基于地理知识迁移的城市增长精细化模拟——以珠江三角洲城市群为例

Spatial data mining techniques are widely used in urban growth modeling and approved as efficient tools in many applications. However, most of spatial data mining algorithms based on one important assumption: training data and test data have a same feature space and a same distribution. If the distribution of data is changed, we have to recollect new sample to train a new model. In many real-world applications, collecting samples is a time and labor consuming work. Due to the limitation of conditions, in certain specific research area or time period it is even impossible to collect any samples. If we can reduce the demand for new samples, make fully use of existing old samples and extract useful knowledge from similar geographic issues (or domain) to assist in the establishment of new geographical model, it will be very valuable and with practical significance..Spatio-temporal knowledge transfer is proposed in this context. For domain adaptation and lack of samples problem of urban sprawl simulation, propose a spatio-temporal knowledge transfer modeling framework. Based on this framework, establish urban land use transfer learning simulation model and apply to the simulation of major cities urban sprawl in Pearl River Delta..This study focus on the limitation that traditional methods cannot training a good model when during domain adaptation or lack of training sample situation. We analyze the feature distribution similarity among different domains, build spatio-temporal knowledge transfer models for monitoring and simulation of urban sprawl establish and apply to real-world applications.

多样化城市发展是未来我国城市增长的主要模式，城市增长过程受多种因素影响，是具有时空异质性的的复杂系统。现有模型通常仅能针对单一增长类型进行变化模拟，而且在数据不充足时，模型的表现效果不佳。在实际应用当中，很难采集到能够充分反映不同类型的精细化数据。本项目拟建立基于知识迁移的城市增长模型，从相似的城市增长领域中迁移样本数据和特征知识，构建多类型的城市增长精细化模拟模型。具体研究内容包括：1）基于宏观数据及社会感知数据的城市增长领域特征分析；2）探索城市增长特征差异对知识迁移的影响机制；3）构建基于知识迁移的CA-ABM协同模拟模型，实现研究区多类型城市增长的精细化模拟。本项目旨在弥补传统模型在跨领域建模方面的不足，采用知识迁移机制使得模型在相类似的研究问题之间共享数据和知识来辅助建模，实现多类型城市增长的精细化模拟。为政府和规划工作人员掌握多样化城市增长的动态过程、制定规则政策提供决策支持。

多样化城市发展是未来我国城市增长的主要模式，城市增长过程受多种因素影响，是具有时空异质性的的复杂系统。现有模型通常仅能针对单一增长类型进行变化模拟，而且在数据不充足时，模型的表现效果不佳。在实际应用当中，很难采集到能够充分反映不同增长类型的精细化数据。本项目拟建立基于知识迁移的城市增长模型，从相似的城市增长领域中迁移样本数据和特征知识，构建多类型的城市增长精细化模拟模型。取得的主要成果包括：1）构建了研究区的城市增长领域数据库；2）提出基于多时相遥感数据的知识迁移城市用地监测模型；3）提出基于众包平台及样本相似性的样本数据缺失值填充方法；4）提出基于集成学习的城市社会经济数据精细化制图方法；5）建立了城市扩张特征定量化分析及比较方法；6）提出基于一类算法的城市更新潜力评价模型；7）提出基于样本知识迁移的城市扩张模拟方法。项目成果能够采用知识迁移机制使得模型在相类似的研究问题之间共享数据和知识来辅助建模，实现多类型城市增长的精细化模拟。为政府和规划工作人员掌握多样化城市增长的动态过程、制定规则政策提供决策支持。