ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING POTENTIAL AS TOOLS FOR GEOREFERENCING AND FEATURES DETECTION USING UAV IMAGERY

Abstract

While Satellite imagery holds the advantage of encompassing expansive geographical regions. spanning square kilometers, its Spatial Resolution (SR) might prove inadequate for specific tasks. Conversely, Unmanned Aerial Vehicles (UAVs) excel in capturing high-resolution images with spatial resolutions in the range of a few centimeters or even millimeters. However, the accuracy of sensor locations during UAV flights is non-accurate enough by the Global Navigation Satellite Systems (GNSS) technology onboard. One of the key objectives of this research is to evaluate a technique aimed at generating precise sensor locations. This technique employs raw data from the drone's GNSS receiver and minimum Ground Control Points (GCPs) placed within a 2-meter diameter circle in the study area. The goal is to achieve accurate Digital Elevation Models (DEM) and orthomosaic images. Another focus of this research is on addressing challenges related to road lane detection. This is achieved through the enhancement of the You Only Look Once (YOLO) v3 algorithm. The proposed approach optimizes grid division, detection scales, and network architecture to enhance accuracy and real-time performance. The experimental results showcase an impressive 92.03% accuracy with a processing speed of 48 frames per second (fps), surpassing the performance of the original YOLOv3. In the rapidly evolving landscape of Artificial Intelligence (AI) and drone technology, this investigation underscores both the potential and complexities inherent in utilizing advanced AI models, such as YOLOv8, for building detection using UAV and satellite imagery. Furthermore, the research delves into robustness and real-time capabilities within building detection algorithms. The outlined strategy encompasses precise pre-processing, Field-Programmable Gate Array (FPGA) validation, and algorithm refinement. This comprehensive framework aims to elevate feature detection in intricate scenarios, ensuring accuracy, real-time efficiency, and adaptability.