Neural network based on rough sets and its application to remote sensing image classification

This paper presents a new kind of back propagation neural network(BPNN) based on rough sets,called rough back propagation neural network (RBPNN).The architecture and training method of RBPNN are presented and the survey and analysis of RBPNN for the classification of remote sensing multi-spectral image is discussed.The successful application of RBPNN to a land cover classification illustrates the simple computation and high accuracy of the new neural network and the flexibility and practicality of this new approach.     

基于GIS和神经网络的森林植被分类

本文综述了国际遥感分类研究,使用Landsat7 ETM＋遥感数据和地理辅助数据,应用BP神经网络方法,将莽汉山林场作为研究区进行了遥感影像的分类研究。比较了BP神经网络分类与最大似然、简单和复杂非监督分类法之间的类型与数量精度。BP神经网络分类的总类型精度是70.5%,总数量精度为84.65%,KAPPA系数是0.6455。结果说明BP神经网络的分类质量优于其他方法,其总的类型精度与其他三种分类方法相比分别增加了10.5%、32%和33%,总的质量精度增加了5.3%。因此,辅以地理参考数据的BP神经网络分类可以作为一种有效的分类方法。     

一种基于自适应谐振理论的TM遥感影像神经网络分类器

提出了一种基于自适应谐振理论建立起来的自组织模糊ARTMAP神经网络分类器。分析了ART神经网络的结构和工作原理,给出模糊ARTMAP神经网络分类的具体算法,并将其运用到TM遥感影像分类的实验中。结果表明模糊ARTMAP神经网络分类器的速度快,精度高,比常用的BP网络具有更好的性能。     

基于PCA-BPNN的多光谱遥感影像分类

基于BP算法的神经网络方法目前已广泛运用于遥感影像分类,提出一种主成分分析(PCA)与BP神经网络相结合的遥感影像分类方法——PCA-BPNN,实验证明该方法是可行并且有效的,在减少计算量和加快收敛的同时,提高了分类的精度。     

变差函数和神经网络结合的遥感影像分类方法研究

提出利用地统计学变差函数对遥感影像纹理信息进行提取，将变差函数得到的纹理信息与光谱信息相结合，运用神经网络进行分类的遥感影像分类方法。将该分类方法应用于试验区，并与最大似然法的分类结果进行对比分析，结果表明，该方法具有较高的分类精度。     

结合对抗网络与辅助任务的遥感影像无监督域适应方法

使用机器学习进行遥感影像标注的一个重要前提是有足够的训练样本,而样本的标注是非常耗时的。本文采用了域适应的方法来解决遥感影像场景分类中小样本量的无监督学习问题,提出了结合对抗网络与辅助任务的遥感影像域适应方法。首先建立了基于深度卷积神经网络的遥感影像分类框架;其次,为了学习到域不变特征,在标签分类器的基础上增加域分类器,并使域损失函数在其反射传播时的梯度与标签损失的梯度相反,从而保证域分类器不能区分样本来自于哪个域;最后引入了辅助分类任务,扩充了样本的同时使网络更具泛化能力。试验结果表明,本文方法优于主流的无监督域适应方法,在小样本遥感影像无监督分类中得到了较好的效果。     

轨道误差对InSAR相位影响的BP神经网络去除法

针对用二次多项式法去除轨道误差对InSAR相位影响时,须对干涉相位其他项分布性质作假设,且自身存在一定的模型缺陷的问题,该文提出用BP神经网络去除轨道误差对InSAR相位影响的方法。研究表明:BP神经网络法在使用时无须对干涉相位其他项分布性质作假设,模型更优。模拟实验中,轨道误差相位拟合残差更小;真实数据实验中,纠正后非形变区相位集中在零值附近,且波动趋势更为平稳。该方法一定程度上降低了传统二次多项式法的应用局限性。     

神经网络优化5G定位结果域的分析与评估

针对室内环境下的5G定位需求,提出了利用神经网络算法对粗略定位结果进行优化的方法,减小了多径、非视距传播造成的定位误差,改善了结果域的定位精度.优化算法利用测距定位中的到达时间(TOA)定位法和到达时间差(TDOA)定位法获得粗略定位结果,分别结合BP神经网络、Elman神经网络及通过遗传算法(GA)优化后的GA-BP神经网络、GA-Elman神经网络共利用4种神经网络进行训练,得到修正后的精确定位结果,并对4种神经网络算法进行了分析与评估. Elman算法相较于BP算法具有迭代收敛快、迭代次数少、误差改正好的特点,更适合5G定位结果域的优化;融入GA后结果精度均有所提高,其中GA-Elman算法能够训练得到最好的定位结果.     

基于BP网络的大坝变形分析与预报

人工神经网络具有表达非线性映射的性质，将BP神经网络模型用于大坝变形的拟合分析和预报研究，并用实例证明了该方法可以取得很好的拟合和预报结果。     

结合面向对象和深度特征的高分影像树种分类

针对传统手工提取特征方法需要专业领域知识,提取高质量特征困难的问题,将深度迁移学习技术引入到高分影像树种分类中,提出一种结合面向对象和深度特征的高分影像树种分类方法。为了获取树种的精确边界,该方法首先利用多尺度分割技术分割整幅遥感影像,并选择训练样本作为深度卷积神经网络的输入。为了避免样本数量少导致过拟合问题,采用迁移学习方法,使用ImageNet上训练的VGG16模型参数初始化深度卷积神经网络,并利用全局平局池化压缩参数,在网络最后添加1024个节点的全连接层和7个节点的Softmax分类器,利用反向传播和Adam优化算法训练网络。最后分类整幅遥感影像,生成树种专题地图。以安徽省滁州市的皇甫山国家森林公园为研究区,QuickBird高分影像作为数据源,采用本文方法进行树种分类。试验结果表明,本文方法树种分类总体精度和Kappa系数分别为78.98%和0.685 0,在保证树种精度的同时实现了端到端的树种分类。     

Study on Ensemble Crop Information Extraction of Remote Sensing Images Based on SVM and BPNN

High resolution remote sensing image contains abundant information, but remote sensing classification only based on spectral information is affected in the complex spectrum area. Crop area and other land-cover objects contain different texture features. This paper extracts texture information based on gray-level co-occurrence matrix and Gabor filters group, sets up spectrum-texture joint feature set. To enhance classification efficiency, Ensemble learning strategy is introduced to improve classical support vector machine and back propagation neural network classifiers in training process. To prove the effectiveness of proposed methods, several experiment images are utilized to execute experiments. Results indicate that proposed methods improve classification accuracy compared with classical algorithms significantly, and promote running efficiency compared with the situation of large sample, support corn area statistical process and yield estimation.     

基于BP神经网络的太阳黑子年均值预测

以1700年至2008年间太阳黑子数年均值为研究对象,在数字信号处理及神经网络的基础上,分析归一化自相关函数以及离散傅里叶变换(DFT)后的图像特征,得到可靠的太阳黑子活动周期,以此周期为依据设置参数建立BP(back propagation)神经网络,将已有的随机时间样本进行分组训练和验证预测,拟合神经网络输出预测值与实际值并计算其拟合优度,验证了将BP神经网络应用在太阳黑子活动方面有较高的科学性和可行性。     

A variable precision rough set approach to knowledge discovery in land cover classification

This paper presents a granular computing approach to spatial classification and prediction of land cover classes using rough set variable precision methods. In particular, it presents an approach to characterizing large spatially clustered data sets to discover knowledge in multi-source supervised classification. The evidential structure of spatial classification is founded on the notions of equivalence relations of rough set theory. It allows expressing spatial concepts in terms of approximation space wherein a decision class can be approximated through the partition of boundary regions. The paper also identifies how approximate reasoning can be introduced by using variable precision rough sets in the context of land cover characterization. The rough set theory is applied to demonstrate an empirical application and the predictive performance is compared with popular baseline machine learning algorithms. A comparison shows that the predictive performance of the rough set rule induction is slightly higher than the decision tree and significantly outperforms the baseline models such as neural network, naïve Bayesian and support vector machine methods.     

神经网络专家系统在大坝变形预测中的应用

探讨了将人工神经网络专家系统用于大坝变形预测的方法，给出了系统功能结构框图，并对各模块的功能进行了分析。     

基于GA-BP算法的大坝边坡变形预测模型

针对大坝安全预测采用传统的统计模型、确定性模型和混合模型存在的不足,应用遗传算法（GA）与基于误差反向传播算法（BP）相结合,构成GA-BP混合遗传算法,建立大坝边坡变形预测的遗传优化神经网络模型（GA-BP模型）。该模型利用神经网络的非线性映射能力、网络推理和预测功能及遗传算法的全局优化特征,克服BP算法易限人局部最小问题。通过该模型对某大坝的实际观测数据进行预测,表明GA-BP模型的预测具有精度高、收敛速度快的优点,在大坝的预测方面具有应用价值。     

分层神经网络分类算法

提高遥感图像分类精度一直是爱到普遍关注的焦点问题。近年来,人工神经网络技术和分 处理技术由于它们的许多优点受到广泛欢迎。本文把两种技术结合起来,提出了分层神经网络的概念,并基于此设计了一种分层神经网络分类算法。通过与最大似然法的对比实验表明,这种分层神经网络分类算法可以明显地提高分类精度,并对不规则分布的复杂数据具有很强的处理能力。     

遥感图像识别中粗糙集理论与神经网络的结合

由于传统神经网络与遥感图像信息量不相匹配 ,为此 ,提出将粗糙集理论集成至遥感图像神经网络识别中。首先分析了神经网络与粗糙集理论结合的可能性以及优势 ,在此基础上提出了基于粗糙集的遥感图像神经网络识别模型 ,并就其中的粗糙集方法处理样本特征集模块和遥感图像识别神经网络模块展开详细的分析。通过对比实验数据说明集成粗糙集理论的遥感图像神经网络识别能够有效提高遥感图像的识别效率 ,具有较强的现实意义     

An iterative approach based on contextual information for matching multi‐scale polygonal object datasets

Object matching facilitates spatial data integration, updating, evaluation, and management. However, data to be matched often originate from different sources and present problems with regard to positional discrepancies and different levels of detail. To resolve these problems, this article designs an iterative matching framework that effectively combines the advantages of the contextual information and an artificial neural network. The proposed method can correctly aggregate one‐to‐many (1:N) and many‐to‐many (M:N) potential matching pairs using contextual information in the presence of positional discrepancies and a high spatial distribution density. This method iteratively detects new landmark pairs (matched pairs), based on the prior landmark pairs as references, until all landmark pairs are obtained. Our approach has been experimentally validated using two topographic datasets at 1:50 and 1:10k. It outperformed a method based on a back‐propagation neural network. The precision increased by 4.5% and the recall increased by 21.6%, respectively.     

基于云理论、粗集和模糊神经网络的区域橡胶种植适宜度评估模型

针对橡胶种植适宜性评估,基于云理论、粗集理论和模糊神经网络理论,提出了一种适宜度评估模型。该模型将转化的样本数据进行粗集简约,通过模糊神经网络得出评价因子的隶属函数,计算评价等级。研究结果表明,此模型能够科学、快速、准确地分析出橡胶种植最适宜区、适宜区、次适宜区和不适宜区。     

Converting local spectral and spatial information from a priori classifiers into contextual knowledge for impervious surface classification

A classification model was demonstrated that explored spectral and spatial contextual information from previously classified neighbors to improve classification of remaining unclassified pixels. The classification was composed by two major steps, the a priori and the a posteriori classifications. The a priori algorithm classified the less difficult image portion. The a posteriori classifier operated on the more challenging image parts and strived to enhance accuracy by converting classified information from the a priori process into specific knowledge. The novelty of this work relies on the substitution of image-wide information with local spectral representations and spatial correlations, in essence classifying each pixel using exclusively neighboring behavior. Furthermore, the a posteriori classifier is a simple and intuitive algorithm, adjusted to perform in a localized setting for the task requirements. A 2001 and a 2006 Landsat scene from Central New York were used to assess the performance on an impervious classification task. The proposed method was compared with a back propagation neural network. Kappa statistic values in the corresponding applicable datasets increased from 18.67 to 24.05 for the 2006 scene, and from 22.92 to 35.76 for the 2001 scene classification, mostly correcting misclassifications between impervious and soil pixels. This finding suggests that simple classifiers have the ability to surpass complex classifiers through incorporation of partial results and an elegant multi-process framework.