Efficient Large-scale Mapping of Acacia Tortilis Trees Using UAV-based Images and Transformer-based Semantic Segmentation Architectures

Abstract. The Acacia tortilis tree, locally known as Al Samr, is one of the native trees in arid and semi-arid ecosystems. This type of tree thrives in challenging climate conditions and considerably contributes to desert ecosystems. However, Acacia trees are increasingly vulnerable to land degradation, degradation, grazing, urbanization, and the demand for wood as a fuel source. Given the ecological significance of Acacia trees and their vulnerability to various environmental threats, current information on their distribution and population is essential for effectively conserving and managing this native species. This study aims to map Acacia trees from unmanned aerial vehicle (UAV)-based images using deep learning techniques. First, a comprehensive field campaign was conducted to record the locations of Acacia trees within the study area. Thereafter, the Segment Anything model was fine-tuned to delineate tree boundaries from the UAV data, facilitating the preparation of ground-truth labels. Subsequently, Mask2Former, a semantic segmentation architecture utilizing a dual-attention vision transformer backbone, was implemented to segment the Acacia trees. The performance of the proposed architecture was compared against those of Mask2Former models based on alternative architectures, including Swin Transformer, Grounded Language-Image Pre-training, and EVA02, a transformer-based visual representation pre-trained model. Results demonstrated that the proposed approach outperformed the evaluated models, efficiently delineating Acacia trees and achieving a mean intersection-over-union of 83.43% and a mean F-score of 90.27.%. The proposed approach offers valuable means for building tree inventories, updating geospatial databases, and promoting sustainable management of native Acacia trees.