Centroid-Guided YOLO for Plant Stem Detection in Digital Agriculture

This study focuses on the utilization of a YOLO (You Only Look Once)-based object detection approach, along with the development of a customized algorithm, to address the identification and localization of invasive plant stems within agricultural crops. The initial phase involved the creation of a dataset comprising 485 manually annotated images, sourced from three distinct data repositories. Subsequent to this, the YOLO model underwent training to accurately recognize the plants depicted in the images. Following the model training, an algorithm was devised that leverages the centroid of the G vector in the RGB image as the designated point of interest for precise stem localization. The algorithm's effectiveness was assessed utilizing key metrics such as mean absolute error (MAE) and root mean square error (RMSE). For images sized at 600×600 pixels, the resulting MAE and RMSE values were calculated at 34.4 pixels and 46.8 pixels respectively. Notably, these outcomes underscored the necessity for refining the detection process, particularly in instances involving plants with delicate branches, where differentiation between the branch and the image back drop proved challenging. In terms of practical application within agrochemical dispersion systems, it was discerned that dispersion takes place across an application area rather than a singular point. By incorporating the proposed algorithm in conjunction with a dispersion cone aperture measuring <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$100\times 100$</tex> pixels, it was projected that the dispersion system could effectively cover approximately 92.4% of the existing plants. This algorithmic approach exhibits the potential to curtail agrochemical dispersion by a substantial factor of up to 36 times, yielding economic and sustainability advantages through the reduction of waste and environmental contamination.