| Wheat as one of the China’s major food crops,it is vital to ensure the increase in wheat production and the safety of supply and demand.However,with the global environmental and climate changes,the unfavorable factors for increasing wheat production continue to increase.The development of high-throughput plant phenotyping algorithms to achieve efficient and high-precision monitoring of wheat structural and functional traits is the key to improving breeding efficiency and yield.Green Fraction(GF)refers to the fraction of green pixels in the crop canopy image.Under field conditions,the growth period of GF is an important trait that characterizes the interaction between genotype and environment of crops under certain climatic conditions.Image segmentation is a key point in plant phenotyping image analysis,which determines the accuracy of GF estimation.However,the existing methods commonly used in plant phenotyping image segmentation,including Excess Green Index(ExG)and Random Forest(RF).The segmentation results affected by occlusion,shadow,complex crop canopy structure and lighting conditions,etc.In response to this problem,this paper proposes a segmentation algorithm based on deep learning U-net model.The specific analysis contents are as follows:(1)An interactive pixel image annotation tool was used to construct a training and validation data set,which was fully considered the variability of the canopy structure,complex lighting conditions and various soil conditions,and get totaling more than 200 data sets.(2)In view of the fact that the existing image segmentation methods cannot effectively segment wheat images under high-resolution and complex background environment conditions,this work proposes a green segmentation method for wheat images based on deep learning U-net.This method is based on U-net network.Optimization based on the framework can accurately extract wheat green vegetation information from RGB images,laying a solid foundation for the study of wheat canopy structural properties.(3)In view of the problem that the accuracy of deep learning models is affected by the number of training samples,this work proposes to use high-precision simulation images to expand the training data and increase the model training samples.Through the mixed training model of simulated data and real data,it is found that the U-net model that uses simulated data for training can obtain high-precision segmentation results,which provides an effective solution to the problem of insufficient deep learning training samples.(4)Analyze the phenotyping characteristics of the improved and developed U-net model at the pixel scale and image scale through independent test data sets.The results show that the accuracy of the U-net model(R2=0.94,RMSE=0.05)is significantly better than other methods at both pixel and image scales,which fully proving the superiority of the U-net model in crop phenotyping image segmentation.In essence,the reason for the good performance of the U-net model is that the deep convolutional neural network has deep feature learning capabilities,which can well generalize the complex information in the wheat image and solve the image segmentation problem.Compared with wheat and other major food crops such as corn and rice,wheat leaves and stems are more slender,combined with more tillers,the overall canopy structure is more complicated.It is foreseeable that the application of U-net model in the segmentation of phenotyping images of crops such as corn and rice is also expected to achieve relatively satisfactory results,which is worthy of further exploration. |
PDF Full Text Request