元学习算法的小样本人口空间化模拟方法  

Spatial simulation method of few-shot population based on meta-learning algorithm

作　　者：李娜 吴怀石 LI Na;WU Huaishi(School of Economics and Management,Tianjin Chengjian University,Tianjin 300384,China)

机构地区：[1]天津城建大学经济与管理学院,天津300384

出　　处：《测绘科学》2024年第12期192-203,共12页Science of Surveying and Mapping

基　　金：天津市哲学社会科学规划项目(TJGL23-004)。

摘　　要：针对城市精细化管理建设重要研究内容的人口空间化模拟,存在人口非对称性映射和小样本的问题,该文提出了基于元学习算法的组合元模型。以多源地理大数据为基础,元学习一种适用于人口小样本预测的中间表征,通过梯度迭代次数优化使其快速适应不同人口分布的复杂情况,从而统计精细尺度下的人口数据。实验结果表明:多种机器学习算法中元模型的训练效果最佳,R^(2)为0.95,相较于其他模型平均提升8%左右。在500 m格网尺度上的模拟结果显示,天津市人口总体呈“南密北疏”的分布特征,同时元模型在行政区级尺度的模拟精度达到97%,在街道级尺度的模拟精度达到94%。此外,为验证模型的整体有效性,通过与公开人口数据集的对比和消融实验,展现了元模型在精细人口空间化场景中的精度优势。This paper proposes a combined meta-model based on meta-learning algorithm for simulating population spatialization,which is an important research topic in the construction of refined urban management.The problem lies in the asymmetric mapping of population and few-shot.Based on multi-source geographic big data,meta-learning is an intermediate representation suitable for predicting few-shot population.It optimizes the gradient iteration times to quickly adapt to complex situations with different population distributions,thereby statistically analyzing population data at a fine scale.The experimental results show that the meta-model has the best training effect among various machine learning algorithms,with an R~2 of 0.95,which is an average improvement of about 8%compared to other models.The simulation results on a 500 meter grid scale show that the population of Tianjin is generally distributed in a“dense in the south and sparse in the north”pattern.At the same time,the simulation accuracy of the meta-model reaches 97%at the administrative level scale and 94%at the street level scale.In addition,to verify the overall effectiveness of the model,the accuracy advantage of the meta-model in fine-grained population spatialization scenarios was demonstrated through comparison and ablation experiments with publicly available population datasets.

关 键 词：多源地理大数据 元学习算法 元模型 人口空间化 小样本 天津市 

分 类 号：P208.2[天文地球—地图制图学与地理信息工程]