Research On Hyperspectral Remote Sensing Based Water Depth Inversion Methods

Hyperspectral remote sensing based water depth estimation technology has the advantages of wide imaging coverage,high spectral resolution,and natural combination of spectral information and spatial information,and has gradually become one of the main estimation technology for shallow water depth research.Hyperspectral remote sensing based water depth inversion methods are mainly divided into data model and physical model methods.Data model methods include look-up table method,neural network method,etc.,and physical model methods include semi-analytical model method,etc.Among them,the neural network method and the semi-analytical model method have attracted much attention in recent years according to their high inversion accuracy.However,there are still some problems to be solved in the hyperspectral remote sensing based water depth inversion based on the above models,such as noise interference,ill-posed problem of water depth inversion,accuracy decline with increasing depth,and so on.In order to solve the above problems,this thesis mainly studies the water depth inversion method based on semi-analytical model and neural network model.The content is organized as follows:Aiming at the problems of ill-posedness,susceptibility to noise interference,and accuracy decrease with depth increase of the water depth inversion method based on the semi-analytical model,an improved water depth inversion method based on the semi-analytical model for hyperspectral remote sensing based water depth retrieval is proposed in this thesis.First of all,due to the ill-posed nature of the traditional semi-analytical model inversion of water depth,the maximum likelihood estimation method is first used to solve the problem to ensure the stability of the solution.Secondly,the measured water depth data is affected by many error sources,and the influence of many error sources is not considered when using the traditional semi-analytical model inversion,so it is necessary to include the error sources into the semi-analytical model to improve the spectral matching.Finally,in order to solve the problem that the accuracy of water depth inversion decreases with the increase of depth,a prior distribution about depth is included.And that means a regularization term is added to the loss function to improve the inversion accuracy.Verified by the simulated dataset and the real dataset of Saipan Island,the proposed method can improve the accuracy of water depth inversion to a certain extent.The advantages and disadvantages of the HOPE nonlinear optimization method,the probabilistic statistical method combined with environmental noise and bottom spectral variability(MILEBI)and the neural network model method in the hyperspectral remote sensing based water depth inversion model are compared.Considering that the HOPE nonlinear optimization method and the MILEBI method cannot solve the problem that the inversion accuracy decreases with the increase of water depth,and to illustrate the advantages of the regularization method in the semi-analytical model,the prior distribution of the water depth is included in these two methods.After testing with the simulated dataset and the real dataset of Saipan Island,the results show that:(1)In the water depth range of 0-20 m,the inclusion of the regularization term can solve the problem of the decrease in accuracy of water depth inversion with the increase of depth to a certain extent;(2)In the water depth range of 0-20 m,the inversion accuracy of the neural network model is slightly higher than that of the HOPE method and the MILEBI method;(3)The accuracy of the regularized MILEBI method is slightly higher than that of the regularized HOPE method.