Verification of a machine learning model for weed detection in maize (Zea mays) using infrared imaging

The potential of the framework of precision agriculture points towards the emergence of site-specific weed control. In light of the phenomena, the search for a cost-effective approach can help the discipline to accelerate the practical implementation. The paper presents a near-infrared data-driven machine learning model for real-time weed detection in wide-row cultivated maize (Zea mays) fields. The basis of the model is a dataset of 5 120 objects including 18 species of weeds significant in the context of wide-row crop production in the Czech Republic. The custom model was subsequently compared with a state-of-the-art machine learning tool You only look once (version 3). The custom model achieved 94.5 % identification accuracy while highlighting the practical limitations of the dataset.