Study On Inversion Method Of Chl-a Concentration In Wuliangsu Lake Based On Machine Learning And Remote Sensing Images

Water is the source of life,and monitoring water quality has always been a concern for countries around the world.The monitoring results of water quality are the main basis for pollution prevention and control,as well as the key to environmental governance.Chla concentration can determine the number of algae in water and indirectly reflect the health status of water quality.Inverting chl-a concentration using satellite multispectral remote sensing data is an effective approach for water quality monitoring.Due to the complexity and optical characteristics of water bodies,different inversion models for chl-a concentration based on remote sensing data need to be established for different regions.This thesis proposes a neural network-based chl-a concentration inversion method and a remote sensing image band selection method for Wuliangsu Lake Wetland,using Sentinel-2 multispectral remote sensing data.The aim is to achieve better chl-a concentration inversion results.The main research content is as follows:(1)Firstly,atmospheric correction,water extraction,and resampling were performed on Sentinel-2 multispectral remote sensing data,and then satellite multispectral remote sensing data was matched with on-site sampling data of chl-a concentration.On this basis,chl-a concentration inversion models based on BP neural network and support vector regression were established using full band remote sensing data.And in response to the periodic characteristics of chl-a concentration in Wuliangsu Lake,a chl-a concentration inversion model was established using months as input features,significantly improving the effectiveness of chl-a concentration inversion.(2)In view of the high correlation between bands of satellite multispectral remote sensing data,two inversion models of chl-a concentration based on time series neural network model,cyclic neural network and short-term memory network were established.When constructing these two models,a different method was used,which utilized the dependencies between bands by inputting each band of satellite multispectral remote sensing data into different time steps.Compared with the chl-a concentration inversion model based on BP neural network and support vector regression,the chl-a concentration inversion model based on short-term memory network performs well.(3)In order to select the bands of remote sensing data more reasonably,this thesis proposes a band selection method based on Bayesian optimization algorithm,which can also optimize the hyperparameter of machine learning.By combining Bayesian optimization with the above three neural networks and support vector regression respectively,and through comparative analysis,the BP neural network model after Bayesian optimization of band selection and adjustment of hyperparameter is finally obtained,which performs best in the inversion of chl-a concentration.