IDENTIFICATION OF COMMON WEED SPECIES USING EfficientNetv2B0
Marwa Y. Mohameda,b, Bayumy AB Youssefa, E.A. Ashmawyb, Islam Elkabanib,c
aComputer Graphics and Multimedia Department, Informatics Research Institute (IRI), City of Scientific Research and Technological Applications (SRTA-City), Egypt
bDepartment of Mathematics and Computer Science, Faculty of Science, Alexandria University, Alexandria, Egypt
cSchool of Information Technology University of Cincinnati, Ohio, USA
Abstract: One of the main threats to agriculture and biodiversity is weeds; also, the management of this issue will become more difficult as a result of climate change. The variety of weed species would alter and some weeds may spread further. Owing to their irregular spatial distribution, a ground or aerial robot is utilized to perform targeted herbicide spraying. This application of herbicides relies completely on computer vision algorithms that aid in identifying weeds in the field before the spot spraying occurs. In this study, we aimed to develop a robust system for site-specific weed control in fields by utilizing color images and a deep learning approach to distinguish four common weeds: kochia, horseweed, ragweed, and redroot pigweed. This study uses a weed dataset collected from four locations in North Dakota, which includes 3,424 digital images showing four types of weeds. The results of using the EfficientNetv2B0 method for identifying these weeds were measured using important statistics like accuracy, marco average of precision, recall, and F1- score, which result in 99.5%, 100%, 100%, and 100% respectively. The findings show that EfficientNetv2B0 performs better than other classification methods like InceptionResNetV2, NASNetMobile, DenseNet121, Xception, and EfficientNetB0 as 97.8%, 95.9%, 98.5%, 97.3%, 99.1% respectively. Also, the study investigated the model’s performance in terms of total parameters, model size, and inference time. The EffcientNetv2B0 is faster, lightweight, and computationally efficient (both during training and inference).
Keywords: Weed classification, Deep learning, EfficientNetv2B0, Transfer learning
VOLUME 9 ISSUE 12 2025: 204 – 218