Time Series Prediction For Phenotypes Of Non-heading Chinese Cabbage

Phenotypes are observable characteristics of crops,which can evaluate the growth of plants.Recently,researchers mainly study the relationship between phenotypes and properties of plants,and analyze the behaviors of plants from the perspective of phenotype.Statistically,the change of phenotypes can be considered as time series data.Precisely predicting this change has many applications in the fields of fertilization,growth assessment,yield prediction and so on.Therefore,taking non-heading Chinese cabbage as object,this paper studies the time series prediction of canopy morphological characteristics.The thesis includes data acquisition system,feature extraction,feature prediction and model application.The main contents are as follows:(1)Data acquisition systems are designed and established.The environmental parameter acquisition system is organized in the form of master and slave computers.The master computer is strawberry pie,providing functions of port configuration,data persistence and interactive interface.The slave computer is Arduino.It communicates with the sensors using RS485 or analog ports,and collects soil water content,temperature,humidity and light intensity.The image capture system utilizes Arduino to control the power supply and shooting switch of the digital camera.The system will also turn on the flat light when the light intensity is weak.(2)Based on traditional machine vision,an interactive Grab Cut segmentation algorithm is proposed.By manually marking the front spots and background spots several times,the algorithm can complete segmentation iteratively.However,the background part contains noise similar to blades,leading to fragmentations in edges.So fully convolutional network is studied.Aiming at the loss of spatial information during feature extraction,the superficial convolution output is added to the reconstruction process.In order to solve the problem of variant scales of features,the dilated convolutions with different expansion rates are connected in parallel to calculate the encoder’s output,so that the feature tensor contains multi-scale information.The experimental results show that the pixel accuracy of the algorithm is 0.9735 and the mean intersection over union is 0.9579,which is better than Grab Cut and original fully convolutional network.(3)A time series model based on long short-term memory is proposed.In order to handle unstable sequences,a bidirectional model structure is designed to extract features of reverse sequence.Meantime,Attention is introduced to provide long-term dependence.For the hyper parameter selection,particle swarm optimization algorithm is utilized to complete optimization.The experimental results show that the model can predict precisely under different time scale.When time step is one hour,the data sequence is unstable,the RMSE of the model is 0.294(88)~2,and R~2 is0.9628.When time step is one day,the randomness in sampling is eliminated,The RMSE is 0.238(88)~2,and R~2 reaches 0.9763.(4)The application of time series prediction in real production is explored.First,the model predict changes of canopy leaf area under different concentrations of nutrient solution.Then,using area sequence as input,a XGBoost classification model is constructed to detect plant lodging;Finally,the canopy leaf area at harvest time is calculated,so that yield index such as fresh weight can be evaluated at an early stage.This paper studies the collection,extraction,prediction and application of canopy morphological characteristics of non-heading Chinese cabbage,which is beneficial for crop growth research.