Automatická detekce zlatobýlu na snímcích pořízených bezpilotním letadlem

Abstract

This bachelor's thesis deals with the automatic detection of Canadian goldenrod in imagery acquired by an unmanned aerial vehicle. Data were collected at a slag and ash disposal site in Dolní Líštná near Třinec using a DJI Mavic 3 Enterprise quadcopter at two flight altitudes — 22 m and 10 m. An orthophoto mosaic covering an area of 9.55 ha was produced from the 22 m flight imagery using Agisoft Metashape Professional. In ArcGIS Pro, 6,169 fully blooming goldenrod plants were annotated and six detection models were trained based on three deep learning architectures — Faster R-CNN, SSD, and YOLOv3. Each architecture was trained separately on data from both flight altitudes. Model evaluation was performed on an orthophoto mosaic excerpt with manually annotated reference data, and the results of individual architectures and datasets were compared. The best results were achieved by the Faster R-CNN model trained on imagery from the 22 m flight altitude, reaching AP value of 0.661 and an F1-score of 0.722 at IoU ≥ 0.5.