| Crop yields are important economic information and are closely linked to the livelihood of the nation.Accurate identification and assessment of crop growth cycles is an important step in promoting the use of agricultural technology.The identification of crop growth is becoming a mainstream technical tool that not only avoids the problems of low accuracy,time and cost of traditional estimation methods,but also enhances the identification of crop growth processes during the growing season with the help of modern image recognition technology.In the growth period of maize,the number of maize plants can be obtained accurately and quickly,and then the development of the maize crop can be determined,which can play a predictive role in the maize yield.The main research elements are as follows:1.Acquisition of data and construction of data setsUsing the maize field in Tongjiang Town,Heilongjiang Province as the data collection point,a high-definition UAV was used for image acquisition.In terms of data set construction,the images were classified and annotated according to different growth stages,and divided into three stages: seedling stage,nodulation stage and maturity stage.Convolutional neural network models were then used to extract and classify features from the images in the dataset to enhance the richness and quality of the dataset.A dataset containing thousands of images of maize growth stages was finally obtained.2.Seedling and nodulation image recognition studyTwo hundred images of maize at the seedling and nodulation stages were collected and divided into healthy and unhealthy maize images.For healthy plant coverage identification,image segmentation techniques and convolutional neural network models were used to segment and identify plants in the maize images and calculate their coverage.For the determination of missing plants at the seedling stage,a convolutional neural network model was used to determine and identify missing plants in the maize images.Ten cross-validation experiments were conducted on the dataset and the results showed good performance in terms of accuracy.By applying digital image processing and convolutional neural network models,we were able to effectively monitor and identify maize growth and health conditions.3.Ripening stage image recognition study1000 images of maize at maturity were collected for local image recognition calculation to obtain the relevant features of each maize grain in the image.The yield and quality of each maize grain were also manually annotated to construct a maize image dataset containing both yield and quality dimensions.Techniques such as data enhancement,balanced sampling and multi-labelled annotation were used.For the convolutional neural network-based regression model,the Tassel Netv2 model was used to predict the yield and quality of each maize grain.Through validation experiments,the results show that our proposed method performs well in terms of average accuracy,mean absolute error values,root mean square error values and coefficient of determination R2. |
