Inversion and Spatiotemporal Variation of Non-optically Active Water Quality Parameters in Daye Lake With Small Samples
Remote sensing technology provides great convenience for monitoring water quality of inland lakes.However,it is difficult to directly find the optical characteristics of certain optically inactive water quality parameters,such as Total Phosphorus(TP),Total Nitrogen(TN)and permanganate index(CODMn).In the complex biological optical environment of inland waters,it is usually hard to fit the measured water quality data with the remote sensing reflectance of such parameters using a simple regression model.Moreover,when the size of measured water quality data sample is small,traditional machine learning methods are prone to under fitting,resulting a low accuracy of water quality inversion results.To address these problems,this paper proposed a dot product attention model for estimating water quality parameters.The model was constructed using 58 measured water quality data and the corresponding Sentinel-2 remote sensing data,in Daye Lake.To verify the accuracy and priority of the proposed model,this paper compared the results with the water quality inversion results of statistical regression model and multi-layer perceptron model under the same measured water quality data sample and image data.Results showed that the accuracy of the inversion results based on the dot product attention model were the best,with the determination coefficients R2 of TP,TN and CODMn of 0.83,0.89 and 0.80,respectively.The proposed model was also employed to estimate TP,TN and CODMn in Daye Lake from 2018-2021.These results help to explore the spatial and temporal variation characteristics of the water quality parameter concentrations in Daye Lake in the past four years and assess the water pollution situation in Daye Lake.This research is expected to provide important data support for the water environment management of Daye Lake.
water quality retrievalsnon-optically active water quality parametersmachine learningattention mechanismSentinel-2
NETL
NSTL
万方数据
