融合多尺度与坐标注意力的城市扩张模拟  

Urban Sprawl Simulation Integrating Multiscale and Coordinate Attention

作　　者：孙令博 刘明皓[1,2] 罗庆喜 许汀汀 陈春 SUN Lingbo;LIU Minghao;LUO Qingxi;XU Tingting;CHEN Chun(School of Computer Science and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Spatial Information Research Center,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;School of Software Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;School of Architecture and Urban Planning,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065 [2]重庆邮电大学空间信息研究中心,重庆400065 [3]重庆邮电大学软件工程学院,重庆400065 [4]重庆交通大学建筑与城市规划学院,重庆400074

出　　处：《西南大学学报（自然科学版）》2025年第2期145-159,共15页Journal of Southwest University(Natural Science Edition)

基　　金：国家自然科学基金项目(42071218);重庆市博士直通车项目(CSTB2022BSXM-JCX0147)。

摘　　要：针对基于机器学习的元胞自动机在土地覆被变化模拟中存在的尺度效应和非平稳性特征提取不充分等问题,构建了ASPP(空洞空间金字塔池化)-CRA(坐标注意力)Unet-CARS(基于多类随机斑块种子)耦合模型。以成渝地区双城经济圈2012、2016、2020年实际城市土地利用变化数据为例,设计2组实验验证了模型的性能,并将其应用于预测2024年及2028年的城市扩张模式。通过模型对比结果显示,ASPP-CRAUnet-CARS模型的Kappa值为0.9123,FoM值为0.4142,Kappa值分别比RF-CMCNN-CA模型和UMCNN-CA模型的高出0.0208和0.0342,FoM值则分别提升了0.0306和0.0679。消融实验表明:去除ASPP和CRA模块后Kappa值与FoM值均有所下降。研究结果表明:ASPP-CRAUnet-CARS模型融合了传统元胞自动机和深度学习模型的双重优势,能较好地学习到城市发展中的复杂空间特征,改善了空间非平稳性建模效果,有效提高了模拟精度。In response to the issues like scale effects and insufficient feature extraction of non-stationarity in land cover change simulation based on machine learning-driven cellular automata,an ASPP(Atrous Spatial Pyramid Pooling)-CRA(Coordinate Attention)Unet-CARS(Cellular Automata for Raster Spaces)coupled model was constructed.Using real urban land use change data from the Chengdu-Chongqing economic circle in 2012,2016,and 2020,two sets of experiments were designed to validate the model s performance.It was then applied to predict urban expansion patterns of 2024 and 2028.Model comparison results demonstrated that the ASPP-CRAUnet-CARS model achieved Kappa value of 0.9123 and FoM value of 0.4142,outperforming RF-CMCNN-CA and UMCNN-CA model in Kappa by 0.0208 and 0.0342,respectively,and in FoM by 0.0306 and 0.0679,respectively.Ablation studies revealed that removing the ASPP and CRA modules resulted in decreased Kappa and FoM values.The study suggests that the ASPP-CRAUnet-CARS model,integrating the advantages of traditional cellular automata and deep learning models,can effectively learn complex spatial features in urban development,improve the modeling of spatial non-stationarity,and enhance simulation accuracy.

关 键 词：ASPP-CRAUnet-CARS模型 多尺度特征 注意力机制 空间非平稳性 

分 类 号：F301.2[经济管理—产业经济]