Objective Plantscape design is an interdisciplinary field blending art and science. Designers often face multiple challenges in the practical application of plantscape design: 1) Difficulty in summarizing design principles; 2) barriers for beginners amidst a high demand for and shortage of designers; 3) low efficiency in revisions; 4) prevalent manual labor, presenting significant opportunities for automation. The design of flower borders, representing a significant aspect of plantscape design, is suitable for case studies. In this context, machine learning offers new technical support. For patterns that are difficult to summarize, machine learning can rapidly process large volumes of experiential data through probability density estimation, thus reducing reliance on designers' intuition while significantly enhancing design efficiency and reducing labor intensity. Previous research confirmed the effectiveness of generative adversarial network (GAN) in layout generation. Research on plantscape generation based on GAN endeavors to rapidly learn from a large dataset of plant combinations with inherent patterns to generate high-quality plant layout designs that meet real-world application needs, thus assisting designers in quickly producing plant layouts at the initial design stages. Further integration of image processing and evaluative feedback can generate plant arrangement schemes that are both standard and artistic. GAN can provide exploratory tools and methods for the landscape design field, promoting innovation and development in the generation of plant arrangement scheme.
Methods The research proposes an experimental framework for a plantscape plan generation model based on GAN. The research design mainly includes three steps: Dataset preparation, model training, and model assessment. Initially, high-quality flower border plan images are collected from a single design firm, and these images are statistically categorized based on spatial structure types and design elements, thereby establishing principles for sample selection to ensure the professionalism and scientificity of the dataset adopted. Then, the selected flower border images undergo preprocessing, including plant classification, merging, image resizing, color tagging, and data augmentation. The cycle generative adversarial network (CycleGAN) algorithm is applied to build the flower border plantscape plan generation model on an open-source programming platform, undergoing multiple training rounds. Lastly, the research objectively analyzes and compares the spatial results of the generated flower border plans with actual design plans. In terms of subjective evaluations, expert scoring and other feedback methods are adopted to assess the generated plantscape images in terms of aesthetics and ecology.
Results The objective evaluation reveals that: 1) The model most accurately identifies the contours of rectangular sites, requiring additional optimization for certain curved sites; 2) The model can accurately learn and reflect about half of the tagged types in the dataset, with most generated color blocks closely matching the pre-determined colors in the dataset, demonstrating high color accuracy; 3) The model needs to enhance the diversity of its outputs when handling similar sites. In terms of spatial structure generation, 1) the unit area of generated results conforms to the application and volume characteristics of plants typically used in flower borders; 2) The results can replicate plant combinations frequently used in the dateset; 3) The results can learn the spatial distribution patterns of plants, replicating the shapes of edge plants, internal plant patches, and sculptural shrubs; 4) The results can display vertical variation. In subjective evaluations, rectangular flower borders show certain advantages in aesthetics and ecology, particularly in low maintenance and sustainability, but there is room for improvement in seasonal variation and texture coordination. Curved flower borders are slightly inferior in aesthetics but perform reasonably well in ecological coordination, needing deeper optimization in plant diversity and aesthetics.
Conclusion CycleGAN exhibits unique advantages in plantscape design represented by flower borders. Although the number of training samples is not extensive, the CycleGAN model performs well in expressing image quality and spatial layout, and demonstrates accurate boundary recognition for rectangular sites and high precision in color reproduction. The generated plan images mimic the dataset well. The spatial layout of the generated plans showcases the spatial distribution characteristics and visual effects of various plants, reproducing some potential combination patterns. The generated flower border designs align with actual design samples in terms of color, seasonal change, layering, and harmony. Ecologically, the generated designs embody sustainability principles, emphasizing the convenience of sustainable maintenance and management. However, the application of CycleGAN also demonstrates certain limitations. First, the aesthetic and ecological quality of GAN-generated flower border designs highly depends on the quality and diversity of the training dataset. Secondly, CycleGAN models tend to produce generic designs, lacking diversity in similar sites. In view of this, future research may focus on improving model algorithms, enriching the training dataset, and investigating the potential of introducing conditional generative adversarial network (CGAN) .
DownLoad: