洪水预报自回归实时校正多步外延方法研究

自回归模型在洪水预报实时校正中应用广泛。针对自回归模型进行连续多时段校正时中间误差系列缺失问题,提出一种基于历史洪水预报误差系列的样本重组自回归外延方法,以淮河流域王家坝断面为背景,选用洪量相对误差、洪峰相对误差、峰滞时间和确定性系数四个指标开展校正效果评估,并与时程递推外延方法对比。结果表明:样本重组外延方法可以提升洪水预报精度,延长洪水预报有效预见期,特别在降低洪量误差和提高洪水过程的拟合精度上优势更为显著。同时,该方法泛化能力较强,具有实用价值。     

基于多模式降水集成的陆气耦合洪水预报

针对基于单模式陆气耦合模型在洪水预报中存在的不确定性问题,利用三种数值天气预报模式MC2、GEM与T213开展基于多模式降水集成的陆气耦合洪水预报研究。通过淮河流域2007～2009年3场暴雨洪水验证表明:基于多模式降水集成的陆气耦合洪水预报可以有效地减小不确定性、提高精度和稳定性;但基于多模式降水集成的陆气耦合洪水预报对小尺度局地强降水仍存在较大的不确定性,今后需对此展开进一步的研究,以给出更加稳定、可靠的洪水预报结果。     

特征加权FCM算法在洪水样本分类中的应用

基于聚类分析和模糊数学的基本原理，对历史洪水建立属性和数值特征的洪水样本，并运用特征加权FCM算法对流域历史洪水特征样本进行聚类分析。在分散式新安江三水源模型的基础上对不同聚类的洪水分别进行参数率定．利用样本特征的模糊识别以提高实时水文预报的精度。     

利用涨落洪规律建立洪水预报模型

为了提高大江河洪水预报精度,结合本地大江河洪水特性,提出利用涨落洪规律建立洪水预报模型。该模型可以解决以下两个问题：一是大江河预报河段上游站不能通过降雨径流关系推求入流过程,而需要作下游站洪水过程预报问题;二是按马斯京根法进行河道流演算的没有预见期的问题。     

多分类支持向量机在滑坡稳定性判识中的应用

如何准确地判识和评价滑坡的稳定性一直是滑坡研究中的关键问题。基于多分类支持向量机的基本理论,利用三峡库区的37个典型滑坡（27个训练样本,10个测试样本）,建立了滑坡稳定性判识的多分类支持向量机模型,并与距离判别分析方法进行了比较。结果表明,SVM模型对测试样本和训练样本的判识准确率均达到100％,而距离判别法对测试样本和训练样本的判识准确率分别为80％和77.8％,前者的判识精度明显优于后者。在此基础上,将SVM模型运用于溪洛渡库区牛滚凼滑坡的稳定性判识中,结果与实际情况吻合较好。     

三峡库区陆气耦合研究及应用

以延长洪水预见期、提高预报精度为目标,研究气象水文耦合机制,利用数值天气预报模式WRF(Weather Research and Forecasting)驱动分布式VIC(Variable Infiltration Capacity)水文模型,构建三峡库区陆气耦合洪水预报系统,并对2007~2008年期间四场暴雨洪水进行日滚动预报试验。结果表明,WRF模式在三峡库区内有着良好的短期降水预报精度,基于数值天气预报模式和分布式水文模型的陆气耦合洪水预报系统能有效延长三峡入库洪水预见期、提高洪水预报精度,具有较大的应用潜力。     

BP神经网络洪水预报模型在洪水预报系统中的应用

采用相关分析法,在区域降水、观测断面流量(或水位)因子中识别出影响预报断面径流过程的主要变量,在多个观测断面的数据均为流量情况下,采用基于时延组合的合成流量为影响预报断面径流过程的变量,采用自相关分析法,识别出影响预报断面径流过程的前期流量(或水位),以这些变量为BP神经网络模型的输入,以预报断面的流量(或水位)为模型的输出,在BP神经网络隐层节点数自动优选的基础上,构建了基于BP神经网络的洪水预报模型。将模型载入中国洪水预报系统中,应用结果表明:模型在历史洪水训练样本具有一定代表性的情况下,可获得较高的预报精度。     

基于状态变量初值修正的洪水预报方法研究

针对现有状态变量初值修正方法的不足,提出了一种基于状态变量初值修正的洪水预报方法(ISVC方法)。该方法以洪水起涨阶段平稳期的预报与实测流量过程间的残差为目标函数,采用粒子群优化算法对状态变量初值进行修正,利用修正后的状态变量初值开展洪水预报。将ISVC方法与新安江模型结合,采用赛塘流域进行了检验。结果表明,ISVC方法切实有效,能提高预报精度。ISVC方法具有一定的独立性、较强的适用性、一定的实用性等特点,具有一定的推广应用价值。     

新安江模型与水箱模型在柘溪流域适用性研究

为提高柘溪流域洪水预报精度,充分合理利用洪水资源,缓解该流域下游的防洪压力,同时为水库防洪调度以及经济运行提供科学合理的决策依据,研究了集总式概念性水文模型新安江模型及水箱模型在该湿润流域的适用性。选取该流域2004~2015年实测数据,应用三目标MOSCDE算法分别优选三水源新安江模型以及水箱模型参数,从而对该流域划分的多个子流域单元进行了场次洪水模拟计算,并对比分析不同流域单元两种模型的模拟结果,探究不同模型结构在柘溪流域场次洪水模拟中的差异,分析总结这两种模型在该流域的适用性。结果表明:两种模型模拟效果比较接近,均可以达到《水文情报预报规范》规定的作业预报精度要求,且对于流域内的大洪水的模拟效果也比较理想。从洪量以及洪水过程方面分析,两个模型的模拟效果比较接近;从洪峰拟合角度分析,新安江模型比水箱模型更适合该流域。     

无资料地区水文模型参数移植不确定性分析

无资料地区降水径流模拟是水文学研究的国际前沿和热点问题。水文模型参数移植是无资料地区降水径流模拟的重要方法,对径流模拟精度具有重要的影响。利用核密度估计和蒙特卡罗随机模拟等方法,构建了一种水文模型参数移植误差驱动的无资料地区径流模拟不确定性定量评估框架。以广西壮族自治区42个有水文监测站点的典型中小河流为研究对象,率定新安江模型参数并模拟日径流和洪水过程,将42个典型流域依次假定为无资料流域,利用基于回归分析、相似流域和机器学习的参数移植方法,模拟无资料地区的洪水过程并识别最优的参数移植方法,分析移植法估算的模型参数值和直接率定值相比误差的概率分布特征,定量评估模型参数移植误差带来的径流模拟不确定性。研究结果表明：(1)基于回归分析的参数移植法模拟无资料地区洪水过程的精度优于相似流域法,优选的机器学习算法比传统回归分析法和相似流域法的计算精度提高了7%～15%;(2)与模型参数率定值相比,移植方法计算的模型参数具有一定的误差,对洪水模拟敏感参数的误差小于不敏感参数;(3)受模型参数移植误差的影响,利用蒙特卡罗法随机模拟的洪水过程具有一定的不确定性,洪量和洪峰相对误差的主要区间分别为10...     

基于因子分析的GA-E LM模型岩溶地面塌陷预测

岩溶地面塌陷是多种因素共同作用的结果,具有隐蔽性和突发性的特点.为了快速、准确地对岩溶地面塌陷进行预测,提出了一种在因子分析的基础上利用遗传算法(GA)优化的极限学习机(ELM)岩溶地面塌陷预测模型.选取8个岩溶地面塌陷影响因素,利用因子分析提取5个公因子,然后输入GA-ELM模型进行预测.利用20组实例作为样本进行学...     

基于极限学习机的GF-2影像分类

遥感图像分类是提取图像有效信息过程中重要的一部分,为了探寻最优的分类方法,许多机器学习算法逐步应用于遥感分类中。极限学习机（extreme learning machine,ELM）以其高效、快速和良好的泛化性能在模式识别领域得到广泛应用。本文采用训练速度快、运算量小的极限学习机算法与支持向量机（support vector machines,SVM）算法和最大似然法进行分类对比,对高分辨率遥感图像进行分类,分析极限学习机算法对于遥感图像分类的准确度等性能。选取吉林省长春市部分区域的GF-2遥感数据,将融合后的影像设置为原始数据,利用3种方法进行分类。研究结果表明,极限学习机算法分类图像总体分类精度达到85%以上,kappa系数达到0.718,与其他分类方法相比分类准确度较高,且极限学习机运行时间比支持向量机运行时间约短2 480 s,约为支持向量机运行时间的1/8,因此具有良好的性能和实用价值。     

基于嵌入式索引的水文时间序列相似性搜索模型

相似模式挖掘已成为水文领域一个重要研究方向。对水文数据的相似性挖掘,有利于掌握水文数据变化规律和趋势,为洪水预报、防洪调度提供支持,是具有重要意义的工作。为此,在引入时间序列嵌入索引的基础上,结合水文时间序列的特点提出水文时间序列的快速搜索方法。该方法通过序列分割、聚类和参考集训练从原始序列中获取参考序列集,在此基础上通过索引计算方法,将相似性搜索过程映射到欧氏向量空间的搜索,从而提高了搜索效率。     

基于岩样细观图像深度学习的岩性自动分类方法

在现阶段的岩土工程中,通常采用人工识别的方法来判别岩样种类,不仅耗时长、专业性强,还易受主观因素影响,准确率不理想。随着计算机技术的发展,机器学习逐渐被应用于岩性的自动识别,开启了岩样分类的新路径。本文以重庆市主城区4种典型岩样（泥岩、砂质泥岩、泥质砂岩和砂岩）的细观图像为研究对象,基于Inception V3卷积网络模型和迁移学习算法,建立了岩样细观图像深度学习模型,并完成了训练学习。结果显示：模型在训练1 000次后,训练集中的分类准确率达到92.77%,验证集中的分类准确率为76.31%。其中,验证集中的砂岩识别准确率为97.28%,泥岩识别准确率为81.85%,泥质砂岩识别准确率为72.59%,砂质泥岩识别准确率为72.35%。与现有的机器学习方法相比,本识别模型不仅可以自动识别岩性极为相近的岩样,而且具有较好的识别准确率、鲁棒性和泛化能力。     

Displacement prediction model of landslide based on a modified ensemble empirical mode decomposition and extreme learning machine

In this paper, an M–EEMD–ELM model (modified ensemble empirical mode decomposition (EEMD)-based extreme learning machine (ELM) ensemble learning paradigm) is proposed for landslide displacement prediction. The nonlinear original surface displacement deformation monitoring time series of landslide is first decomposed into a limited number of intrinsic mode functions (IMFs) and one residual series using EEMD technique for a deep insight into the data structure. Then, these sub-series except the high frequency are forecasted, respectively, by establishing appropriate ELM models. At last, the prediction results of the modeled IMFs and residual series are summed to formulate an ensemble forecast for the original landslide displacement series. A case study of Baishuihe landslide in the Three Gorges reservoir area of China is presented to illustrate the capability and merit of our model. Empirical results reveal that the prediction using M–EEMD–ELM model is consistently better than basic artificial neural networks (ANNs) and unmodified EEMD–ELM in terms of the same measurements.     

Strength of ensemble learning in multiclass classification of rockburst intensity

Rockbust is a violent expulsion of rock due to the extreme release of strain energy stored in surrounding rock mass, leading to considerable damages to underground strucures and equipment, and threatening workers' safety. As the operational depth of engineering projects increases, a larger number of factors influence the mechanism of rockburst. Therefore, accurate classification of rockburst intensity cannot be achieved based on conventional criteria. It is urgent to develop new models with high accuracy and ease to implement in practice. This study proposed an ensemble machine learning method by aggregating seven individual classifiers including back propagation neural network, support vector machine, decision tree, k-nearest neighbours, logistic regression, multiple linear regression and Naïve Bayes. In addition, we proposed nine data imputation methods to replace the missing values in the compiled database including 188 rockburst instances. Five-fold cross validation and the beetle antennae search algorithm are used to tune hyperparameters and voting weights of the individual classifiers. The results show that the rockburst classification accuracy obtained by the classifier ensemble has increased by 15.4% compared with the best individual classifier on the test set. The predictor importance obtained by the classifier ensemble shows that the elastic energy index is the most sensitive input variable for rockburst intensity classification. This robust ensemble method can be extended to solve other classification problems in underground engineering projects.     

Improved flood susceptibility mapping using a best first decision tree integrated with ensemble learning techniques

Improving the accuracy of flood prediction and mapping is crucial for reducing damage resulting from flood events. In this study, we proposed and validated three ensemble models based on the Best First Decision Tree (BFT) and the Bagging (Bagging-BFT), Decorate (Bagging-BFT), and Random Subspace (RSS-BFT) ensemble learning techniques for an improved prediction of flood susceptibility in a spatially-explicit manner. A total number of 126 historical flood events from the Nghe An Province (Vietnam) were connected to a set of 10 flood influencing factors (slope, elevation, aspect, curvature, river density, distance from rivers, flow direction, geology, soil, and land use) for generating the training and validation datasets. The models were validated via several performance metrics that demonstrated the capability of all three ensemble models in elucidating the underlying pattern of flood occurrences within the research area and predicting the probability of future flood events. Based on the Area Under the receiver operating characteristic Curve (AUC), the ensemble Decorate-BFT model that achieved an AUC value of 0.989 was identified as the superior model over the RSS-BFT (AUC = 0.982) and Bagging-BFT (AUC = 0.967) models. A comparison between the performance of the models and the models previously reported in the literature confirmed that our ensemble models provided a reliable estimate of flood susceptibilities and their resulting susceptibility maps are trustful for flood early warning systems as well as development of mitigation plans.     

Mapping flood susceptibility in an arid region of southern Iraq using ensemble machine learning classifiers: a comparative study

This study examined the efficacy of three machine ensemble classifiers, namely, random forest, rotation forest and AdaBoost, in assessing flood susceptibility in an arid region of southern Iraq. A dataset was created from flooded and non-flooded areas to train and validate the ensemble classifiers using a binary classification scheme (1—flood, 0—non-flood). The prepared dataset was then partitioned into two sets with a 70/30 ratio: 70% (2478 pixels) for training and 30% (1062 pixels) for testing. A total of 10 influential flood factors were selected and prepared based on data availability and a literature review. The selected factors were surface elevation, slope, plain curvature, topographic wetness index, stream power index, distance to rivers, drainage density, lithology, soil and land use/land cover. The information gain ratio was first utilised to explore the predictive abilities of the factors. The predictive performances of the three ensemble models were compared using six statistical measures: sensitivity, specificity, accuracy, kappa, root mean square error and area under the operating characteristics curve. The results revealed that the AdaBoost classifier was the best in terms of the statistical measures, followed by the random forest and rotation forest models. A flood susceptibility map was prepared based on the result of each classifier and classified into five zones: very low, low, moderate, high and very high. For the model with the best performance, i.e., the AdaBoost model, these zones were distributed over an area of 6002 km² (44%) for the very low–low zone, 2477 km² (18%) for the moderate zone and 5048 km² (40%) for the high–very high zones. This study proved the high capabilities of ensemble machine learning classifiers to decipher flood susceptibility zones in an arid region.