Monitoring Inland Water Quality Using Remote Sensing And Machine Learning

Water is the foundation of all living things on the earth and the source of the human civilization development.Water quality monitoring is crucial for timely detection,tracing and controlling of water pollution,so it is key to water environment management and water resource protection.Traditional water quality monitoring methods are time-and labor-consuming,and the water quality of point scale is difficult to reflect the water quality of the whole water surface.Therefore,it is difficult to help the water environmental department to effectively and accurately grasp the temporal and spatial changes of water quality.Remote sensing provides large-area,periodic,dynamic,efficient and low consumption water quality monitoring.Since the 1980s,remote sensing has been progressively used in water environment management.After half a century of development of water color remote sensing,scholars have proposed mature atmospheric correction technology,constructed stable and accurate water quality estimation methods,and realized the operational of remote sensing monitoring of open ocean water quality.Compared with ocean waters,inland waters such as lakes and reservoirs are usually small in scale,severely affected by climate and human activities,and the water quality has strong spatiotemporal heterogeneity.As a result,there is still a lack of mature algorithms for inland water quality retrieval.Meanwhile,the existing research mainly focuses on optically active constituents（OACs）,and the research on non-optically active constituents（non-OACs）is limited.Machine learning has been proved to have the ability to capture complex nonlinear relationships,so it is expected to be used to accurately capture the potential relationship between OACs,non-OACs and image spectra,so as to realize the remote sensing estimation of non-OACs.However,the machine learning models developed based on regional data are usually only suitable for local areas,and are difficult to be extended to other water bodies.In addition,machine learning models lack physical interpretability,and the existing research is limited in the interpretation of model estimates.Moreover,due to the limited research on remote sensing retrieval of non-OACs,there is also few studies on the spatiotemporal changes of many important non-OACs and their response mechanism to human activities and climate change.Based on the measured water quality data of the Lake Bolong in Tianjin City,Lake Taihu in Jiangsu Province,Qinglinjing reservoir in Guangdong Province,China,Lake Huron and Lake Simcoe in Canada and the synchronous MODIS,Landsat series,and Sentinel-2 images,this study constructs machine learning models including random forest,support vector regression,back propagation neural network,multimodal deep learning and enhanced multimodal depth learning,which are used to remotely estimate seven water quality parameters,i.e.,Chla,SDD,TP,TN,DO,DOC,and COD.We split the data into training set,test set and independent set to train and rigorously verify the model performances,and compare the performances of the proposed models with the most commonly used empirical algorithms and other machine learning and deep learning models.Meanwhile,with the help of the latest machine learning interpretability algorithms,we quantitatively calculate the contribution of the input features of each deep learning model based on artificial neural network to the model estimates,effectively separate the interaction between the estimated water quality parameters,and improve the physical interpretability of the models.In addition,this study uses the models constructed and strictly evaluated to reconstruct and reveal the spatiotemporal variations of different water quality parameters at multiple spatiotemporal scales,and quantify and explore the impact mechanism of human activities and climate change on different water quality parameters.The results show that,（1）compared with Landsat series,MODIS,and other remote sensing image data sources,the high spatiotemporal resolution of the Sentinel-2 images makes it possible to retrieve the water quality parameters of small water bodies.Sentinel-2 image bands with the highest contribution to TP,TN,and COD estimation are B₃,B₄,and B₅.The most suitable band combinations for TP,TN,and COD estimation are“B₃+B₄+B₅+B₆+B₇+B₈”,“B₃+B₄+B₅+B₆+B₇+B₈”,and“B₂+B₃+B₅+B₆+B₇+B₈”,respectively.The optimized machine learning models and band selection significantly improve the estimation accuracy of non-OACs,especially TP and TN.（2）The SVR model optimized by the kernel-trick has good robustness in the dissolved oxygen（DO）estimation at multi spatiotemporal scales using Landsat and MODIS data.The average R²,RMSPE,and MAPE of the model constructed from Landsat and MODIS image data were 0.91%,2.65%,and 4.21%,respectively.Inputting water surface temperature and sampling point coordinates into the model significantly improves the generalization performance of the model.The monthly DO estimated by Landsat and MODIS data is highly consistent（average R²=0.88）.From1984 to 2019,the surface DO of Lake Huron decreased by 6.56%.Since 2000,the DO in Lake Huron has shown an obvious downward trend in the same month of different years（p<0.05）.AT,SRAD,and PRCP are the main climate factors affecting the long-term change trend of DO in Lake Huron.（3）The MDL models developed and verified using the atmospherically corrected R_rs data of Landsat raw images and the synchronous water quality measurements of Lake Simcoe can accurately estimate Chla（MAE=32.57%,Bias=10.61%）,TP（MAE=42.58%,Bias=-2.82%）,and TN（MAE=35.05%,bias=13.66%）,and outperform the other machine learning and deep learning models and empirical algorithms.Although TP and TN belong to non-OACs,MDL models effectively capture the potential correlation among them,OACs,and R_rs,and promote the direct retrieval of non-OACs.The MDL model can reconstruct the spatial distribution of Chla,TP,and TN in Simcoe Lake since 1984,help to effectively monitor the water quality of Simcoe Lake by using Landsat R_rs data,and provide a basis for quantitative analysis of the water quality improvement effect of LSEMS implemented in 1990 and LSPP implemented in 2009.In addition,quantifying the contribution of each feature to the model contributes to the effective splitting of Chla,TP,and TN estimates,and improves the model interpretability.（4）The EMDL models developed and verified using the R_rs data of HLS images,the synchronous water quality measurements of LSLMP,WST,and climate data perform well in the estimation of Chla,TP,TN,SDD,DOC,and DO of Lake Simcoe with Slope being close to 1（0.84-0.95）,MAE≤20.17%,and Bias≤14.68%.Adding climate data to the model input features enhances the performance of the model.EMDL model has the potential to reconstruct the spatial distribution and time series dynamics of water quality,and can effectively capture the gradient of spatial variation of water quality.The fusion of Landsat 8 and Sentinel-2 images improves the temporal resolution of meter-scale resolution satellite images,which is helpful for water quality monitoring in inland and coastal waters.The contribution of different input features to EMDL model is calculated using the SHAP library,and the possibility of splitting the interactions between the estimated water quality parameters is discussed,which improves the physical interpretability of the model.We reconstructed and analyzed the spatial distribution and time series variations of water quality parameters in Lake Simcoe from 2013 to 2019,and quantitatively analyzed the impact mechanism of 12potential natural and human activity factors on the water quality of Lake Simcoe.The results show that the water quality of Lake Simcoe and its two most concerned estuaries is mainly affected by human activities,such as urban development and agricultural production.Natural factors also have a great impact on the water quality in these areas.This study introduces machine learning into the water color remote sensing to realize the remote estimation of non-OACs.The model generalization is greatly improved by inputting climate variables and remote sensing indices into the machine learning models.Besides,by calculating the contribution of each input feature to the model estimation,this study effectively splits the interactions among the estimated water quality parameters,and improves the physical interpretability of the model.Finally,the historical spatial distribution of water quality is reconstructed using the combination and fusion of images at different spatiotemporal scales.By quantifying the correlation between multiple potential factors and water quality,the impact mechanism of human activities and climate change on long-term water quality change is revealed.This inland water quality monitoring framework based on machine learning and remote sensing provided in this study contributes to the efficient monitoring of water environment and water resources.In addition,the water quality monitoring framework proposed in this study not only contributes to the efficient monitoring of water quality,but also helps the integration of water quality online monitoring system,which facilitates the environmental protection.