Detection of palm tree from high-resolution UAV images using deep learning technique

The palm tree is a significant contributor to Saudi Arabia’s Gross Domestic Product (GDP) in the agricultural sector. Much energy is required to manage palm-tree plantations. An audit process is required to keep track of all palm trees in a plantation that uses irrigation and fertilizer and plans to grow more trees. An effective and efficient method is required because of the manual nature of the audits. A key component of good precision palm tree field management is the precise mapping and counting of individual palm crowns using remote-sensing images. This paper presents a method for detecting, mapping, and quantifying palm crowns using high-resolution Unmanned Aerial Vehicles (UAVs) imagery using a Deep Learning (DL) instance-segmentation approach (Mask Region Convolutional Neural Networks; Mask R-CNN). Before being split into training and validation sets, the acquired high-resolution images were cropped and sampled into smaller tiles. Model performance was evaluated on three sites (S1, S2, and S3). For these sites, the model precision ranged from 92.74 to 97.22%, recall ranged from 90.91 to 96.67%, accuracy ranged from 86.32 to 94.27%, and the F1-scores were of 0.92 to 0.97. Beyond reporting accuracy, the study documents a practical processing recipe (tiling/overlap, confidence filtering, and duplicate removal) and uses field-verified, manually delineated crown polygons as ground truth, with independent-site testing to assess transferability under similar acquisition conditions. The results show that the approach can detect palm crowns with high accuracy in the tested areas and quantify their number from UAVs imagery under comparable conditions; additional validation across denser or less-structured plantations, seasons, and illumination conditions is needed for broader generalization.