Water Depth Inversion In Shallow Water With WorldView-2 Stereo Image Pair

With the development of high-resolution satellite remote sensing technology,using high-resolution remote sensing images to obtain shallow seawater bathymetry has become a current research focus.In order to study the shallow sea sounding capability of high-resolution remote sensing images,this paper uses WorldView-2 multi-spectral stereo pairs and in-situ bathymetry data to the bathymetry around North Island.Main contents include:(1)Researched the geometric conformation model of World View-2 stereo pair.The geometric relationship of camera site,image point,and underwater target point from single-media photogrammetry are analyzed.Based on the collinear equation and Snell’s Law the basic principles of dual-media photogrammetry is studied.The rational function model is introduced in detail.The RPC parameters and rational function model are used to obtain the three-dimensional information of the underwater observation point.(2)The refraction correction model is studied.The geometric relationship between the homonymous light and the underwater target point is studied in the situation where the underwater homonymous light of the stereoscopic images does not intersect.a refraction correction model is constructed according to the incident angle and the refraction angle,which calculates seawater depth from high-resolution satellite images using photogrammetry directly.Also,the error source of this method is analyzed systematically.(3)For dual-media photogrammetry,data processing steps mainly includes solar flare removal,water-land separation,epipolar image production,image dense matching and other operations.The clarity of underwater feature points is improved and the False matching rate of the homonymous points is reduced to overcome the shortage of SIFT algorithm that can only extract feature points relatively rare.The experimental results show that the dual-media stereophotometry sounding accuracy is low in the extremely shallow area of 0-5m,the MAE is 1.405m,and the R2 is only 0.22,but the MAE reaches 1.270m in the whole area,The R2 reaches 0.870,the overall effect is better.(4)We studied single-band linear regression model,dual-band ratio linear regression model,multi-band combined linear regression model,BP neural network water depth inversion model.The accuracy of these four water depth inversion models is analyzed using high-resolution remote sensing images and in-situ data.The result shows that the multi-band combined linear regression model and the BP neural network model has the highiest accuracy.The MAE and R2 reach 1.140m and 0.907.the accuracy of dual-media photogrammetry is higher than the three linear regression modelsIts’ accuracy,the accuracy as similar as the BP model,The single-band and double-band ratio linear regression inversion model has poor accuracy.