Deep Ensemble Learning for Weed Recognition Using YOLO Variants in Precision Agriculture

Abstract

Weed competition in paddy fields is a critical challenge to crop yield and agricultural sustainability because conventional control methods, whether Broad-spectrum weeding or manual control herbicides, are Work-intensive, environmentally harmful, lack the precision needed for modern farming. To overcome the limitations mentioned above, then we suggest a new deep ensemble building blocks for learning that integrates 3-YOLO variants (YOLOv8 with a high detection accuracy, YOLOv11 with real-time processing capability, and GE-YOLO with good generalization) into one unified detection system optimized for automated weed recognition tasks in a paddy field. For each model, the ensemble architecture employed NMS and WBF strategies to combine predictions and hence improve the bounding boxes' localization accuracy. In this work, we built and manually annotated a paddy field dataset and further augmented the images using rotation, flipping, scaling, and brightness adjustments to represent various real-world scenarios that include changing lighting conditions, high-density crops, water reflections, and rice-weed morphological similarities. Each variation was independently trained across 100 epochs.  using the SGD optimizer by tuning its hyperparameters to achieve maximum performance in weed detection. The results achieved by the proposed deep ensemble framework were outstanding: it obtained an accuracy of 98.75%, precision of 88.89%, Consider of 99.35%, an F1-score of 93.42%,  mAP@50 of 88.89%. These results are better than any previous YOLO models and any recent literature on the best weed detection results. This shows that our proposed system can work well in a variety of difficult agricultural situations, such as when the light changes, there are a lot of weeds, and the background is complicated. These results show that the proposed framework is very good for use in real-life precision agriculture applications. The proposed approach can also function as a front-end module for autonomous weeding robots, UAVs, and IoT-based smart farming platforms without modifications. This will make it possible to manage weeds on a site-by-site basis with fewer herbicides, less harm to the environment, and more support for sustainable farming practices that boost crop productivity overall.