Performance evaluation of deep learning object detectors for weed detection for cotton

Abstract

Alternative non-chemical or chemical-reduced weed control tactics are critical for future integrated weed management, especially for herbicide-resistant weeds. Through weed detection and localization, machine vision technology has the potential to enable site- and species-specific treatments targeting individual weed plants. However, due to unstructured field circumstances and the large biological variability of weeds, robust and accurate weed detection remains a challenging endeavor. Deep learning (DL) algorithms, powered by large-scale image data, promise to achieve the weed detection performance required for precision weeding. In this study, a three-class weed dataset with bounding box annotations was curated, consisting of 848 color images collected in cotton fields under variable field conditions. A set of 13 weed detection models were built using DL-based one-stage and two-stage object detectors, including YOLOv5, RetinaNet, EfficientDet, Fast RCNN and Faster RCNN, by transferring pretrained the object detection models to the weed dataset. RetinaNet (R101-FPN), despite its longer inference time, achieved the highest overall detection accuracy with a mean average precision ([email protected]) of 79.98%. YOLOv5n showed the potential for real-time deployment in resource-constraint devices because of the smallest number of model parameters (1.8 million) and the fastest inference (17 ms on the Google Colab) while achieving comparable detection accuracy (76.58% [email protected]). Data augmentation through geometric and color transformations could improve the accuracy of the weed detection models by a maximum of 4.2%. The software programs and the weed dataset used in this study are made publicly available (https://github.com/abdurrahman1828/DNNs-for-Weed-Detections; www.kaggle.com/yuzhenlu/cottonweeddet3).