Study Of The Water Depth By Using The High Resolution Satellite Stereoscopic Detection In Shallow Water

Water depth is an important parameter of the ocean.It is a necessary reference index for various marine exploration and navigation.The mainstream method of ocean depth detection is the use of acoustic instruments such as single-beam,multi-beam,etc.However,in some areas,because of environmental or political reasons,it is impossible to carry out the depth measurement of ship borne water,thus the remote sensing detection of water depth is proposed.At present,the accuracy of the remote sensing water depth inversion is up to 80%,but there are still two problems:the single inversion model has a large error in 0-5m,there must have the known data of water depth for inversion.For the first question,in this paper,several different kernel functions are selected,and used a small machine learning method Support Vector Regression to construct the water depth inversion model,then selected the fuzzy membership degree as the standard to fuse the inversion results of multiple single models to improve the accuracy of the depth inversion,especially in the shallow water;For the second phenomenon,a method of water depth detection without prior data is proposed.This method uses the stereoscopic image pair to extract the underwater ground feature DEM visible in the image,selects the reference sea surface according to the extracted water line position,and then refracts the underwater part.Correct,obtain the depth value of the underwater point from the sea level,and obtain the elevation value of the underwater point on the depth datum after the tide correction.The experimental results are as follows:1.The multiple kernel SVR based on the fuzzy membership degree for remote sensing water depth fusion detection model has the mean absolute error of 0.99 m and the mean relative error of 8.2%in shallow waters of 25 m depth;its average relative error is compared with RBF,Sigmoid,polynomial and linear kernel functions respectively.The four single-core SVR models improved by 1.7%,4%,4.4%,and 4.8%,an increase of 5.5%over the classical log-linear model;For different depth sections,the mean relative error of the multi kernel SVR fusion detection model in the 3 water depth sections of 0 to 15 m is 0.7%to 54.9%higher than that of the four single-kernel SVR models,and 5 depths are 0 to 25 m.Intra-segment log-linear model increased by 1.1%to 20.4%;2.The optical depth seawater depth detection method without prior data was used to detect the depth of water in Panshiyu and Beidao.The accuracy before refraction correction was 42.94%and 31.13%respectively.After refraction correction using stereoscopic image pairs,The accuracy is 31.80%and 23.51%respectively,the accuracy have been improved by 11.14%and 7.7%;when the average incidence angles of stereo image pairs in Panshiyu Experimental Area are 26.60 and 0.7°,respectively,the accuracy of single-piece correction and double-chip correction is less than 0.01%in all verification points and different depths,that is,when stereoscopic image pairs When the incident angle of the scene image is extremely small,the single-piece correction model can replace the two-piece model to correct the depth of water;when the average incidence angles of the Beidao experimental area are 17.30 and 11.4°,respectively,the single-chip correction and double-chip correction are all the accuracy difference between verification points and different depths of water differs by about 0.2%.Comparing with the results in Panshiyu,when the difference in the average angle of incidence of stereo images to the two images is different,the accuracy of the water depth obtained by using the single-chip correction model is slightly lower than that of the two-piece correction model,and the accuracy is corrected with the differences in perspective between stereo pairs.