Automated landscape element recognition and layout optimization based on image segmentation and object detection

Urban landscape planning faces growing complexity that demands more effective tools for element recognition and spatial layout optimization. This paper presents an automated framework integrating image segmentation and object detection to streamline the identification and design of landscape elements. We develop a multi-scale feature fusion segmentation network with domain-adapted attention modules and a boundary-aware loss, achieving 81.2% mean Intersection over Union (mIoU), a gain of + 5.4% over the DeepLab v3 + baseline. A dual-branch architecture that jointly performs object detection and instance segmentation reaches 75.3% mean Average Precision (mAP) for element localization and classification. Building on these recognition outputs, a hybrid genetic algorithm–particle swarm optimization strategy under spatial relationship constraints yields layout improvements of 18.4% to 31.2% across quality metrics. All models are trained on a purpose-built dataset of 17,077 annotated landscape images spanning ten categories, using a ResNet-based encoder with PyTorch on NVIDIA V100 GPUs under 5-fold geographically stratified cross-validation. Ablation studies confirm the contribution of each proposed module. Expert evaluation by certified landscape architects using blind scoring rubrics (Likert 1–10 scale, n = 35) and structured user satisfaction surveys (n = 120, five-point Likert items) yield scores above 8.0 and satisfaction rates above 82%, respectively, both with statistically significant improvements over manual baselines (p < 0.01). The framework addresses persistent limitations of manual landscape design—high labor costs, subjective inconsistency, and poor scalability—and establishes a practical foundation for intelligent urban planning.