【Objective】Taking the distribution of cultivated land in central Yunnan urban agglomeration as an example, the present research analyzed the spatial multi-scale cultivated land models to explore the appropriate land use simulation model. 【Method】From 1,5,10,20 and 30 kilometers, five grid scales and 16 driving-force factors, goodness-of-fit, residual spatial autocorrelation, factor quantity and space scales, such as indicators, compared the diversity and precision of four kinds of arable land models which were classic regression, spatial lag, spatial error and geographically weighted regression. 【Result】Both cultivated land and its driving factors had positive global spatial autocorrelation at multiple spatial scales, and had spatial non-stationarity. The smaller the weight distance, the higher the spatial dependence degree, and it showed strong spatial heterogeneity with the increase of weight distance. The aggregation characteristics of cultivated land showed a high consistency with the change of spatial scales, but there were some differences in the specific spatial locations and the degree of detail expression. The smaller the scale, the more obvious the agglomeration was, and the more significant the local spatial difference was. Selection of goodness of fit, since the model of residual space relevance, indicators such as impact factor quantity and space scales, comparison and analysis the spatial multi-scale models, it was concluded that characterizing local spatial patterns and distribution of cultivated land order model and sequence was geographically weighted regression model, spatial error model, spatial lag model, regression model. 【Conclusion】Geographical weighted regression model takes into account the local characteristics and differences of geographical space, which can characterize the local characteristics of cultivated land space in more detail, determine the goodness of fit degree of different spatial location models, and obtain the weight of each driving factor of land use with spatial location. It has good model construction ability in terms of local spatial characteristics, simulation accuracy and driving factor coefficient, and can be combined with other mathematical models to build a land use prediction simulation model that is suitable for the scale of plateau and mountain urban agglomeration. [ABSTRACT FROM AUTHOR]