Seedling Monitoring And Tasseling Stage Identification Of Field Maize Based On CNN

As one of the main food crops in China,the stable production of corn is the key task of chinese food security.Rapid,nondestructive and accurate acquisition of large area maize yield data can provide effective data and technical support for current agricultural decision-making management and production.The seedling rate and tassels of maize are closely related to yield.Therefore,the detection of maize seedlings and tassels in the two growth periods is very important for yield monitoring.To this end,this study mainly uses deep learning methods to detect and count the seedlings and tassels in the top view of maize.The work done is summarized as follows:(1)Corn top view image acquisition and data set construction under two different platforms.Firstly,visible image was obtained for 13 maize varieties in the plots covered by the HTTP in the field.In order to ensure the clarity of data acquisition by imaging equipment,the distance between the imaging system and the top of corn canopy was adjusted to 1.5m after several experiments.From the beginning of tasseling stage to the end,visible images of corn in the monitored field were acquired twice a day to form visible image sequence of corn during the whole tasseling stage.Secondly,the visible light images of corn plants of different varieties in the experimental fields covered by the UAV platform were also obtained once a day during the tasseling stage.Finally,the visible light images of HTTP and UAV platform were used to construct the visible light image data set of corn during the tasseling stage under different light conditions.(2)Deep learning based method was used to construct the detection and counting model of corn tassel.Res Net 50 was used as a new feature extraction network instead of VGG 16 in the original Faster R-CNN to optimize it.The improved model was used to detect and count the tassels of corn in two kinds of visible images.The mean average precision of the improved model on the HTTP test set reached 90.14%,which was 1.6%higher than the original Faster R-CNN model.The mean average precision on the UAV platform test set also reaches 82.14%,which is 2.39% higher than the original Faster R-CNN model.(3)The model constructed in this study was used to identify the tasseling stage of maize plants.Based on the high time sequence and continuous image sequence of maize obtained by the HTTP,the improved Faster R-CNN model was used to detect and count the tassels of corn in the image according to the determination method that the tasseling proportion of single row corn plants in the field is more than half when they enter the tasseling stage.The statistical results showed that there were some differences in the time of entering the tasseling stage of different varieties of corn,and the time needed from the beginning to the end of tasseling of most varieties of corn was generally 5 to 7days.(4)The improved Faster R-CNN model was further applied to detect and count maize seedlings in images of different platforms and compared with YOLOv3.The results showed that the improved model in this study performed better than YOLOv3 in the field maize seedlings detection and counting task,but the detection speed was slower than that of YOLOv3.Moreover,the detection results of the two models on the HTTP were better than that of the UAV platform.Finally,the dynamic monitoring of the diurnal changes in the number of maize seedlings of different varieties and densities before and after the seedlings of 10 days was realized by using the time-series images of the HTTP.The results showed that the time from the beginning to the end of the seedling of XD20 was shorter than that of AD268 under the same density in this study,that is,the seedling was more orderly.