A Deep Dive Assessment into Recent YOLO Models Efficiency and Robustness for Agricultural Dense Scene Analysis: A Wheat Head Study Case

Abstract

Applications of artificial intelligence (AI) in agriculture are becoming more and more omnipresent. When applied to detecting problems in dense scene analysis in agricultural challenges, they offer answers to issues that previous metaheuristics were unable to address with a degree of accuracy that permits the industrialization of the suggested solutions. Therefore, counting wheat heads in field conditions with spike and spikelet occlusions and overlaps is one of the most difficult looked into and explored challenges in computer vision for agriculture in the post-pandemic period of awareness regarding staple crop security. In order to enable farmers, breeders, and agronomists to monitor the emergence and growth of wheat heads over time and space, as well as to extrapolate yield estimations, the ultimate goal is to provide a reliable and essential informational input among many other essentials. This paper proposes implementing YOLOv10, YOLOv11 and YOLOv12 models to GHWD, a global wheat head dataset compiled from various continents to ensure phenotype diversity, in order to address mentioned issues using new deep learning advancements. We then test the robustness of the obtained results on a self-compiled Tunisian wheat dataset. Finally, we compare and discuss the results.