Spatial distribution prediction model for lake water quality parameters from ultra-sparse sampling data with recent satellite data

Daily spatial distributions of water quality parameters (WQPs) are a crucial indicator in delicacy management of lake water environment in China. However, because of the high cost of construction and maintenance, there are a few automatic monitoring stations (AMSs) for continuous water sampling with only ultra-sparse spatial sampling data can be obtained. Meanwhile, although WQPs from satellite inversion are spatially continuous, they are missing on most days due to severe affected by weather and satellite revisit cycle, besides the high purchase cost. Thus, the model called MLP-GPR is formulated in this study to address the issue of predicting the spatial distribution of WQPs from ultra-sparse sampling data of AMSs combined with the spatial correlation information based on ultra-sparse satellite inversion data. This model includes the multilayer perceptron model (MLP) and Gaussian process regression model (GPR). MLP is trained to learn the interaction factors of WQPs between different locations through the satellite inversion data for predicting WQPs at more locations based on the data measured by AMSs. Then, GPR regresses the spatial distribution of WQPs at all locations on the entire lake from the predicted data by MLP. Based on the above model, we applied the satellite inversion, AMSs and artificial sampling data to predict the WQP spatial distribution of Shahu Lake in Wuhan, China. The experiment results show that the spatial resolution of WQP distribution predicted by MLP-GPR trained by satellite inversion data can reach 15 m by sampling data from artificial stations on the lake with 3.197 km2. Moreover, for each of these WQPs, compared with the satellite inversion data, the average MRE, R, and NSE of prediction results are below 3.95%, above 0.85, and above 0.69, respectively.