Multi-class weed detection based on deep learning using unmanned aerial vehicle images in potato fields

Abstract

Early weed detection is a critical task in precision agriculture, as weeds compete with crops for resources and significantly reduce yield and quality. This study proposes a multi-class weed detection approach based on deep learning using unmanned aerial vehicle (UAV) images acquired in real potato fields under uncontrolled environmental conditions. A custom dataset containing five classes (potato crop and four weed categories) was created, annotated, and made publicly available to support reproducibility and future comparisons. Two modern object detection architectures, based on convolutional neural networks and Transformer models, were evaluated through multiple training variants to analyze the impact of class balancing, image resolution, and hyperparameter optimization on detection performance. The models were assessed using standard object detection metrics and complemented with statistical validation to verify the reliability of the predictions in real-field conditions. Experimental results show that both architectures achieve robust and comparable performance, demonstrating their suitability for near real-time weed detection and precision agriculture applications. The findings confirm the effectiveness of combining advanced deep learning models with UAV imagery for accurate multi-class weed identification in complex agricultural environments.