AGAVE PLANT DENSITY USING CONVOLUTIONAL NEURAL NETWORKS ON AERIAL IMAGERY

Abstract

Agave plants (Agave tequilana Weber) are an indispensable element in the tequila production chain. Traditionally, plantation monitoring has been done manually; however, having accurate information on agave inventories is crucial for planning and estimating production volume. In this context, it was proposed that deep learning algorithms can achieve high detection rates of agave plants, improving the management and control of plantations. For this purpose, YOLOv4 and YOLOv4-tiny convolutional algorithms were implemented and evaluated using high-resolution RGB aerial images captured by a remotely piloted aircraft system for the determination of agave plant density. Three flight plans were planned and carried out, with ground sampling distances of 1.10, 1.64, and 2.19 cm pixel-1, respectively. The database was created, and the algorithms were evaluated for a confidence level of 0.25 and an intersection threshold over the junction of 0.50. The results showed an average mean accuracy of 0.99 for both algorithms and an F1 score of 0.95 for YOLOv4 and 0.96 for YOLOv4-tiny. Furthermore, a high detection rate (Rc) of 99 % and precision values (Pr) between 90 and 92 % were obtained. A decrease in the performance of the algorithms was observed when detecting agave tillers in images with a spatial resolution of 2.19 cm pixel-1. The implemented YOLO convolutional algorithms proved to be highly robust and able to generalize agave plant characteristics at different phenological stages, allowing accurate detections. In addition, the coordinates of the detected plants were used to estimate the distance between them, with a maximum error of 20 cm.