Performance evaluation of the deep learning system for weed recognization

Abstract Numerous approaches based on machine learning have emerged in recent years to enhance crop protection efficiency. One example is the utilisation of deep neural networks to differentiate between various weed types in actual event scenarios. Nevertheless, these methods often need substantial input from experts who work iteratively to design robust deep learning systems. To simplify such processes and conserve resources, researchers have explored a fresh method known as automated deep learning. The our technology's recognition of weeds through the use of machine learning was evaluated using plant seedlings and weed collection from plants as a dataset to address the issue of weed recognition. The study compared various configurations, including plant segmentation, using a collection of classifiers in place of Softmax, and training on datasets that contain noise. The findings indicated ensuring performance, with F1-scores of 93.1% and 90.2% based on the dataset utilised. These results align with AutoML-linked studies while falling short of manually fine-tuned deep-learning-based systems created by human specialists. To conclude, exploring the potential of combining manual expert work and automated deep learning could be a promising direction for enhancing efficiency in plant defence.