Inversion And Prediction Of Dissolved Oxygen In Taihu Lake Based On Space Ground Monitoring Data

Improving water quality monitoring capabilities and timely understanding of changes in the water environment are of great significance for lake water environment protection and pollution prevention.With the development of artificial intelligence and remote sensing technology,the use of remote sensing satellite data and ground measurement data to invert the water quality of inland lakes has made up for the limitations of traditional water quality monitoring such as limited areas and poor timeliness.Benefiting from the rapid progress of deep learning,using long-term measured data to more accurately predict water quality can help predict potential pollution events and timely identify water quality issues within the region.This thesis provides an important basis for managing and maintaining the water quality status of inland lakes,improving the accuracy of lake water environmental pollution monitoring and warning,and strengthening the level of water environment governance.This thesis uses the remote sensing image of Himawari-8(H8)and the measured data of Dissolved Oxygen(DO)in the"Xuhuxin"section of the Taihu Lake to achieve accurate inversion and prediction of DO concentration in the"Xuhuxin"section.The specific contents include preprocessing the data used,constructing a Dissolved Oxygen Inversion Model based on Time Series Decomposition Algorithm and Multimodal Deep Neural Network(DOI-TSD-MDNN),and constructing a Dissolved Oxygen Prediction Model Based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise and Temporal Convolutional Network(DOP-CEEMDAN-TCN).In terms of data pre-processing,this thesis uses meteorological data after synchronization time and K-Nearest Neighbor(KNN)algorithm to fill the missing values in DO data.In addition,this thesis preprocessed the H8 remote sensing images and extracted spectral index information significantly related to water DO concentration.Various remote sensing features and processed DO measured data formed an inversion dataset.The preprocessing results indicate that the DO concentration in the"Xuhuxin"section shows a trend of decreasing in spring and increasing in autumn,with uneven distribution at the seasonal level.The DO data filled by the KNN model tends to be more stable,providing a reliable data foundation for subsequent inversion and prediction model construction.To achieve DO concentration inversion,this thesis proposes a DOI-TSD-MDNN model.The Seasonal-Trend Decomposition Procedure Based on Locally Weighted Regression(STL)is used to decompose the DO measured data into periodic components,trend components,and residual components.The H8 remote sensing data is divided into three modes based on the central wavelength and spectral index.The experimental results show that the performance of the inversion model after time series decomposition is better than that before decomposition.The Adjusted Coefficient of Determination(adj＿R~2)of DOI-TSD-MDNN reaches 0.89,and the Root Mean Square Error(RMSE)is only 0.46 mg/L,which is superior to other comparison models.The density map of the comparison experiment shows that DOI-TSD-MDNN has a high degree of fitting with the measured data and is superior to other comparison models,Helps to improve the accuracy of DO inversion.To achieve DO concentration prediction,this thesis proposes a DOP-CEEMDAN-TCN model.Based on hourly measured DO data,the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)algorithm is used to decompose historical DO data,resulting in 14 Intrinsic Mode Functions(IMF)and 1 residual component.Then K-means clustering algorithm is used to reconstruct the decomposed components,and the distance calculated by Dynamic Time Warping(DTW)algorithm is used as the distance measurement of clustering.Finally,the reconstructed different components are used as inputs to the Temporal Convolutional Network(TCN)for prediction,and all prediction results are reconstructed to obtain the final predicted value.The results show that the CEEMDAN algorithm performs slightly better than other decomposition algorithms of the same type.Compared with other time series prediction models,the DOP-CEEMDAN-TCN proposed in this thesis performs the best when the number of clustering clusters is 8,adj＿R~2reaches 0.974,while RMSE is only 0.203 mg/L,which can improve the accuracy of DO prediction.