Rice Panicles Counting Method Based on YOLOv7 Using Unmanned Aerial Vehicles Images

Abstract

【Objective】Based on deep 1earning techno1ogy, rapid counting of rice panic1es in RGB images co11ected by unmanned aeria1 vehic1es (UAV) is beneficia1 for 1abor saving, time saving, and efficiency. And it provides a basis for downstream harvesting, drying, warehousing and variety comparing and eva1uation.【Method】Images of rice panic1es were collected by UAV from the rice full heading stage to the fill stage. Images were annotated, grouped, and trained, and a network structure model based on YOLOv7 was obtained. The accuracies were evaluated by test datasets of rice panicles in different rice subspecies and validated by field investigation. The detection results were compared and validated with field investigation results.【Result】Comparing the predicted results of the model with the actual results, for the same training dataset, the median values of the Intersection of union (IoU) calculated from YOLOv7 model are generally higher than those of the YOLOv5 model. Models trained using images of rice subspecies Oryza sativa sp. japonica performed best on recognition of rice subspecies Oryza sativa sp. japonica. And the mean average precision (mAP), [email protected] was 80.75% and [email protected] was 93.01% of YOLOv7 model, while [email protected] was 73.36%, and [email protected] was 91.16% of YOLOv5 model. But recognition of two models for rice subspecies Oryza sativa sp. indica is not high. Models trained using images of rice subspecies Oryza sativa sp. indica performed best on recognition of rice subspecies Oryza sativa sp.indica. And [email protected] was 73.19% and [email protected] was 83.71% of YOLOv7, while [email protected] was 72.77%, and [email protected] was 81.66% of YOLOv5 model. But recognition of two models for rice subspecies Oryza sativa sp. indica is not high. Validation tests were conducted between the predicted results and observed results. The models trained using only images of rice subspecies Oryza sativa sp. japonica had accurate recognitions on rice subspecies Oryza sativa sp. japonica. Predicted results of the models had significant correlations with observed results. The correlation coefficient R2 was 0.9585 and the root mean square error (RMSE) was 9.17 of YOLOv7 model, while R2 was 0.9522 and RMSE was 11.91 of YOLOv5 model. The models trained using only images of rice subspecies Oryza sativa sp. indica had accurate recognitions on rice subspecies Oryza sativa sp. indica. Predicted results of the models had moderate correlations with observed results, R2 was 0.8595, RMSE was 24.22 of YOLOv7 model, while R2 was 0.7737, RMSE was 32.56 of YOLOv5 model.【Conclusion】This study preliminarily established a rapid survey method for the number of rice panicles per unit area in the field from images captured by UAV. And it had considerable accuracy and could be applied to practical work. This method is conductive to overcoming the problems of heavy workload, low efficiency and artificial errors. In future, it could be used to develop mobile devices of rice yield investigation and estimation for different scenarios.

Keywords

• rice
• panicles detection
• yolov7
• object detection
• deep learning