Remote Sensing Inversion Of Island Shallow Water Depth Based On Machine Learning And Empirical Methods

Knowledge of shallow bathymetry information is of great significance to the safe navigation of ships and marine mapping.Although,the conventional bathymetry methods can obtain shallow bathymetry data with high accuracy,it is difficult to achieve a wide range of shallow bathymetric distribution mapping due to the draught depth of ships,the difficulty of field measurement,high time and labor cost and low efficiency.The bathymetry inversion by satellite remote sensing can realize the dynamic observation of bathymetry on a large scale,which provides an effective supplement to the conventional shallow bathymetry methods.At present,the classic logarithmic ratio model(Stumpf empirical models)and statistical models based on machine learning methods were widely used in remote sensing bathymetry inversion research,which perform well in highly transparent water types and can achieve rapid estimation of water depth.However,previous studies are often limited to one type of model or a single study area,there is a lack of comparative and analytical studies on these two types of models,and the relationship and pattern between model inversion accuracy and water depth range,water substrate and the number of samples at different water depths are not clear,and various statistical models constructed based on machine learning algorithms also lack comparison and analysis.To address the above problems,this study takes the two typical optical shallow water islands of Ganquan Dao and Dazhou Dao as the study area,and constructing the water depth inversion model.We construct the Stumpf empirical model and the machine learning models based on the random forest,neural network and support vector machine algorithm respectively.At the same time,the accuracy indexes were selected,the accuracy performance and applicability of the above models were compared and analyzed on the basis of the analysis of model inversion accuracy,exploring the relationships and laws between the water depth ranges,water substrate types and numbers of water depth samples and the accuracy of various water depth inversion models.Providing a reference for Stumpf empirical model and various machine learning models to conduct remote sensing water depth inversion research in different water body types.Expanding the application scenarios of bathymetric inversion algorithms and to deeply understand the uncertainty of remote sensing bathymetric inversion.The main research results of this paper are as follows:(1)The bathymetric inversion experiments were conducted using Sentinel-2 satellite data.In the image pre-processing process,the obvious sun-glint phenomenon in the water part of the image,the algorithm is used to correct the sun-glint with the reflectivity information of nearinfrared band,which can better remove the sun-glint on the image and improve the accuracy of bathymetric inversion.(2)The Stumpf empirical model was constructed based on optimization algorithms,and various machine learning models were constructed using grid search.Comparing and analyzing the accuracy of the Stumpf empirical model and machine learning model in the Ganquan Dao research area.It is found that the Stumpf model can achieve good inversion accuracy at water depths lower than about 15 m.There is no significant difference between the Stumpf model and various machine learning models,and the root mean square error(RMSE)is 1.42 m in the water depth range from 0 to 5 m,1.27 m for the water depth range from 5 to 10 m,2.2 m for the water depth range from 10 to 15 m,and the RMSE of the model as a whole is 2.59 m.It indicates that the Stumpf model can be applied to water bodies with mainly sandy and coral substrates and high reflectivity at the bottom,and the model is less influenced by the number of bathymetric samples,and can construct a reasonably accurate bathymetric inversion model based on a small number of accurate and reliable bathymetric sample points to achieve rapid estimation of water depth.(3)Comparing and analyzing various water depth inversion models in the study area of Dazhou Dao.It indicates that the Stumpf model is limited by the substrate type of the water body,the accuracy of the model in Dazhou Dao is poor,and the prediction results of the model cannot reflect the real water depth.The Stumpf model is difficult to apply to the water body type with mainly bedrock substrate of low bottom reflectivity.The prediction results of the model constructed based on various machine learning algorithms can reflect the real water depth,the model accuracy is reasonable,and the applicability in water depth inversion is stronger,which can be applied to various water body types.(4)By comparing the accuracy of various models in the area of Ganquan Dao and Dazhou Dao,the bathymetric inversion models constructed based on various machine learning algorithms have the better accuracy performance,which are better than the Stumpf empirical model.The accuracy of the machine learning model is significantly better than the empirical model when the water depth is greater than 15 meters,and the applicable water depth range of the machine learning model is greater than that of the Stumpf empirical model.The overall accuracy results of the random forest model were optimal,with RMSE of 1.41 m and 2.73 m,and the RMSE of the BP neural network model were 1.46 m and 2.78 m,the RMSE of the support vector machine model were 1.50 m and 2.89 m,respectively.Among various types of machine learning algorithms,the random forest algorithm performs the best in bathymetric inversion and is more suitable for application in remote sensing bathymetry inversion.