Water Depth Inversion Method Of Optical Remote Sensing Using Multilevel Decision-making Scheme Based On Worldview-2 Image

At present,the international maritime situation is becoming extremely grim.It's for marine resources contention in the final analysis.As a sea power,how to defend our national maritime rights and interests,fully efficiently use marine resources,develop marine economy is an important issue today.Submarine terrain information is the basic information of ocean,which can provide essential support for marine transportation,engineering build,resources exploitation and marine disasters prevention and control,and provide geographic information guarantee for national defense construction and national security.Therefore,accurate ocean depth detection,as a direct acquisition method of seabed terrain information,is of great significance.Among various depth detection means,optical remote sensing is an effective one of shallow water depth measurement.In the past studies,traditional models such as simple theory and semi-empirical semi-empirical are often used to invert the shallow water depth,and the acceptable results are obtained.However,there are still some problems to be improved.Such as the selection of water depth inversion factor is not enough attentioned,the model is often simplified and the spatial differences of the model accuracy is often ignored in the inversion process.So,the precision of the model still has the potential of further optimization.Focusing on issues above,using WorldView-2 image and single-beam sounding data,according to the principle of optical remote sensing sounding and combining the characteristics of the traditional model with that of the artificial neural network model(ANN),this paper proposed an optical remote sensing depth inversion scheme based on multilevel decision-making to improve the accuracy of depth inversion and provide a method for future research on shallow bathymetry.The main research contents are as follows:(1)Denoising using assemble BM3D method.In order to remove the image noise and enhance the image quality of WorldView-2 image,an integrated BM3D denoising method is proposed according to the characteristics of the origin BM3D.And its effectiveness in denoising is verified.At last,removing noise in WorldView-2 image is completed by using the proposed method.(2)Bathymetry using an optical remote sensing based multilevel decision-making scheme.A multilevel decision-making depth inversion scheme,which combine traditional model and artificial neural network model,is proposed.Firstly,the water depth inversion factor library is constructed,and the partial factor with the largest correlation coefficient is selected from the library.The inversion effect via traditional model is analyzed to determine the optimal depth inversion factors which are to composing"strong correlation factor".Then,the feature space,made of strong correlation factor,as input features,the neural network depth inversion model is established,and the secondary water depth inversion using the single network model and result analysis are complied to evaluate the drawbacks of the single network model.Then according to the sencondary inversion result,depth space is split.And then the multi-network model is built according to different depth ocean space,and is used to deprive the water depth.Hereafter,the result is assessed.As these,the research realises whole process of optical remote sensing bathymetry based on multistage decision.Finally,the whole scheme is evaluated.The study concluded that:(1)The quality of WorldView-2 image enhances greatly after preprocessing such as radiation correction,sun-glint removing and deep water correction is carried out.The object-oriented method is suitable for the separation of land and water for high-resolution remote sensing images.The multi-band threshold segmentation can effectively identify wave area and prevent the fail inversion of the locations with white-hat covered.(2)BM3D method is obvious adept in removing gaussian noise,but it has no effect on removal of impulse noise.However,the improved integrated BM3D method can remove the mixed noise efficiently.Not only its denoising effect is better,but also the edge,texture and gradient changes in the image are better kept when denoising than other methods.WorldView-2 image is enhanced significantly after denoising by integrated BM3D method.(3)Among all the multi-spectral bands of WorldView-2,the correlation between blue/green band and water depth is large;logarithmic processing can significantly improve the correlation,and the correlation between the red band and the water depth is greatly increased after treatment;Spectral linear processing technique can remove linear background and noise so as to has great potential in the inversion of water depth and the spectral linear processing factor has a high correlation with water depth.Subject to constraints of light's water permeability capacity,the traditional models don't have a good relult for full depth inversion,however opponent for 30 meters(minimum relative error is 23%).In all water depth inversion factors,the logarithmic processing factor,the ratio factor and the spectral linear processing factor are more sensitive to water depth,and setting the logarithm of the water depth value as output of model is more conducive to water depth inversion.(4)The probability of result model approaching to the global optimal solution can be more maximized when using the "early stop" method to avoid overfitting of the neural network model and alleviating the local minimum problem by combining the momentum method,the Levenberg-Marquart optimizing BP method and the multi-initial parameter networks method.It can effectively alleviate the problem that the single model bathymetry will reduce the detail precision of different water-depth space under keeping model generalization ability when taking the above-mentioned over-fitting and local minimum problem into considering and using the multi-network model.(5)The "multi-level decision-making optical remote sensing depth inversion scheme",of which hard-core is artificial neural network model,using the traditional model as a supplement,has taken the light transmission characteristics and the water space differences into account.By using the scheme,the relative inversion error of derived shallow sea depth of study area in South China Sea when taking factors such as atmosphere,water environment and tidal into consideration,reduces to 8%and RMSE to sub-meter,which shows us that the inversion precision is greatly improved and reveals the scheme an effective depth inversion strategy.