基于双变量分区的KMeans++GConViTGCA建模方法及土地覆被变化模拟  

KMeans++-ConViT-CA modeling method based on bivariate zoning and its application in simulating land cover change

作　　者：王恩铭 刘明皓[1] 王琨 代俊 许汀汀 WANG Enming;LIU Minghao;WANG Kun;DAI Jun;XU Tingting(School of Computer Science and Technology,Chongqing University of Posts and Telecommunications/Chongqing Engineering Research Center for Spatial Big Data Intelligent Technology,Chongqing 400065;School of Software Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学计算机科学与技术学院/空间大数据智能技术重庆市工程研究中心,重庆400065 [2]重庆邮电大学软件学院,重庆400065

出　　处：《地理与地理信息科学》2025年第2期31-39,共9页Geography and Geo-Information Science

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：[JP+2]分区建模和多尺度信息挖掘是当前解决空间异质性建模的主要途径。该文针对现有模型对空间异质性挖掘不充分的问题,提出一种耦合分区与多尺度信息挖掘的KMeans++-ConViT-CA建模方法,在综合考虑因变量地类的邻域元胞活跃度和自变量驱动因子的空间相似度并采用KMeans++聚类算法进行分区基础上,应用ConViT模型挖掘土地适宜性的尺度信息,并以重庆都市圈为例从分区效应、尺度效应、模型效应3个角度进行验证。研究结果表明:①该模型Kappa系数达0.8362,FoM达0.4178,比未分区的ConViT-CA模型分别提升0.0541、0.0507,比单独基于因变量和自变量的分区模型分别提升0.0221、0.0291和0.0381、0.0442(分区效应);与挖掘转换规则时只关注局部特征的双变量单一尺度的KMeans++-CNN-CA模型及KMeans++-ViT-CA模型相比分别提升0.0294、0.035和0.0206、0.0244(尺度效应),比其他基于深度学习的模型分别提升0.0884、0.0826(模型效应)。②分区方法对建模精度影响较大,考虑邻域元胞活跃度分区的双变量建模方法能在一定程度上弥补单变量建模方法对空间异质性规律学习不足的问题,模拟精度更高。③分区与多尺度模型并非简单的替代关系,通过模型耦合可发挥该模型在区域尺度和像元尺度的综合优势,从而改善建模效果,提高土地覆被变化模拟精度。In the current research,zoning-based modeling and multi-scale information mining stand as the primary methods for addressing the challenges of spatial heterogeneity modeling.Taking Chongqing metropolitan area as a case study,this paper confronts the issue of inadequate exploration of spatial heterogeneity in existing models and proposes a modeling method named KMeans++-ConViT-CA,which integrates zoning with multi-scale information mining.Specifically,the method commences with a comprehensive consideration of bivariate factors,namely the neighborhood activity of land cover types(the dependent variable)and the spatial similarity of driving factors(the independent variables).Subsequently,the KMeans++clustering algorithm is employed as the basis for zoning.Afterward,the ConViT model is utilized to extract the scale-related information of land-use suitability.The validation process is conducted from three distinct perspectives:the zoning effect,the scale effect,and the model effect.The research findings are as follows.①The proposed KMeans++-ConViT-CA model demonstrates remarkable performance,with a Kappa coefficient reaching 0.8362 and a FoM of 0.4178.When compared with the unzoned ConViT-CA model,the Kappa coefficient and FoM exhibit increases of 0.0541 and 0.0507 respectively.In contrast to the zoning models based solely on the dependent variable or the independent variable,the improvements in Kappa and FoM are 0.0221 and 0.0291,as well as 0.0381 and 0.0442 respectively,highlighting the zoning effect.When compared with the KMeans++-CNN-CA model and the KMeans++-ViT-CA model with bivariate single-scale,which focus only on local features during the mining of transformation rules,the increments in Kappa and FoM are 0.0294 and 0.035,and 0.0206 and 0.0244 respectively,signifying the scale effect.Moreover,compared with other deep-learning-based models,the enhancements in Kappa and FoM are 0.0884 and 0.0826 respectively,underscoring the model effect.②The choice of zoning method exerts a substantial influence on the mod

关 键 词：土地覆被变化模拟 深度学习 分区建模 空间异质性 

分 类 号：P208.2[天文地球—地图制图学与地理信息工程] TP391.4[天文地球—测绘科学与技术]