Adaptive Residual Group Dual-Channel Pulse-Coupled Neural Network for Autonomous Drone Navigation in Complex Urban Wind Zones

Drones play a crucial role in civil engineering tasks such as logistics and data collection. While existing autonomous drone research focuses on safe path planning in static environments, urban navigation remains a challenge due to dynamic factors like wind disturbances, complex building layouts, and signal interference. Traditional autonomous navigation methods struggle in urban settings, where wind variations, GPS limitations, and unpredictable obstacles affect flight stability. Most existing approaches rely on extensive external data sources, increasing computational complexity. Therefore, an efficient, sensor-based solution is required to enable drones to navigate dynamic wind zones with real-time adaptability and high precision. This study introduces a Residual Group Adaptive Dual-Channel Pulse-Coupled Neural Network with Cat Hunting Optimization (RGADCPCNNet-CHO), which leverages local RGB camera and GPS data to enhance drone navigation without relying on external inputs. The framework integrates Window-Aware Guided Image Filtering (WGBF) to refine GPS-estimated wind force with RGB camera data, while a Cswin Transformer extracts features such as obstacles, wind drift, velocity, and position. The RGADCPCNNet handles path planning, and Cat Hunting Optimization (CHO) optimizes hyperparameters for better adaptability. The proposed model achieves 99.15% accuracy, 99.14% recalls, 0.99 explained variance ratio (EVA), and a low RMSE of 0.41, demonstrating superior performance over existing methods.