Geospatial Deep Learning for Pomelo Tree Detection in Mixed Orchard Systems: A UAV-Based Assessment of YOLO Models

Abstract

This study addresses the global transition in agricultural area monitoring and assessment from traditional field survey methods to the widespread use of geospatial technology and artificial intelligence under the concepts of precision agriculture and digital agriculture. Accordingly, this study aims to apply unmanned aerial vehicle (UAV) imagery in combination with deep learning techniques to classify pomelo trees in mixed orchard systems, while comparatively analyzing the performance of YOLO models in order to identify the most suitable model for complex terrain. The findings are expected to support the development of a precise farm-plot-level spatial database, which is important for the management of geographical indications (GI) pomelo, yield estimation, and effective regional agricultural planning. The results indicate that YOLOv5 is more suitable for tasks requiring high accuracy in tree counting and spatial density analysis, whereas YOLOv8 demonstrates architectural stability and potential for further development in the future. Overall, the findings will facilitate the selection of an appropriate model for developing an accurate farm-plot-level spatial database, which is important for the management of various orchard systems, as well as for yield estimation and effective regional agricultural planning