Precision Drone Landing Recognition: ResNext50 Transfer Learning Combined with Hybrid Ensemble Random Forest for Optimal Scene Classification

A vision-based landing structure to enhance Global Positioning System (GPS) performance during the final descent phase of an Unmanned Aerial Vehicle (UAV). Instead of relying solely on satellite signals, a camera would be placed on the drone in systematic order to obtain real-time photographs as the drone approaches the landing site. The images above are processed in order to locate a predefined important place which is followed to a target location. Once detected, the organization calculates the positional error between the centre of the marker and the image frame. That error is familiar to a proportional regulation scheme which continuously adjusts the drone's horizontal motion continuously during descent, thus minimizing the positional deviation. Compared with the standard GPS-only method, the vision-based strategy provides additional organic wisdom, especially in the case of signal impairment in alternately dense environments where satellite location might be unreliable. During testing, our model reached 91.31% classification accuracy and brought the average landing error down to just 2.34 centimeters. Although GPS is essential for navigation in general, it is not sufficient for more precise landing. The recommended solution will improve the reliability and safety of UAVs while remaining cost-effective and compatible with the existing UAV platforms.