Detection And Counting Method Of Rice Panicle Based On Deep Learning

Rice is a key crop that is a staple food in many countries.The detection and counting of rice panicles in agricultural fields is one of the most direct and effective methods for determining rice growth and breeding conditions and predicting rice yield.Currently,UAV remote sensing technology combined with deep learning has been widely used in agriculture,enabling an appropriate balance between image quality,perceptual efficiency and operational cost to be achieved.However,the variable phenotypic characteristics,small size and localised occlusion of rice panicles during the growth cycle,and the irresistible imaging interference factors all add to the difficulty of rice panicles identification.Therefore,it is very challenging to achieve accurate rice spike counts in the field quickly.On the other hand,current research lacks publicly available UAV images of rice panicles taken at different flight altitudes and at different fertility stages,making the use of UAV images for rice panicles detection and counting inadequate.In addition,the automated image-based methods for detecting and counting rice panicles that currently exist are not user-friendly and poorly accessible to non-computer specialist researchers.It is therefore particularly important to facilitate non-computer specialist researchers to easily obtain the results of rice crop counting through images.To address the above problems,this paper focuses on the research and application of field rice panicle detection and counting in terms of the construction of rice panicle UAV image dataset,the construction of Panicle-YOLO rice panicle detection network,and the construction of rice panicle detection and counting cloud platform and yield grade prediction,based on deep learning target detection as the theoretical basis.The main research work and contributions of this paper are as follows:1.In view of the fact that a publicly available,standardized and diverse UAV rice panicles dataset does not exist,this paper collects UAV images of rice panicles taken at different flight altitudes and at four fertility stages,and establishes a morphologically diverse UAV image dataset of rice panicles on this basis.The UAVs were equipped with visible light sensors,and three flight altitudes of 7 m,12 m and 20 m were set to capture rice panicles from the beginning of the crop to the late filling stage,pre-process the rice panicles images and label them.The established UAV rice panicles dataset has 5282 UAV images and 237659 rice panicles labels.The dataset contains various morphologies of rice panicles,which can be used as a benchmark evaluation dataset for the research of rice panicles detection and counting algorithms.2.Panicle-YOLO rice panicle detection model is proposed to address the problems of large variation of rice panicle phenotypic features and complex background environmental interference.The model utilizes a multi-branch structure to make the model perceptual field more diverse and to obtain more different contextual semantic information,and at the same time,channel attention is introduced and the loss function is adjusted to make the model better optimized.Through several rice panicle detection and counting experiments,the accuracy,effectiveness and generalization of the proposed algorithm for rice panicle detection and counting are demonstrated.3.To address the current problem that it is difficult for non-computer professional researchers to use the available rice panicle detection and counting algorithms,a cloud platform for rice panicle detection and counting that is easy to use by non-specialists is developed,and the proposed Panicle-YOLO and four other well-trained rice panicle detection models are embedded in this cloud platform.At the same time,to verify the growth stages of suitable detected rice panicles,the features of rice panicles at different growth stages were extracted and analyzed.Finally,the number of rice panicles is obtained by the cloud platform and the rice yield level is effectively classified using a supervised machine learning method with an accuracy of 84.03% on the validation set.The cloud platform designed in this paper can facilitate breeding experts to conduct in-depth research on the agronomic characteristics of rice panicles and provide technical support for rapid and intelligent breeding.