Regional-scale Acacia tortilis crown mapping from UAV remote sensing using semi-automated annotation and a lightweight hybrid segmentation framework

• Introduced a semi-automated annotation workflow for efficient large-scale UAV labeling. • Proposed a lightweight hybrid segmentation framework integrating Mamba, Transformer, and CNN. • Achieved a superior balance between segmentation accuracy and computational efficiency. • Demonstrated strong performance and generalizability across diverse landscapes. • Enabled scalable regional mapping and monitoring of Acacia tortilis from UAV images. Acacia tortilis (Umbrella thorn acacia), native to the dry and semi-arid zones of Africa and the Middle East, can withstand extreme climatic conditions and provides substantial ecological and socioeconomic benefits. Reliable, regional-scale mapping and monitoring are increasingly critical as populations face mounting environmental and anthropogenic pressures. This study presents an efficient and lightweight deep learning framework for the regional-scale delineation of A. tortilis from ultra-high spatial resolution unmanned aerial vehicle (UAV) images. A comprehensive field campaign was conducted to collect ground-truth samples across diverse urban, agricultural, and mountainous settings. Aiming to expedite the labor-intensive labeling process and generate consistent, reliable training data, a semi-automated annotation workflow was developed, expanding the annotated dataset nearly fourfold to approximately 36,800 A. tortilis instances. The study also introduces a hybrid U-shaped Mamba-transformer architecture, designed to maximize the advantages of Mamba units, Transformers, and convolutional networks for joint modeling of local detail and extensive contextual information. The proposed framework was evaluated against several representative CNN–, Transformer-, and Mamba-based models, as well as alternative decoder designs, to assess segmentation quality, efficiency, and generalizability. Using an independent testing dataset, the hybrid model family (tiny, small, and base variants) demonstrated a strong balance between accuracy and efficiency, achieving mean intersection over union values ranging from 85.30% to 85.44% and mean F-scores between 91.52% and 91.61%, maintaining consistent performance across geographically distinct regions. The lightweight CNN-based decoder offered an optimal balance between accuracy and computational cost, with substantially reduced training time compared to more complex decoders. Across an extensive regional survey extent (∼250 km 2 ), the framework consistently delineated A. tortilis crowns, underscoring robust and scalable performance. Thus, the proposed framework can facilitate systematic monitoring of A. tortilis and support the conservation, restoration, and sustainable management of native tree species across arid and semi-arid ecosystems.