INTEGRATING UAV PHOTOGRAMMETRY AND DEEP LEARNING FOR VEGETATION MONITORING ON THE DRIED BED OF THE ARAL SEA

Abstract

The dried bed of the Aral Sea is vast, hard to access and covered by sparse, mosaic vegetation that 10–30 m satellite pixels cannot resolve. We tested a workflow that combines UAV Structure-from-Motion photogrammetry, deep-learning vegetation detection, LiDAR-based 3D analysis and a continuously updated flora database. Two key sectors (1.01 km² and 0.524 km²) were surveyed with a DJI Matrice 30T, giving a ground sampling distance of 5.09 and 6.77 cm/pixel and reprojection errors of 0.847 and 0.796 pixels. A YOLOv5-based detector and a five-class CNN classifier gave preliminary active-vegetation cover of about 45% and 34%, concentrated in micro-depressions. Results are stored in a Darwin Core–compatible database published through the geoai.uz platform, which is used at least weekly by 66.5% of 287 surveyed users. A three-level monitoring system linking field plots, UAV surveys and satellite data is proposed.