基于YOLOv7的无人机影像稻穗计数方法研究

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

【目的】利用深度学习技术开展基于无人机采集的水稻稻穗RGB图像进行稻穗快速计数技术研究,利于建立省工、省时、高效的产量评估预测,为后续收获、烘干、仓储工作以及品种试验评估等提供依据。【方法】在水稻齐穗-灌浆期,使用无人机采集水稻稻穗图片,通过对图片中稻穗的标注、分类以及训练,获得基于YOLOv7的网络结构模型,与田间实际调查结果进行对比和验证,针对该方法对不同亚种水稻稻穗的数穗计数精度作出评价。【结果】将得到的模型的预测结果与真实结果进行比较,对于相同的训练集,YOLOv7模型的重叠率(Intersecion of union,IoU)值的中位数普遍高于YOLOv5模型。仅使用粳稻数据训练得到的模型对粳稻有较好的识别精度,YOLOv7模型的mAP@0.5为80.75%、mAP@0.25为93.01%,优于YOLOv5l模型的mAP@0.5值73.36%、mAP@0.25值91.16%;两种模型对籼稻识别精度不高。对籼稻识别最佳的模型为使用籼稻数据训练得到的模型,YOLOv7模型的mAP@0.5为73.19%、mAP@0.25为83.71%,优于YOLOv5l模型的mAP@0.5值72.77%、mAP@0.25值81.66%;但两种模型均对粳稻识别精度不高。对预测结果与实际调查结果进一步比较验证表明,仅使用粳稻数据训练得到的模型对粳稻有较好的识别精度,模型预测值与观察值显著相关。其中YOLOv7模型对粳稻预测精度最高,R2为0.9585、RMSE为9.17;其次为YOLOv5模型,R2为0.9522、RMSE为11.91。对籼稻识别最佳的模型为使用籼稻数据训练得到的模型。其中YOLOv7模型对籼稻预测精度最高,R2为0.8595、RMSE为24.22。其次为YOLOv5模型,R2为0.7737、RMSE为32.56。【结论】本研究初步建立的基于无人机的田间水稻单位面积穗数快速调查方法,具有较高精度,可应用于实际田间测产工作,有利于克服人工田间估产工作量大、效率低、人为误差等问题,未来可进一步应用于可移动水稻估产装置的开发。

【Objective】Based on deep learning technology,rapid counting of rice panicles in RGB images collected by unmanned aerial vehicles(UAV)is beneficial for labor saving,time saving,and efficiency.And it provides a basis for downstream harvesting,drying,warehousing and variety comparing and evaluation.【Method】Images of rice panicles 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),mAP@0.5 was 80.75%and mAP@0.25 was 93.01%of YOLOv7 model,while mAP@0.5 was 73.36%,and mAP@0.25 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 mAP@0.5 was 73.19%and mAP@0.25 was 83.71%of YOLOv7,while mAP@0.5 was 72.77%,and mAP@0.25 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.