At present, studies of rural texture lack practical guiding significance for rural planning, resulting in a single form of new rural residential areas, causing serious damage to the original rural texture, moreover, there are many research methods for spatial planning, of which the method of machine learning has been very common in the design field, especially the ability to learn spatial features and the quantitative study of features. The method of using machine learning to influence decision-making in planning and design has not been deeply studied and applied in rural areas. The GANs (Generative Adversarial Networks) model is established in this paper by creating a four-level scale model of the boundary definition and factor-labeling paradigm of the rural space, clarifying the boundary, factor, and calculation form that affect the scale at all levels, and annotating the samples using the standard as a reference. Through the case study of a rural area in Shanghai, the labeling method was adjusted, various sample augmentation methods were continuously used to increase the sample size, the sample quality was adjusted by screening samples, and multiple pieces of training were finally generated to meet the needs of the new residential areas planning scheme in the given area. Through the establishment of area indicator and spacing indicator, the rationality of the scheme is quantitatively evaluated. The findings of this study show that machine learning approaches have a lot of potential for new settlement planning in rural areas; the GANs model is very effective in creating planning schemes, and this research method generates fresh ideas for rural area planning.