Abstract
The Liaodong Shoal in the east of the Bohai Sea has obvious water depth variation. The clear shallow water area and deep turbid area coexist, which is characterized by complex submarine topography. The traditional semi-theoretical and semi-empirical models are often difficult to provide optimal inversion results. In this paper, based on the traditional principle of water depth inversion in shallow areas, a new framework is proposed in combination with the deep turbid sea area. This new framework extends the application of traditional optical water depth inversion methods, can meet the needs of the depth inversion work in the composite sea environment. Moreover, the gate recurrent unit (GRU) deep-learning model is introduced to approximate the unified inversion model by numerical calculation. In this paper, based on the above-mentioned inversion framework, the water depth inversion work is processed by using the wide range images of GF-1 satellite, then the relevant analysis and accuracy evaluation are carried out. The results show that: (1) for the overall water depth inversion, the determination coefficient R2 is higher than 0.9 and the MRE is lower than 20% are obtained, and the evaluation index shows that the GRU model can better retrieve the underwater topography of this region. (2) Compared with the traditional log-linear model, Stumpf model, and multi-layer feedforward neural network, the GRU model was significantly improved in various evaluation indices. (3) The model has the best inversion performance in the 24–32 m-depth section, with a MRE of about 4% and a MAE of about 1.42 m, which is more suitable for the inversion work in the comparative section area. (4) The inversion diagram indicates that this model can well reflect the regional seabed characteristics of multiple radial sand ridges, and the overall inversion result is excellent and practical.
Highlights
Water depth is an important element for marine scientific research, transportation and shipping, resource development, engineering construction, and environmental protection
root means square error (RMSE), mean absolute error (MAE), and mean relative error (MRE) are used to evaluate the error between the inversion results and the observed values
From the perspective of the overall water depth, the four accuracy evaluation methods given in Section 5.1 are applied to evaluate the inversion effect of the four water depth inversion models, root mean square error (RMSE), mean absolute error (MAE), mean relative error (MRE) and determination coefficient (R2) are obtained, respectively
Summary
Water depth is an important element for marine scientific research, transportation and shipping, resource development, engineering construction, and environmental protection. The above-mentioned water depth optical remote sensing inversion models express the relationship between the reflected light information of the seabed and the sea water depth and has been widely used in waterway engineering and reef detection [24,25,26,27,28] They are only applicable to shallow sea areas, and the inversion effect depends on the penetration ability of sunlight into the water body. This paper considers the use of the GRU deep learning method to regress this complex piecewise function uniformly, that is, to express the depth of shallow water and deep water at the same time This model can effectively learn the abundant spectral dimension sequence features of multispectral remote sensing images and establish the complex mapping relationship between the spectral features of remote sensing images and sea depth values. Comparative experiments are designed with the log-linear model, Stumpf model and other traditional methods, the experiments are carried out from the overall and segment aspects, and the correlation analysis and accuracy evaluation are followed
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
