of remote sensing images, relatively small area waters are possible to be monitored.The Weihe River flows through Guanzhong, the most economically developed region of Shaanxi Province. Therefore, it has important realistic meanings to monitor its water quality parameters accurately by remote sensing technology. In the thesis, based on the traditional statistical theory and the machine learning approach,linear regression models and nonlinear regression models were separately built to predict the concentration of water parameters(chemical oxygen demand(COD),permanganate index(CODmn),dissolved oxygen(DO) and ammonia nitrogen (NH3-N)),using the historical measured data of the Weihe River water quality and the data obtained from SPOT-5 remote sensing images.The results of the regression inversion showed that there are complicated nonlinear relationships between the measured data and the image data. To improve the inversion precision of the Weihe River's water quality parameters, a new machine learning paradigm,ensemble learning,was introduced in this thesis, and on the basis of some existing ensemble methods, two new ensemble learning method:the weighted average ensemble of Artificial Neural Networks based on sample reconstructing and the selective ensemble of Support Vector Machines which adopted the double disturbance mechanism is proposed, which further improve the inversion precision of ensemble models.The study of the thesis includes the following aspects:(1)The overview of domestic and foreign remote sensing monitoring of water quality is summarized, and then the ensemble learning and its primary algorithms is introduced briefly.(2) Based on the normal distribution analysis of the measured data and the image data, the stepwise linear regression models are built, and outliers are detected pre and post the modeling.(3) The generalization ability of the Artificial Neural Networks always can be improved by the typical Bagging,the typical Boosting and the simple average ensemble based on sample reconstructing.To obtain stronger generalization ability,on the basis of the simple average ensemble based on sample reconstructing,the weighted average ensemble based on sample reconstructing is proposed by assigning weights to the member learners according to their respective prediction accuracy and difference measurement.(4) At least in the case of this study,Support Vector Machine,a kind of stable strong learner,its generalization ability is not improved by the typical Bagging and Boosting.The generalization ability of single SVM is strengthened significantly by ensembling some member SVMs with high prediction, which selected from member SVMs that are obtained by disturbancing the base learner's training samples and model parameters.(5)The generalization ability of the models mentioned above is compared,and the selective ensemble of Support Vector Machines which adopted the double disturbance mechanism is employed to regress the four water parameters of the Weihe River,the inversion results are coincident with the actual measured data,which suggesting that the ensemble methods are feasible and effective.