A Comparative Analysis of Deep Learning Methods for Ghaf Tree Detection and Segmentation from UAV-based Images

Abstract

The Prosopis cineraria, commonly known as the Ghaf tree, is an ecologically significant species that prevents desertification, enhances soil fertility, and supports biodiversity within arid ecosystems. Mapping and monitoring Ghaf trees using unmanned aerial systems (UAVs) and deep learning are essential for advancing conservation efforts through reliable, automated assessments. In this study, we performed a comparative analysis of several transformer-based deep learning models, including Mask DETR with Improved Denoising Anchor Boxes (Mask DINO) and Mask R-CNN models based on the Vision Transformer (ViT), Swin Transformer, and Enhanced Multiscale ViT (MViTv2), for mapping Ghaf trees from UAV images captured in diverse urban and agricultural environments. Results demonstrated strong potential for the assessed instance segmentation architectures in mapping Ghaf trees, achieving mean average precision values of 80% to 84.2% for detection and 82.2% to 85.1% for segmentation, with F1-scores ranging from 83.55% to 88.3% for detection and 85.5% to 88.6% for segmentation. This study underscores the effectiveness of transformer-based deep learning architectures for mapping Ghaf trees from UAV images, with findings and refinements that can be applied and extended to map other native tree species and support broader conservation initiatives.