Detection and Counting of Corn Plants in the Presence of Weeds with Convolutional Neural Networks

Corn is an important part of the Mexican diet. The crop requires constant monitoring to ensure production. For this, plant density is often used as an indicator of crop yield, since knowing the number of plants helps growers to manage and control their plots. In this context, it is necessary to detect and count corn plants. Therefore, a database of aerial RGB images of a corn crop in weedy conditions was created to implement and evaluate deep learning algorithms. Ten flight missions were conducted, six with a ground sampling distance (GSD) of 0.33 cm/pixel at vegetative stages from V3 to V7 and four with a GSD of 1.00 cm/pixel for vegetative stages V6, V7 and V8. The detectors compared were YOLOv4, YOLOv4-tiny, YOLOv4-tiny-31, and YOLOv5 versions s, m and 1. Each detector was evaluated at intersection over union (IoU) thresholds of 0.25, 0.50 and 0.75 at confidence intervals of 0.05. A strong F1-Score penalty was observed at the IoU threshold of 0.75 and there was a 4.92% increase in all models for an IoU threshold of 0.25 compared to 0.50. For confidence levels above 0.35, YOLOv4 shows greater robustness in detection compared to the other models. Considering the mode of 0.3 for the confidence level that maximizes the Fl-Score metric and the IoU threshold of 0.25 in all models, YOLOv5-s obtained a mAP of 73.1% with a coefficient of determination (R-2) of 0.78 and a relative mean square error (rRMSE) of 42% in the plant count, followed by YOLOv4 with a mAP of 72.0%, R-2 of 0.81 and rRMSE of 39.5%.

Automated summary

This study developed a method using aerial RGB images and deep learning algorithms to detect and count corn plants, and compared the performance of different detectors.