Weed and crop species classification using computer vision and deep learning technologies in greenhouse conditions

Abstract

Site-specific weed management in Precision Agriculture is becoming a popular topic among researchers and farmers. The objective of this study was to classify weeds and crop species using RGB image texture features with the comparison of the Support Vector Machine (SVM) classification model and deep learning-based visual group geometry 16 (VGG16) classification models. A total of 3792 RGB images of crop and weed samples were captured from the greenhouse, including 2271 weed images and 1521 crop images. ReliefF feature selection algorithm was applied to select the most important features for prediction models. The SVM and VGG16 deep learning classifiers were used to classify four weeds (horseweed, kochia, ragweed, and waterhemp) and six crop species (black bean, canola, corn, flax, soybean, and sugar beets). Accuracy, f1-score, and kappa score metrics were used to evaluate model performance and data reliability. The VGG16 model classifiers had outperformed all the SVM model classifiers. The results showed that average f1-scores of the VGG16 model classifier were obtained between 93% and 97.5%. The f1-score value of 100% was obtained for the corn class in the VGG16 Weeds-Corn classifier, which seems outstanding for the corn crop production system. This study shows promising results of using a deep learning algorithm (VGG16) for weed identification in site-specific weed management in precision agriculture.

Keywords

• Deep learning
• Feature selection
• Machine learning
• Precision agriculture
• Weed classification