Feed-forward vs recurrent neural network models for non-stationarity modelling using data assimilation and adaptivity

Abstract

Artificial neural networks (ANN) are nonlinear models widely investigated in hydrology due to their properties of universal approximation and parsimony. Their performance during the training phase is very good, and their ability to generalize can be improved by using regularization methods such as early stopping and cross-validation. In our research, two kinds of generic models are implemented: the feed-forward model and the recurrent model. At first glance, the feed-forward model would seem to be more effective than the recurrent one on non-stationary datasets, because measured information on the state of the system (measured discharge) is used as input, thereby implementing a kind of data assimilation. This study investigates the feasibility and effectiveness of data assimilation and adaptivity when implemented in both feed-forward and recurrent neural networks. Based on the IAHS Workshop held in Göteborg, Sweden (July 2013), the hydrological behaviour of two watersheds of different sizes and different kind of non-stationarity will be modelled: (a) the Fernow watershed (0.2 km²) in the USA, affected by significant modifications in land cover during the study period, and (b) the Durance watershed (2170 km²) in France, affected by an increase in temperature that is causing a decrease in the extent of glaciers. Two methods were applied to evaluate the ability of ANN to adapt on the test set: (i) adaptivity using observed data to adapt parameter values in real time; and (ii) data assimilation using observed data to modify inaccurate inputs in real time. The goal of the study is thus re-analysis and not forecasting. This study highlights how effective the feed-forward model is compared to the recurrent model for dealing with non-stationarity. It also shows that adaptivity and data assimilation improve the recurrent model considerably, whereas improvement is marginal for the feed-forward model in the same conditions. Finally, this study suggests that adaptivity is effective in the case of changing conditions of the watershed, whereas data assimilation is better in the case of climate change (inputs modification).     

Modeling of tensile strength of rocks materials based on support vector machines approaches

In the predicting of geological variables, artificial neural networks (ANNs) have some drawbacks including possibility of getting trapped in local minima, over training, subjectivity in the determining of model parameters and the components of its complex structure. Recently, support vector machines (SVM) has been found to be popular in prediction studies due to its some advantages over ANNs. Because the least squares SVM (LS‐SVM) provides a computational advantage over SVM by converting quadratic optimization problem into a system of linear equations, LS‐SVM method is also tried in study. The main purpose of this study is to examine the capability of these two SVM algorithms for the prediction of tensile strength of rock materials and to compare its performance with ANN and linear regression (MLR) models. Total porosity, sonic velocity, slake durability index and aggregate impact value were used as input in modeling applications. Favorite performance evaluation measures were employed to assess developed models. The results determined in study indicate that the SVM, LS‐SVM and ANN methods are successful tools for prediction of tensile strength variable and can give good prediction performances than MLR model. Although these three methods are powerful artificial intelligence techniques, LS‐SVM makes the running time considerably faster with the higher accuracy. In terms of accuracy, the LS‐SVM model resulted in error reductions relative to that of the other models. Copyright © 2012 John Wiley & Sons, Ltd.     

Determination of liquefaction susceptibility of soil: a least square support vector machine approach

This article adopts least square support vector machine (LSSVM) for determination of liquefactions susceptibility of soil based on standard penetration test data. Two models (Models I and II) have been developed. For Model I, input variables are cyclic stress ratio and standard penetration test value (N). Model II employs peak ground acceleration and N as input variables. The developed LSSVM models (Models I and II) give equations for determination of liquefaction susceptibility of soil. The performances of Models I and II are the same. The developed LSSVM gives probabilistic output. The results of LSSVM have been compared with the artificial neural network model. This article shows that N and the peak ground acceleration are sufficient input parameters for determination of liquefaction susceptibility of soil. Copyright © 2012 John Wiley & Sons, Ltd.     

基于概率神经网络的智能找矿方法——以四川雅江县木绒锂矿为例

概率神经网络是一种分类准确率高、泛用性强、可以包容一定数量错误样本的人工神经网络,极其适合勘查地球化学找矿中的预测找矿靶区。本文以四川雅江县木绒锂矿为例,运用概率神经网络搭建智能找矿模型,以已知区的Li元素及与其相关性强的Rb-Cs-Al-Fe元素作为训练指标,对模型进行训练,经过多次训练后将Spread值确定为0.08,使模型在训练集和测试集的准确率均大于80%,实现非线性的指标与成矿潜力的对应,得到本矿区的PNN模型,然后对预测区的样本数据进行预测,成功圈定了1处靶区。为检验靶区准确性,以Li、Rb、Cs元素数据累计频率的80%作为异常下限,圈出的异常区域与靶区位置基本重叠。对预测区进行了实地查证工作,发现两条红柱石带,其中一条与靶区位置吻合,表明该神经网络模型准确性高,可用于矿产勘查的预测研究。     

卷积神经网络和视觉注意力语义分割模型在高分辨率遥感影像分类中的性能分析

随着深度学习语义分割的快速发展,基于计算机视觉语义分割模型的高分辨率遥感影像分类方法也大量涌现。为系统定量地研究经典的和先进的视觉语义分割模型在遥感影像分类中的性能,在总结深度学习语义分割进展的基础上,选择9种基于卷积神经网络(CNN)和视觉注意力的语义分割算法,对米级和厘米级2个尺度的遥感数据集进行分析研究。在模型构建上基于计算机视觉通用的语义分割框架,训练时采用红绿蓝3波段遥感图像并基于ImageNet预训练权重进行迁移学习训练。研究结果表明：通用的语义分割模型通过常规训练设置进行训练能取得较好的遥感影像分类效果,部分地物的交并比(IoU)可以达到90%以上;基于视觉注意力的遥感影像分类模型的精度普遍高于基于CNN的模型,且MaskFormer能更有效地提取离散的地物信息;不同类别的精度最高值并不全在总体最优模型中,部分会存在于次优模型中;类似的地物在更高分辨率遥感数据集中可以获得更高的精度。     

非对称双环神经网络系统的稳定性和Hopf分岔

环型结构在神经网络中普遍存在,目前对环型神经动力学分岔研究大多数局限于单环情形.值得注意的是,神经网络由成千上万个神经元耦合而成,这些复杂的神经元网络结构不可能只由一个环形结构来准确表述,因此研究具多环拓扑的神经网络模型更具实际意义.本文提出了一种非对称双环神经元网络模型,选择单环的时滞和为分岔参数,分析了双环模型的稳定性和Hopf分岔.最后给出数值仿真对结论进行了验证.     

参数不匹配下含有泄露时滞与多个传输时滞的分数阶忆阻器神经网络的拟同步

本文主要讨论了一类在参数不匹配之下带有多个传输时滞与泄露时滞的分数阶忆阻器神经网络的拟同步问题.首先,给出了时滞线性反馈控制器;然后,根据分数阶微分包含、集合值映射理论和分数阶微分不等式理论等,得出了主从系统达到拟同步所满足的条件以及所对应的误差边界;最后,给出了一些数值模拟验证所得结论的有效性.     

阵风因子与大气边界层要素的关系及预报试验

利用2008—2016年舟山4个海岛气象站大风资料和欧洲中心的ERA-interim再分析资料,分析了阵风因子随平均风速、风向、小时、月份等分布的气候特征,统计阵风因子与边界层的大气稳定度、250～1000m风速与10m风速的比值、6h变温等要素的相关后,选取最佳预报因子,利用BP人工神经网络方法,根据不同因子组合对阵风进行循环试报。结果表明:①平均风速较小时阵风因子波动范围大;靠近大陆站点的阵风因子及来自陆地方向气流的阵风因子偏大。②白天11:00—16:00受太阳辐射影响大气湍流相对较强,阵风因子偏大。7—9月沿海受台风影响频繁,其阵风因子要大些,而11—12月阵风因子偏大则与来流方向的地表粗糙度较大有关。③阵风因子和边界层不同高度的风速与10m风速比值,及气温具有明显正相关,与边界层大气稳定度参数具有负相关,相关结果印证了阵风主要物理成因与动量的垂直湍流输送有关。④阵风循环试报表明最佳组模型试报的绝对误差及方差均比对比组模型减少约11%～25%,具有较好的预报效果。     

基于神经网络模型的生物扰动碳酸盐岩储集层识别与孔隙度预测——以塔里木盆地塔河油田奥陶系生物扰动碳酸盐岩储集层为例

塔里木盆地塔河油田奥陶系生物扰动碳酸盐岩储集层非常发育,但利用常规测井数据识别生物扰动储集层发育段和准确预测孔隙度难度较大。本文在对研究区16口取芯井奥陶系岩芯上生物扰动区域扰动等级划分的基础上,通过岩性标定测井,优选常规测井参数,基于BP神经网络模型分别建立了适合研究区生物扰动碳酸盐岩储集层识别和孔隙度预测的模型,并对建立的模型进行了有效性检验。结果表明：① 选择自然电位、自然伽马、井径、深侧向电阻率、浅侧向电阻率、补偿中子和密度等常规测井数据作为生物扰动碳酸盐岩储集层识别模型输入层的参数值,生物扰动指数（Bioturbation Index, BI）作为输出结果;选取rprop、sigmoid symmetric和sigmoid stepwise函数分别作为训练函数、隐含层和输出层的激活函数,建立节点数为3、层数为3的神经网络识别模型,识别效果好,适用于研究区奥陶系生物扰动碳酸盐岩储集层的识别。② 选择自然电位、自然伽马、井径、声波、补偿中子和密度值等常规测井数据作为输入层的参数值,对应深度上岩芯柱塞孔隙度测试结果和利用孔隙度样品检验模型计算得出的孔隙度结果作为输出结果,选取incremental、gaussian和sigmoid分别作为训练函数、隐含层和输出层的激活函数,建立节点数为4,层数为3的生物扰动碳酸盐岩储集层孔隙度预测模型,预测效果良好,适用于研究区奥陶系生物扰动储集层孔隙度的预测。该研究对定量表征研究区生物扰动储层特性、储量估算、油藏描述和储层地质建模等具有重要的借鉴意义。