A Comparative Study for Wheat Head Detection Through Testing the Robustness of Two Global Dataset Trained YOLO Models on a Tunisian Wheat Dataset

Artificial intelligence (AI) applications are getting more and more ubiquitous in the field of agriculture. They provide responses to problems that old metaheuristics were unable to respond to, with a level of precision allowing the industrialization of the proposed solutions when applied to detection problems in dense scene analysis in agricultural challenges. Hence, in the era of post-pandemic awareness towards staple crop security, one of the most challenging problems studied in computer vision for agriculture is the wheat head counting in field conditions with spike and spikelet occlusions and overlaps. The final target is to offer a solid key informational input between many necessary others, to make farmers, breeders and agronomists able to estimate the wheat head density and supervise their emergence and growth over time and space allowing them to interact on time in the process and to have an idea about yield estimations. To address these issues using new deep learning advances, this paper proposes the implementation of YOLOv7 and YOLOv8 models to the GHWD-2021, a global wheat head dataset compiled from different continents to guarantee phenotypes diversity, then we test the robustness of the results on a self compiled Tunisian wheat dataset. Finally, we compare and discuss the results.