Site Landscape Design Optimization Based on YOLOv13 and Multi-Task Bayesian Federated Learning (YOLOv13-HyperACE-BFL)

Current distributed processing of site landscape images suffers from uneven sample distribution and multi-task gradient conflicts, which can easily lead to poor detection performance in federated learning. To address this issue, this paper proposes the YOLOv13-HyperACE-BFL (BFL, Bayesian Federated Learning) method. This method applies the Hypergraph-based Adaptive Correlation Enhancement (HyperACE) module to construct an adaptive adjacency matrix from node similarities. A multi-head hypergraph attention mechanism is then used to update node features, capturing high-order spatial dependencies. Full-Pipeline Aggregation-and-Distribution (FullPAD) is also used to establish fully interconnected paths between features at different scales and align low-resolution features. Within the multi-task framework, latent variables are parameterized using a Gaussian distribution to assess inter-task uncertainty. Furthermore, a dynamic weight adjustment mechanism is incorporated to correct inter-task gradient conflicts and adjust the gradient weights of each task in real-time. In the federated phase, FedAvg-M with knowledge distillation ensures stable model aggregation and reduces communication costs. Experiments demonstrate that this method achieves 83.7% mean average precision (mAP) for large-scale object detection, 93.2% accuracy for site classification, and 82.3% boundary segmentation IoU (Intersection over Union) for multi-task recognition of site landscapes. This method provides high-precision scene understanding capabilities and a distributed collaboration foundation for site protection planning, restoration design, and digital display.