Evaluation of YOLOv10n for Tomato Plant and Weed Detection in Agricultural Fields Using a Lightweight Deep Learning Approach

Abstract

Deep learning techniques have emerged as powerful tools in precision agriculture, enabling automated crop monitoring and weed management. This study investigates the performance of the YOLOv10n model, a lightweight and computationally efficient deep learning architecture, for object detection in tomato crop fields. A custom dataset containing field images of tomato plants and diverse weed species was collected and annotated. The model was trained and tested using a dataset consisting of 1,091 training images, 326 validation images, and 174 testing images. A total of 2157 annotated instances across two classes; plant and weed, were collected from field conditions. The model achieved a mean Average Precision (mAP@50) of 62.5% and mAP@50–95 of 40.7%, with an inference time of 5.6 milliseconds per image. Class-wise evaluation revealed high detection accuracy for tomato plants, with a precision of 87.3% and recall of 83.5%, indicating the model’s strong potential to identify and preserve crop regions. Weed detection, however, showed relatively lower performance, primarily due to intra-class variability and class imbalance. These findings suggest that YOLOv10n can effectively detect tomato plants in complex backgrounds, providing a reliable basis for future integration into real-time precision agriculture systems. Further enhancement in weed detection may be achieved through data augmentation and improved class-specific representation.