Transformer-Based Sunflower (Helianthus annuus L.) Recognition from Multi-Temporal UAV Orthomosaics

Abstract

The first appearance of inflorescence in sunflowers (Helianthus annuus L.) signifies the transition of the sunflower from the vegetative stage to the reproductive (R). At this growth period, accurate and automated detection of sunflower inflorescences is of utmost significance for sunflower yield estimation. Unmanned aerial vehicles (UAVs) have become essential in agricultural product detection due to their high spatial and temporal resolution data collection ability. With the rapid enhancements in deep learning, transformer architectures have emerged as a revolutionary paradigm, showing remarkable success in precision agriculture applications, including crop recognition and mapping. The main goal of this study is to investigate the potential of the DETection TRansformer (DETR) model in identifying sunflowers at the reproductive stage using multi-temporal UAV orthomosaics. To this end, orthomosaics were produced using high-resolution aerial photos collected with a DJI Phantom 4 Pro V2 UAV in a sunflower field located in Akyazı district of Sakarya province, during two reproductive periods of sunflower (R5.1 and R5.9). Utilizing the orthomosaics, two sunflower detection datasets were constructed to train and evaluate the model. The results revealed that the DETR performed better on the R5.9 growth stage (AP<inf>0.50</inf> = 92.40%, AR<inf>100</inf> = 68.00%) than the R5.1 (AP<inf>0.50</inf> = 83.70%, AR<inf>100</inf> = 53.90%). Furthermore, given increasing IoU thresholds, DETR demonstrated 16.4% and 29.8% improvements in AP and AP<inf>0.75</inf>, respectively, at the R5.9 stage. The results highlighted that DETR could be a powerful tool for identifying sunflowers, especially at advanced growth stages, likely due to more distinct and developed features of inflorescences. © 2025 Elsevier B.V., All rights reserved.