为探究城市扩展规律，模拟城市的空间变化趋势，对于客观判断城市未来发展、制定合理且可行的土地政策具有重要意义。近年来，元胞自动机作为一种研究复杂系统随时空变化的典型方法，虽已有广泛的地理学应用，但仍存在转换规则难以自动获取、时间难以界定等具体问题。人工神经网络适宜于解决复杂非线性问题，在城市系统建模问题上较线性回归方法具有优越性，将其引入元胞自动机模型可以较好地解决在城市扩展模拟中获取转换规则的问题。因此，本文提出应用基于神经网络的元胞自动机模型来模拟和预测城市扩展。论文内容包括：①综述了元胞自动机和神经网络在国内外城市系统应用研究的研究进展和面临的问题；②介绍了基于CA和ANN模拟城市扩展的理论基础；③基于MATLAB平台，耦合ANN和CA，构建了CA-ANN-URBAN模型，该模型包括分为数据预处理、空间变量选取、ANN参数设置、样本采集、神经网络训练、模型校正、模型模拟和预测等主要流程；④利用LCM-CT数据验证模型并通过实验分全图与子区测试了相关参数，分全图与子区重建了1986年至2000年的扩展过程，比较了模型内不同迭代时间间隔设置方案的模拟精度；⑤利用遥感技术提取北京市朝阳区的城市扩展变化信息，并利用GIS技术分析研究区域土地利用变化特征，在此基础上应用该模型重建了朝阳区1991年至2007年城市扩展过程，并预测了该区域2008年至2020年的扩展情况，讨论了模拟结果与实际分类图的空间差异及可能的原因。论文结论包括：①通过实验发现在采样样本固定的前提下，神经网络结构中隐层神经元的数量变化对于模型精度变化没有实质的影响；②通过LCM-CT数据两种区域（全图与子区）的对比实验，发现无论对逐年还是等间隔方案，子区的模拟精度指标值均高于全图，表明模型对于城市较为集中的区域更有效，这一结论也在朝阳区得到验证；③对于LCM-CT和朝阳区两个不同数据集的模拟结果说明，模型具有较好的适应性，证明了本文提出的CA-ANN-URBAN是一种有效的城市扩展模拟方法。

Exploration the spatial pattern, simulation and prediction of urban expansion are critical for reasonable and workable future design and land policy. In recent years, Cellular Automata (CA), as one of the typical methods studying complex systems with the temporal and spatial variation, has been more and more widely applied in the analysis of geographical phenomena. However, there still exist difficulties to automatically obtain the rules of conversion and define of time step. Artificial Neural Network (ANN) shows great advantages in solving highly complex nonlinear problems. It can simulate better than the linear regression method when confronting complex urban system. This advantages help to solve the problem of obtaining conversion rules in CA model. Therefore, in this thesis, we propose an ANN-based CA model to simulate and predict urban growth. We first review present researches in CA model in China and around the world, as well as the problems in using ANN to obtain CA model parameters. Secondly, we introduced the theory of Cellular Automata in urban system and Neural Network in obtaining CA conversion rules. Thirdly, we establish the CA-ANN-URBAN model under MATLAB environment. This model have seven principal parts, that is dataset pre-processing, selection of dataset, parameter setting up, sampling for training, training of neural network, calibration of model and simulation and prediction of model. Subsequently, we compare the result of our model and ArcGIS LCM module by testing its example dataset CT in year 1994, which validates our model. Thereafter, we apply the model to reconstruct the expansion process from 1986 to 2000 for both the entire region and sub area of CT. Then, simulation accuracies with different time intervals are discussed, so as to provide more scientific and reliable conclusions. Finally, we utilize CA-ANN-URBAN model to the Chaoyang District, which located in the eastern part of Beijing urban fringe area. In detail, we extract the maps of Chaoyang land use from remotely sensed imagery (1991, 2001 and 2007) and analyze urban land change with GIS spatial analysis techniques. We reconstruct the urban growth from 1991 to 2007, and then predict its possible expansion maps from 2008 to 2020. By comparing the simulation results with the actual classification, spatial differences and possible reasons are given.In short, main conclusions are: ①Under the premise of fixed samples, experiments prove that the accuracy of simulation in model do not sensitive to changes of hidden layer neuron number in neural network structure. ②Through experiments on the entire region and sub area of CT data, we find that for interval of one year and of fix interval years, the simulation accuracy in sub-area are always higher than the entire region. This indicates that the model is more effective when focuses on compact urban development pattern. In addition, test of Chaoyang District confirms this discipline. ③Examples of both LCM-CT data and Chaoyang District of Beijing city demonstrate the CA-ANN-URBAN model has a good adaptability. The simulation results manifest the model for the simulation and prediction of urban expansion is practicable and effective.