Multi-modal deep learning for intelligent landscape design generation: A novel CBS3-LandGen model

Area of Science:

• Computer Science
• Artificial Intelligence
• Landscape Architecture

Background:

• Urbanization presents complex challenges for traditional landscape design methods, necessitating more efficient, precise, and sustainable approaches.
• Manual landscape design processes struggle to meet the demands of modern urban development in terms of speed and accuracy.

Purpose of the Study:

• To propose an intelligent landscape design generation model, CBS3-LandGen, leveraging multimodal deep learning to address the limitations of traditional methods.
• To enable the generation of landscape plans that meet design objectives within resource constraints by integrating diverse data types.

Main Methods:

• The CBS3-LandGen model integrates image data processed by ConvNeXt, text data analyzed by BART, and generation optimization via StyleGAN3.
• A multimodal deep learning architecture was employed to fuse image and text information for landscape plan generation.
• The model was trained and validated using the DeepGlobe and COCO datasets.

Main Results:

• CBS3-LandGen demonstrated strong performance in image generation quality, text consistency, and multimodal data fusion.
• Quantitative results include a Frechet Inception Distance (FID) of 25.5 and an Inception Score (IS) of 4.3 on the DeepGlobe dataset.
• On the COCO dataset, the model achieved an FID of 30.2 and an IS of 4.0, indicating superior generation capabilities.

Conclusions:

• The proposed CBS3-LandGen model offers a novel approach to intelligent landscape design, effectively integrating deep learning technologies.
• The model's performance highlights its potential for improving efficiency, diversity, and multimodal data fusion in landscape generation tasks.
• Future research will focus on optimizing performance, enhancing training efficiency, and expanding applications in urban planning and ecological conservation.