Document Type : Original Article

Authors

• Nadia Saadati ¹
• Ali Najar ²

¹ Department of Agricultural Biosystems Engineering, University of Mohaghegh Ardabili, Ardabil, Iran

² Department of Computer Engineering, Sharif University of Technology, Tehran, Iran

Abstract

With advancements in technology, particularly in electronics and mechanics, the agricultural sector is increasingly looking to adopt these innovations. One of the areas of interest for researchers is the use of modern technologies to optimize herbicide spraying in agricultural fields. Manual removal of weeds and the use of herbicides are time-consuming and costly, and cause more resistance in weeds. It also has many consequences for the environment and humans. As a result, it is necessary to use herbicides optimally and appropriately. One of these ways is the machine vision system. In this study, we developed a video-based machine vision system designed to identify two common weeds found in potato fields: Chenopodium album (Common lambs quarters) and Polygonum aviculare (Knotweed). After video recording, preprocessing, and segmentation, 1688 individual objects were detected. Using a hybrid of an artificial neural network and simulated annealing algorithm, four key features were selected from an initial set of thirteen, including the third moment invariant, perimeter, fifth moment invariant, and sum entropy. These features were then used in a hybrid classifier combining an artificial neural network and a bee colony algorithm to classify the weeds. To evaluate the classifier’s performance, we calculated sensitivity, precision, specificity, F1-score, accuracy, and false positive rate. For test data, the sensitivity for Chenopodium album was 97%, and for Polygonum aviculare, it was 89%. The overall precision was close to 94%, while the specificity for Chenopodium album and Polygonum aviculare was 89% and 97%, respectively.

Keywords

• Artificial intelligence
• Classification
• Site-specific application
• Video processing
• Weeds