Study On Cotton Nitrogen Nutrition Diagnosis Based On Deep Learning

Timely and accurate estimation of nitrogen content and real-time monitoring of nitrogen nutrition status of cotton are not only important for cotton production management,but also have important strategic significance for cotton trade and agricultural policy planning.Therefore,it is necessary to find a fast,accurate and efficient diagnostic method for cotton nitrogen nutrition.At present,automatic nitrogen nutrition diagnosis of cotton mainly includes remote sensing technology,digital image technology and industrial equipment and instrument technology.However,remote sensing technology has some problems such as low resolution and poor timeliness,which can not guide precise fertilization.Industrial equipment(such as SPAD instrument,etc.)is expensive and difficult to be accepted by farmers at present.But the traditional digital image technology is more complicated,the data processing speed is slow,the accuracy is easily affected;Based on this,this study used smart phone platform to collect digital images of cotton leaves,integrated the characteristics of digital image processing technology and the advantages of deep learning model,and established a multi-scale nutritional diagnosis model based on cotton leaves and canopy in flowering and bolling period.The main research results are as follows:(1)Combined with deep learning target detection technology to achieve the positioning and segmentation of functional leaves(the fourth stereotypical leaf of the main stem)of cotton.Using the improved mask-RCNN network as the reference network,functional leaves in cotton canopy digital images were detected and segmented.The result pixel array was input into Alex Net,VGGNet,Mobile Net,Shuffle Net,Efficient Net,Inception Net and other models for training.The experimental results show that:Shuffle Net was the best monitoring model for nitrogen nutrient level in flowering and bolling stage.The identification accuracy of Shuffle Net was more than 92% under five different nitrogen nutrient gradients,and its F1 fraction and Kappa coefficient were both above 0.9.The optimal nitrogen content estimation model was Efficient Net network,and the R^2 curve of the predicted value was 0.9755,RMSE was only1.0841mg/g and MAPE was only 6.361%.(2)Diagnosis of nitrogen nutrition in canopy vegetative organs of cotton under complex field environment.The cotton canopy image was taken as the deep learning data set,and the characteristics of all nutrient organs in the cotton canopy were learned by combining the backbone network structure of the model with better performance in the previous experiment and deepening the network layers.The cotton nitrogen nutrition diagnosis model under a complex background was established to realize the end-to-end diagnosis of cotton nitrogen nutrition level.In this study,an improved Res Ne Xt structure based on identity mapping and depth separable convolution was proposed to optimize the nutritional diagnosis model.The results showed that: Compared with Xception Net,Resnet-V2,Res Ne Xt,Inception resnet-v2 and other models,The Improved-Res Ne Xt model proposed in this thesis achieved the highest recognition accuracy,average accuracy and Kappa coefficient of 97%,97% and 0.96 in the classification of nitrogen deficiency level of cotton leaf images,respectively.The number of model parameters(25250017)and memory consumption(136.5M)were the lowest except Shuffle Net and Xception Net,with the best overall performance.It provides algorithm support for cotton nutrition monitoring research on embedded off-line equipment.