Research On Intelligent Diagnosis Method Of Rice Nitrogen And Potassium Nutrient Level Based On Deep Learning

Nitrogen and potassium are key mineral nutrients in the growth process of rice.Accurately diagnosing nitrogen and potassium nutrient levels and optimizing field fertilization are important ways to promote high-quality and high yield of rice.At present,the diagnosis method of rice nutrition level based on deep learning is difficult due to the dynamic changes in the phenotype of rice under different nutrition levels.In this paper,a spatio-temporal convolutional network model Res Net101-SE-LSTM is proposed based on the RGB sequence images of the rice canopy taken by UAV in 2021and 2022,to identify the nitrogen and potassium nutrient levels of rice at the early-panicle initiation stage（EPIS）.The main research content and results are as follows:（1）Acquisition and preprocessing of rice canopy images.In the 2-year experiment,using the conventional indica rice Huanghuazhan（HHZ）and the conventional japonica rice Xiushui（XS134）as the research objects,16 groups of nitrogen and potassium interaction treatments were set up on different farmlands,and 16 groups of rice populations with different nitrogen and potassium contents were constructed.Then the sequence images of rice in different growth stages were obtained by UAV,and time series and non-time series data sets were produced for model training and testing.（2）Research on diagnosis model of nitrogen and potassium nutrient level in rice based on deep learning.On the same HHZ dataset,using the transfer learning method,six models of VGG16,Alex Net,Google Net,Dense Net,inception V3 and Res Net101were trained and tested on the rice data set at EPIS,and the average recognition accuracy rates were 64.35%,60.12%,68.19%,67.58%,68.23%and 70.81%,respectively,indicating that it is difficult to achieve the ideal diagnostic effect only with the CNN model,and the Res Net101 model is more suitable for feature extraction of rice canopy images than the other five models.（3）Research on diagnosis algorithm of nitrogen and potassium nutrient level in rice based on spatiotemporal convolutional network model.The difficulty in diagnosing nitrogen and potassium nutrition levels in rice lies in:（1）rice canopy images under different nitrogen and potassium interactions have certain similarities;（2）rice growth is a dynamic process,and the nitrogen and potassium phenotypes of rice plants at different growth stages are different,so it is difficult and limited to diagnose a single growth cycle.In this paper,the channel attention mechanism of SENet is added on the basis of the feature extraction model of the Res Net101 model,and combined with the long-term short-term memory network（LSTM）to form a spatio-temporal convolutional network model Res Net101-SE-LSTM,when tested on the datasets of HHZ and XS134in 2021,the Res Net101-SE-LSTM achieved the highest accuracy of 85.38%and88.38%,and through the cross-dataset method,the average accuracy on the HHZ and XS134 datasets tested in 2022 are 81.25%and 82.50%,respectively,showing a good generalization.In addition,using the SVM model to identify the nutrient level of rice from 18 items of rice physiological data collected manually,the accuracy rates of identifying the nutrient level of HHZ and XS134 were 82.71%and 81.83%,respectively.By comparing the various evaluation indicators of the Res Net101-SE-LSTM model and the SVM model,the results show that the misjudgment and F₁ scores of the two models are basically the same,but the collection of rice physiological data is cumbersome and damaged.The method proposed in this paper is Harmless and convenient.