Weed detection and classification using UAVs and deep neural networks: mapping for localized treatment

Abstract

Accurate detection and identification of weeds are essential in the implementation of Precision Agriculture (PA). In this study, we addressed the detection and classification of weeds in maize and tomato in their early growth stages by integrating images acquired from UAVs and analysis based on advanced convolutional neural networks. Subsequently, the objective of this research was the creation of geo-referenced maps that would allow localized and selective post-emergence management of weed species. Our results indicate that the combination of UAV-captured imagery and deep learning algorithm-based image analysis provides an effective solution for this purpose. The accuracy and efficiency achieved in the identification of weeds in their early stage were promising, both in the set of species present in the crop and in individual species. This advance is of great relevance in the field of PA, as it would allow a more efficient management of weeds by selecting the herbicide according to the type of weed present, reducing the use of herbicides and ultimately contributing to the sustainability and profitability of agriculture. In addition, the generation of geo-referenced maps facilitates real-time decision making. In summary, this study suggests that the combination of emerging technologies, such as drones and deep neural networks, may be applicable tools in localized weed management in the context of PA.