Dynamic monitoring and counting for lotus flowers and seedpods with UAV based on improved YOLOv7-tiny

Abstract

The numbers of lotus flowers and seedpods are the important agronomic traits for lotus breeding and field management. However, the traditional measurement method is manual, subjective, inefficient and labor-intensive. Therefore, this study proposed a novel approach for dynamic monitoring and counting lotus flowers and seedpods with UAV based on improved YOLOv7-tiny. To improve YOLOv7-tiny, the fusion mechanism of SPD layer and convolutions was studied to construct novel SPD-Conv blocks which outperformed the traditional SPD-Conv. Besides, a small target detection layer was adopted to enhance the performance for the lotus flower and seedpod detection. Subsequently, a group of Convolutional Block Attention Modules were embedded in the neck of YOLOv7-tiny to optimize the utilization of channel-spatial information. The P, R, mAP@.5 and mAP@.5:.95 of the improved YOLOv7-tiny model achieved an increase of 0.6%, 4.0%, 3.5% and 3.3% comparing with YOLOv7-tiny. In addition, the MAE, RMSE, and R2 for flowers measurement were 1.90, 2.72, and 0.96 respectively, while 1.96, 2.78, and 0.97 for seedpods respectively. The results demonstrated that the improved YOLOv7-tiny model had satisfactory performance for lotus flowers and seedpods detection and counting. In addition, the specialized software base on the improved YOLOv7-tiny was developed, and a sliding window method was proposed to monitor the large-scale lotus field. This study provides an efficient and convenient measurement method for flowers and seedpods in lotus breeding and field management.