Journal of Agronomy and Field Innovations

Background: The Glycine max (L.) Merr. (soybean) is a significant oilseed crop grown worldwide on about 121 million hectares annually. The competition from weeds reduces the soybean yields by 20–40% and raises the cost of production due to labor-intensive manual identification and herbicide application. Weed control is performed traditionally, which is based on visual identification and suffers from poor accuracy due to variability and using ineffective care method.
Objective: To develop and implement a CNN-based classification framework for automatic and efficient recognition of soybean plants and the most common weed species. 
Methods: A wide-set image dataset comprised of 12500 RGB images (soybeans: 3500 images, broadleaf weeds: 4200 images, grassy weeds: 3100 images, seeds: 1700 images) was gathered from different soybean cultivation areas with using handheld cameras and UAV technologies at the stage of soybean growth stages V3-V8. The images were then processed, normalized, and augmented. Neural networks for weed recognition were created using transfer learning and ResNet50 model trained initially with ImageNet image classification task.
Results: The ResNet50 model that has been developed has successfully obtained a classification accuracy of 94.8% overall with average precision of more than 92% for each class. The accuracy for broadleaf weeds class was obtained at 96.2%, grassy weed - 93.5% and soybean plant - 95.1%. The average time of inference per image was established at 0.041 seconds thus allowing real-time implementation of the technology. The ROC-AUC values for every class exceeds 0.96. The comparative analysis proved that the proposed solution is more efficient than traditional machine learning solution which had accuracy at 81.3% as well as MobileNetV3 (91.4%).
Conclusion: CNN weed classification based on deep learning technology provides fast, reliable and cheap automated weed identification for soybean production in terms of herbicide application. The use of UAV platform and autonomous sprayers can lead to changes in the field of agriculture and environmental protection.