Deep learning models for detection and recognition of weed species in corn crop

Weed detection and control are important challenges in modern agriculture. Weed infestation can significantly reduce crop yields. The identification of weeds by species, along with their location, is important to reduce production costs and the environmental impact resulting from the use of chemical control across the plantation. In this study, we assessed four deep learning models for detection and recognition of weed species in corn crop. UAV flights were carried out over six corn farming areas at an altitude of 10 meters. Using LabelImg, we labeled almost 10,000 samples of six weed species with high incidence in corn crops. The resulting WEED6C-Dataset was made available for academic purposes. Model assessment was carried out using a 5-fold cross-validation, three metrics for classification evaluation, and six metrics for detection evaluation. Experimental results showed evidence for statistically significant differences between the assessed models. In our experiments, the Faster R-CNN architecture obtained the best results for recall, f-score, RMSE, MAE, R${}^2$, mAP50, mAP75 and mAP50-95. On the other hand, the SABL, FoveaBox and YOLOv3 architectures achieved higher precision rates for weed recognition in corn.