【作者】 范晓；

【导师】 田淑芳；

【作者基本信息】 中国地质大学(北京) ， 资源与环境遥感， 2012， 硕士

【摘要】 随着社会发展和人口的过度膨胀，人类对土地利用的改造程度不断加深，土地类型变化加速。如何快速识别土地利用类型，有效地进行土地分类，成为土地资源研究的重点问题，而遥感技术的快速发展为土地利用监测提供了便利。多层感知器前馈神经网络（BP网络）是应用最广泛的神经网络模型之一。然而标准BP网络具有收敛速度缓慢、训练过程易陷入局部极小值等问题。本论文依托MATLAB软件平台，对标准BP网络在网络结构和学习算法等方面进行改进，并在土地利用信息提取中实现。论文取得的研究成果如下。1）论文针对标准BP网络的局限性，从以下几个方面进行了改进：①对BP神经网络结构进行深入研究，提出采用黄金分割法与遗传算法结合寻求最优隐含层节点个数，并且获得网络初始化权重参数；②针对标准BP算法运算过程中收敛缓慢和局部极小值的问题，论文将现有各种改进算法进行归纳总结，对比各种改进算法的运算效率和内存需求，并通过实验确定弹性算法为最优学习算法；③为保证BP网络能够有效的学习，对输入数据统一进行规则化处理，同时对BP网络分类输出数据进行去模糊化处理；④探讨了影响BP网络泛化能力的因素和提高泛化能力的方法。2）通过对标准BP网络在网络结构、学习算法、泛化能力、数据处理等方面的分析和改进，提出了改进型BP神经网络在土地利用分类中的流程，并以RapidEye为遥感数据源，以北京市密云县露天采矿区为研究区域进行土地利用分类，取得了较好的效果。3）为了全面评价分类结果，论文将改进型BP网络与标准BP网络、极大似然法分类结果进行对比，从分类精度、分类效率和分类图美观度三个方面分别进行定量评价。结果证明改进后的BP算法无论在分类精度、分类效率和分类图的美观可读性方面都具有优势，可以作为一个具有潜力的分类方法继续应用和研究。更多还原

【Abstract】 With social development and population expansion, the land use reform isbecoming deeper and the types of land use are rapidly changed, therefore how fastand efficiently do we identify and classify the types of land use is the hot spot in landmanagement research. Fortunately, the rapid development of remote sensingtechnology provides a convenience.Back propagation neural network (BP network in short) is one of the most widelyused neural networks. Nevertheless, standard BP network has some drawbacks likeslow convergence speed and tendency towards converge to local minimum, etc. Basedo the MATLAB software, this thesis puts forward a new BP network by improving thestandard BP network in the respects of network structure, learning algorithm, dataprocessing and generalization ability, and then conducts land use classification inMiyun mining area, Beijing. The research includes:1) Some improvements are made according to standard BP network:①Aftermakingdeep research on the structure of BP network, the thesis combines golden sectionmethod with genetic algorithm to obtain the node number of hidden layer andinitial connection weight;②Determine the optimal learning algorithm bydiscussing some frequently used learning algorithms in terms of operationalefficiency and memory requirement;③Normalizethe input data to ensure BPnetwork learning effectively, and de-fuzzy the output data.④Discuss both thefactors that influence the BP network generalization ability and the methods whichcan improve it.2) This thesis conducts land cover classification after preprocessing the RapidEyeremote sensing data in opening mining area of Miyun county, Beijing using theimproved BP network and achieves good results.3) In order to evaluate the classification results comprehensively, this thesiscompares results of improved BP network with those of maximum likelihoodmethod and standard BP network from three perspectives, which are classificationaccuracy, classification efficiency and appearance of classification maps. Theresults prove that the improved BP network proposed in this thesis has more advantages than the other two methods, and can be used as a potentialclassification method in Information Extraction and land use classification.更多还原