基于数值预报和随机森林算法的强对流天气分类预报技术

随机森林算法是当前得到较为广泛应用的机器学习方法之一,有着很高的预测精度,训练结果稳定,泛化能力强,解决多分类问题有明显优势。本文将随机森林算法应用于强对流的潜势预测和分类,分短时强降水、雷暴大风、冰雹和无强对流四种类别,基于2005—2016年NCEP 1°×1°再分析资料计算的对流指数和物理量,开展强对流天气的分类训练、0～12 h预报和检验,经2015—2016年独立测试样本检验表明,针对强对流发生站点的点对点检验,整体误判率为21. 9%,85次强对流过程基本无漏报,模型尤其适用于较大范围强对流天气。随机森林算法筛选的因子物理意义较为明确,和主观预报经验基本相符,模型准确率高,可用于日常业务。     

随机介质模型的一种构造方法

为更好地描述真实的地下介质特点,提出了一种局部随机位置或固定位置任意形状截取法并结合多尺度建模技术来共同构造随机介质模型的新方法。从现有的随机建模理论出发,在有效的建模区域内产生或者给定局部扰动位置,以此作为局部扰动分布区的中心点,引入局部扰动半径R和局部随机或给定扰动角度θ两个模型特征量,采用局部位置任意形状截取法并充填各种随机介质,从而构造各种不同的随机介质模型。给出了局部位置任意形状截取法的基本步骤和应用实例。     

海底双相随机介质声反射的SVM神经网络分类识别研究

考虑海底沉积介质为双相介质,为了更好地模拟实际海底底质的不均匀性,将随机介质理论引入双相介质理论。首先,通过基于随机-双相介质理论的高阶有限差分数值技术模拟计算海底底质分别为泥质砂、泥、泥质砾时的地震反射波信号。然后利用小波变换分别求取不同底质的一次反射波的包络作为其特征向量,最后利用基于粒子群智能算法优化的支持向量机神经网络对这些反射波信号进行分类识别。为了进一步考察所用方法的抗噪能力,对正演得到的海底底质反射波信号分别加入10%、30%、50%的高斯白噪音之后再进行分类,支持向量机仍然取得了较好的分类预测效果。基于上述正演模拟及分类识别方法的论证,提出了一套行之有效的微机软件模拟海底沉积物分类识别的一般化流程,这将有利于开展海底沉积物反射特征的进一步研究。     

基于随机集合的非传统型有效波极值模型

The analysis and design of offshore structures necessitates the consideration of wave loads. Realistic modeling of wave loads is particularly important to ensure reliable performance of these structures. Among the available methods for the modeling of the extreme significant wave height on a statistical basis, the peak over threshold method has attracted most attention. This method employs Poisson process to character- ize time-varying properties in the parameters of an extreme value distribution. In this paper, the peak over threshold method is reviewed and extended to account for subjectivity in the modeling. The freedom in selecting the threshold and the time span to separate extremes from the original time series data is incorpo- rated as imprecision in the model. This leads to an extension from random variables to random sets in the probabilistic model for the extreme significant wave height. The extended model is also applied to different periods of the sampled data to evaluate the significance of the climatic conditions on the uncertainties of the parameters.     

基于f-x域时频非凸正则化低秩矩阵近似的共偏移距道集去噪方法

随机噪声压制是地震数据处理的关键环节,而时频稀疏低秩近似算法逐道处理地震数据过程中无法利用信号的道间相干性。为此,将时频稀疏低秩近似与f-x域去噪结合,提出一种f-x域时频非凸正则化低秩矩阵近似算法。该算法对f-x域中每一单频分量作时频分解后,再对时频系数矩阵作低秩矩阵近似计算,能够利用信号和噪声的时频谱差异实现非平稳信号去噪处理。与共炮点道集和共中心点道集相比,共偏移距道集具有平缓甚至接近水平的同相轴结构,基本满足f-x域去噪的线性同相轴假设前提,建议将所提算法应用于共偏移距道集去噪处理。通过数值模拟和实际地震数据试算,证明本文方法能够有效压制随机噪声,同时保持有效信号不被损害。     

Numerical Simulation of Multi-Directional Random Seas

Multi-directional irregular waves are simulated on the basis of the given directional spectrum using a double summation model, a single direction per frequency model and a single summation model. Their results are compared. It is shown that the single direction per frequency model proposed in this paper can generate a realistic wave field. The effects of the model parameters on the simulated results are also studied in this paper and corresponding suggestions are given.     

结合MST划分和RHMRF-FCM算法的高分辨率遥感图像分割

针对基于像素的HMRF-FCM算法抗噪性差以及对地物复杂边界分割精度低的问题,提出一种结合形状信息的静态MST区域划分和RHMRF-FCM算法的高分辨率遥感图像分割方法。该方法定义一种静态MST同质区域划分准则,借助MST能较好表达边界和形状信息、能较好抑制几何噪声的特点,解决地物复杂边界的表达和降低分割结果中几何噪声问题。首先,利用MST静态划分将图像域划分成若干个均质区域,假设每个均质区域内光谱测度服从独立同一的多元高斯分布。然后,在此基础上构建了区域隐马尔可夫随机场模型,以及建立基于信息熵和KL信息正则化项的模糊聚类目标函数。最后,采用偏微分方法对分割模型参数进行求解,从而得到全局最优分割结果。为验证本文方法,对WorldView-3高分遥感图像进行分割试验。定性、定量分析了尺度参数、光谱相似性参数和区域紧致度参数对最优分割结果的影响,并对比分析本文算法和eCognition软件中的多分辨率分割算法、分水岭算法。     

Otolith Microstructure Analysis Reveals Different Growth Histories of Japanese Sardine (Sardinops melanostictus) in the Oyashio Waters

After decades of low year classes,the stock of Japanese sardine(Sardinops melanostictus)has begun to recover since the mid-2000s.The hatch dates and otolith growth rates of age-0 juvenile sardine,which were collected in the subarctic Oyashio waters in autumn 2018,were determined from an otolith microstructure analysis.The sardines were hatched from late January to late April,while mostly in February and March.The otolith growth rate increased continuously up to 60 d after hatching and thereafter de-creased.The revealed growth rate in a crucial growth period is faster than that reported for juvenile sardines collected in the 1990s,which is coincided with the recent recovery trend of the sardine stock.Two groups with different hatch dates,growth histories,and migration routes were identified using unsupervised random forest clustering analysis.They were considered inshore and offshore migration individuals in accordance with recent researches.In the offshore group,a high proportion of sardine juveniles hatched late and grew faster in the Kuroshio-Oyashio transitional waters,a finding consistent with the hypothesis of growth-rate-dependent re-cruitment.This finding on the population composition and growth rate of juvenile sardine in the Oyashio waters can be a basis for an improved prediction of their survival and provides us with valuable information on the recruitment processes of this stock during the period of stock recovery.     

Flexible spectral methods for the generation of random fields with power-law semivariograms

Random field generators serve as a tool to model heterogeneous media for applications in hydrocarbon recovery and groundwater flow. Random fields with a power-law variogram structure, also termed fractional Brownian motion (fBm) fields, are of interest to study scale dependent heterogeneity effects on one-phase and two-phase flow. We show that such fields generated by the spectral method and the Inverse Fast Fourier Transform (IFFT) have an incorrect variogram structure and variance. To illustrate this we derive the prefactor of the fBm spectral density function, which is required to generate the fBm fields. We propose a new method to generate fBm fields that introduces weighting functions into the spectral method. It leads to a flexible and efficient algorithm. The flexibility permits an optimal choice of summation points (that is points in frequency space at which the weighting function is calculated) specific for the autocovariance structure of the field. As an illustration of the method, comparisons between estimated and expected statistics of fields with an exponential variogram and of fBm fields are presented. For power-law semivariograms, the proposed spectral method with a cylindrical distribution of the summation points gives optimal results.