一种基于BP算法学习的小波神经网络

为发展 Szu的基于信号表示的小波神经网络 ,提出一种多输入多输出的小波网络模型 ,网络隐层采用框架小波函数、输出层采用 Sigmoid激励函数 ,并选用“熵误差函数”以加速网络的学习速度。奇偶判别和混沌时间序列预测例子的实验结果表明了它具有良好的函数逼近能力和推广能力 ,收敛速度和均方误差均优于相同结构的多层感知器模型。     

深度学习的BP神经网络在GNSS水准拟合中的应用

针对误差逆向传播BP (back propagation)神经网络在GNSS水准拟合中存在梯度消失、陷于局部最小点的问题,通过使用深度学习中的分段线性整流函数Relu(rectified linear units)作为神经元激活函数,自适应矩估计Adam (adaptive moment estimation)算法作为网络优化函数,提出了一种基于深度学习的BP神经网络模型。研究结果表明:改进后的BP神经网络内外符合精度分别提高近50%和25%,可达0.9 cm和2.4 cm,为GNSS水准拟合提供了新的思路。     

基于粒子群优化神经网络的水深反演模型

针对直接采用BP神经网络反演水深收敛速度慢,且易陷入局部最优的问题,提出了一种基于粒子群(PSO)优化BP神经网络的水深遥感新模型。该模型首先利用粒子群算法对BP神经网络的权重和阈值进行优化,然后将该优化值作为BP神经网络的初始值,最后再将PSO优化后的模型用于测试海区的反演精度评估。实验结果表明,该模型的网络收敛速度明显加快,水深反演的精度也得到提高。     

基于头脑风暴优化算法与BP神经网络的海水水质评价模型研究

针对基于传统BP神经网络的海水水质评价模型存在易陷入局部极小等问题,提出了一种新的利用头脑风暴优化算法(BSO)优化BP神经网络的海水水质评价模型(BSO-BP)。该模型引入具有全局寻优特点的头脑风暴优化算法,用于模拟人类提出创造性思维解决问题的过程,具有强大的全局搜索和局部搜索的能力,同时利用BP神经网络所具有良好的非线性映射能力、学习适应能力和容错性,最大程度上考虑到海洋水质评价因素的非线性和非平稳的关系,得到BP神经网络的各层权值、阈值的最优解,使得海水水质评价结果准确合理。并以胶州湾海域的12个监测站位的监测数据作为评价样本进行水质评价,实验结果表明该评价模型能够克服局部极小问题,评价结果准确性较高,并具有一定的实用性。     

基于混合密度网络的NSCAT散射计海面风场反演

以NSCAT散射计数据为例,介绍了一种神经网络反演海面风场的方法.风速的反演是基于多层感知器网络;多解风向的反演是基于多层感知器网络和混合密度模型组合而成的混合密度网络,其中的核函数采用高斯函数的形式.通过与欧洲中期天气预报模式风场和现场浮标数据对比,证明了该神经网络反演海面风场的有效性.     

基于人工神经网络的赤潮预测模型

本文利用非线性时间序列预测模型,将海洋预报和人工神经网络BP算法相结合,提出了基于神经网络的海洋预报模型;运用改进的三层BP(Back Propagation)神经网络模型对海洋气象进行赤潮灾害监测和预报;同时针对仿真结果进行分析,结果表明该模型具有较好的预测能力。     

基于灵敏度分析的海洋油气资源BP神经网络预测模型的优化

作者针对BP神经网络结构设计中存在的问题,提出利用灵敏度分析方法对BP神经网络预测模型进行优化。通过BP算法与参数灵敏度分析的结合,寻找网络输入属性与输出属性之间的影响因子;在保证精度的前提下优选网络输入属性,简化网络结构,以增强网络的泛化能力,减少人为主观因素对网络设计的影响。最后以海洋油气资源预测为例,结合实测资料建立BP神经网络预测模型并进行了优化及预测精度评价,表明优化后的模型既能有效提高油气资源预测结果的稳定性,又不损失预测精度。     

雷达高度计海况偏差估计非参数模型研究

本文基于Jason-2高度计,采用核函数估计（NW）和局部线性回归估计(LLR)两种非参数估计方法,选用高斯(Gaussian)核函数和球谐(Epanechnikov)核函数及固定带宽和局部可调带宽。对不同组合形式的模型进行优选,确定LLR估计方法的Epanechnikov核函数、局部可调带宽为最优非参数模型。通过对最优非参数模型和参数模型结果进行对比分析表明,非参数模型在北高纬度区域表现更优,而在中低纬度及南纬区域参数模型不失优势。将非参数模型应用于我国HY-2A高度计,得到与以上同样的结论。     

基于决策树与二分分割算法的BP神经网络在赤潮等级预测中的应用研究

针对赤潮灾害等级预测难的现状,提出了一种基于C4.5决策树与二分分割算法优化的BP(反向传播)神经网络赤潮等级预测模型。该模型针对传统BP神经网络输入参数难以选择和隐含层节点数量难以确定的问题,通过决策树分类获取最优的属性组合,来解决输入参数难以选择的问题;通过"二分分割算法",来解决隐含层节点数难以确定的问题。实验结果表明,该模型在青岛近海海域赤潮灾害等级预测中,预测结果的均方根误差(RMSE)小于传统BP神经网络的预测误差,并且在网络训练时间上有所缩短,预测精度上有所提高,能够获得良好的预测结果,可为赤潮等级预测提供新的解决方法。     

基于遗传算法和BP神经网络的海洋工程材料腐蚀预测研究

为提高海洋工程材料腐蚀速率预测的精度,提出了一种基于遗传算法(GA)优化反向传播(Back Propagation,BP)神经网络的海洋工程材料海洋环境腐蚀速率预测模型。通过遗传算法对BP神经网络的权值和阈值进行优化,利用优化后的BP神经网络对试验数据进行预测。GA-BP模型选取具有代表性的2Cr1312不锈钢、Q235B碳钢和6082铝合金三种基本海洋工程材料数据进行试验,预测结果误差小于传统BP神经网络,并且在网络训练时间上有所缩短,预测精度上有所提高。本模型在海洋工程材料于海洋环境中腐蚀速率的实际预测中具有良好的推广价值。     

基于交叉证认的海洋磁力测量数据小波去噪研究

针对海洋磁力测量数据中存在的高斯噪声,采用基于交叉证认的小波阈值去噪方法进行噪声去除。首先对磁力数据进行小波分解,利用交叉证认方法自动识别小波分解的信号层与噪声层,然后对噪声层小波系数进行阈值处理,进一步提取噪声层中的有用信息,最后进行小波重构得到去噪信号。同时提出了一种改进阈值处理函数,能够改善传统软、硬阈值函数存在的缺陷,提高重构信号的精度。实验分析表明,当噪声水平小于1.5nT时,相比于传统软、硬阈值函数的小波阈值去噪方法,该方法可更好地去除海洋磁力数据中的高斯噪声,并可较好地保留数据中的有用细节信息,使数据质量得到提高。     

Toward parameterization of the eddy field near the Gulf Stream

Analysis of measurements from two long-term moored arrays in and near the Gulf Stream suggests a simple parameterization of eddy spatial covariance statistics: a parameterization that can be referred to as “quasi-homogeneous and isotropic”. Taking the normalized covariance function (i.e. the correlation function) for streamfunction to be homogeneous and isotropic and assuming motions to be horizontally nondivergent and hydrostatic permit the velocity and temperature covariances to be derived from the streamfunction covariance. Statistical tests indicate that deviations from these assumptions are indistinguishable from Gaussian random noise. The spatial correlation function used in Gaussian with a decay scale of about 140 km, which is only weakly depth dependent. A simple form is also suggested for the vertical lag dependence. This parameterization permits calculation of derived quantities such as the eddy vorticity flux divergence which is discussed in the context of the mean potential vorticity balances for the depth integrated circulation and for the subthermocline layer. The divergence of the relative vorticity flux is found capable of driving two counter-rotating gyres of strength 30–40 Sv on either side of the Stream, as are observed. The “thickness flux” dominates the lower layer eddy potential vorticity flux and is of the correct sign to make the recirculation more barotropic. The lower layer eddy forcing is weak and the gyres exist in a region of nearly uniform mean potential vorticity.     

Range and Doppler information from fourth-order spectra

In an active sonar system setting, a novel method is evaluated that extracts range and Doppler information from a Doppler-spread active sonar echo. The Doppler spreading is based on H. Van Trees' (1971) Gaussian amplitude modulating model, and the new method is based on the Fourier transform of a special case of the fourth-order cumulant. Specifically, from the envelope of the Gaussian amplitude modulated echo of a transmitted coded pulse train, the second-order spectrum and the Fourier transform of a special case of the fourth-order cumulant are derived and analyzed for this ability to extract range and Doppler information. It is shown that the method can theoretically extract range and Doppler information without degradation. The reason for this result is that a special case of the fourth-order cumulant is independent of the covariance of the Gaussian amplitude modulating function. These methods are also simulated and compared with the simulated results of the range-Doppler ambiguity function. This shows that the ambiguity function and the second-order spectrum are degraded due to the Gaussian amplitude modulation. The results are further demonstrated by simulating the three range-Doppler extraction methods for the received echo in noise     

Vertical structure of the upper ocean from profiles fitted to physically consistent functional forms

A method for characterizing the upper ocean structure is developed. Each temperature (density) profile is fitted by an ideal function based on the assumption that the permanent and seasonal thermoclines can be approximated respectively by steady state and transients of turbulent-diffusive processes and that the mixed layer can advance sharply under external forcing. The ideal profile is composed of two pieces joined at the mixed layer depth (MLD). The upper part is a constant; the part below the MLD is a product of an exponential decay and a Gaussian, representing the seasonal thermocline and decaying asymptotically to a straight line that describes the permanent thermocline. The composition of an exponential decay and a Gaussian accurately fits a wide family of solutions of the diffusion equation and includes the case of a shift of the boundary. The ideal fit for each profile relies on six adjustable parameters including the MLD. As the function is non-linear and non-differentiable, a Differential Evolution optimization algorithm is proposed to make the fitting. The solution gives a good estimate of the MLD based on the topology of the profile. It also provides a measure of the gradient and the shape of each profile, which are intuitive parameters for characterizing the upper ocean structure with direct applicability in ecosystem models. The algorithm is applied to a time series of monthly conductivity–temperature–depth (CTD) profiles from a hydrographical station in the southern Bay of Biscay. The construction of a local climatology of the vertical structure evolution (mixed layer development) is presented as a practical application. Other potential uses of the method are also discussed.     

基于主成分分析的BP神经网络模型的形变预测方法

主成分分析可以提取形变主要信息,BP神经网络具有很强的预测功能,提出将两者相结合用于形变监测数据处理。通过MATLAB编程实现了该算法,并用实测数据进行验证,证明此方法能够提高预测数据的精度和可靠性。结果表明：与其他方法相比,基于主成分分析的改进BP神经网络能取得更好的预测效果。     

Curvelet变换域侧扫声纳图像增强算法

针对海底侧扫声纳图像对比度低、纹理弱、噪声严重等问题,提出了一种基于第二代Curvelet变换的声纳图像增强算法。首先对原始声纳图像进行多尺度、多方向的Curvelet变换分解,得到低频子带和高频子带;然后引入非线性S型函数对低频系数进行处理,提高图像整体的对比度;采用一种可以避免过度增强的新型非线性函数对各尺度的高频子带系数进行处理,提高图像整体的对比度,增强图像边缘和纹理细节,并通过估计噪声水平设定阈值进行阈值降噪。最后经Curvelet逆变换得到增强图像。实验表明,该方法不仅改善了海底侧扫声纳图像对比度低的问题,而且降低了噪声,突出了声纳图像的边缘和纹理细节。     

Sea surface temperature retrieval based on simulated microwave polarimetric measurements of a one-dimensional synthetic aperture microwave radiometer

Compared with traditional real aperture microwave radiometers, one-dimensional synthetic aperture microwave radiometers have higher spatial resolution. In this paper, we proposed to retrieve sea surface temperature using a one-dimensional synthetic aperture microwave radiometer that operates at frequencies of 6.9 GHz, 10.65 GHz,18.7 GHz and 23.8 GHz at multiple incidence angles. We used the ERA5 reanalysis data provided by the European Centre for Medium-Range Weather Forecasts and a radiation transmission forward model to calculate the model brightness temperature. The brightness temperature measured by the spaceborne one-dimensional synthetic aperture microwave radiometer was simulated by adding Gaussian noise to the model brightness temperature.Then, a backpropagation(BP) neural network algorithm, a random forest(RF) algorithm and two multiple linear regression algorithms(RE1 and RE2) were developed to retrieve sea surface temperature from the measured brightness temperature within the incidence angle range of 0°–65°. The results show that the retrieval errors of the four algorithms increase with the increasing Gaussian noise. The BP achieves the lowest retrieval errors at all incidence angles. The retrieval error of the RE1 and RE2 decrease first and then increase with the incidence angle and the retrieval error of the RF is contrary to that of RE1 and RE2.     

A hybrid Lagrangian-Eulerian numerical model for sea-ice dynamics

A hybrid Lagrangian-Eulerian(HLE) method is developed for sea ice dynamics,which combines the high computational efficiency of finite difference method(FDM) with the high numerical accuracy of smoothed particle hydrodynamics(SPH).In this HLE model,the sea ice cover is represented by a group of Lagrangian ice particles with their own thicknesses and concentrations.These ice variables are interpolated to the Eularian gird nodes using the Gaussian interpolation function.The FDM is used to determine the ice velocities at Eulerian grid nodes,and the velocities of Lagrangian ice particles are interpolated from these grid velocities with the Gaussian function also.The thicknesses and concentrations of ice particles are determined based on their new locations.With the HLE numerical model,the ice ridging process in a rectangular basin is simulated,and the simulated results are validated with the analytical solution.This method is also applied to the simulation of sea ice dynamics in a vortex wind field.At last,this HLE model is applied to the Bohai Sea,and the simulated concentration,thickness and velocity match the satellite images and the field observed data well.