Research On Diagnosis Model Of Nitrogen Content Of Rubber Leaf Based On Hyperspectral Spatial Features And Channel Feature Extraction

Due to its unique physical and chemical properties,natural rubber has a wide range of applications in military,economic,medical and other fields,and is an important economic crop in our country.In the process of planting natural rubber,in order to achieve refined fertilizer application,it is often necessary to monitor and evaluate the nutritional status of crops.The existing methods for evaluating the nutritional status of agricultural crops include direct observation of appearance,physical and chemical analysis,and spectral analysis.Observation methods often rely on experienced operators to a large extent,but are also largely affected by their subjective factors.Physical and chemical analysis methods are often destructive.At the same time,due to the long analysis period and high cost,despite their high accuracy,physical and chemical analysis methods still have major drawbacks in terms of timeliness and environmental issues.The spectroscopic analysis method has received extensive attention because it meets the requirements of agricultural refinement and automation.Nitrogen is an important nutrient element of natural rubber.In order to improve the yield and quality of crops,these elements need to be effectively and accurately monitored during the planting process.In this paper,through the establishment of a rubber leaf nearinfrared spectroscopy and physical and chemical analysis of the true value data set,from the perspectives of statistical learning methods,convolutional neural networks and self-attention models,experiments for a variety of non-destructive diagnostic models for the nitrogen content of rubber tree leaves and their effectiveness are carried out.The specific research content of this article is as follows:1.This paper explores the information extraction methods of the spatial and spectral dimensions in the near-infrared spectroscopy imaging of rubber blades.Based on the kernel principal component analysis and the Gaussian mixture model in the statistical learning method,the near-infrared spectral imaging of the leaves is extracted based on the low-dimensional representation and clustering of the spatial dimensional information.Based on the spatial information,the weights of the global spectrum are updated to obtain the weighted average spectrum corrected by the spatial information.Therefore,the least square linear regression model that is most suitable for diagnosing the nitrogen content of leaves is trained.The discussion on the nonlinear correlation between the characteristics of the various channels in the leaf spectrum was verified,and an effective non-destructive diagnosis of the nitrogen content of the rubber leaf was realized.2.This paper proposes a method of using convolutional neural network to update the weight of each channel data of the leaf near-infrared spectrum.The leaf area information is embedded in the reflectance vector of a single pixel,and the multilayer information is extracted through a two-dimensional convolution operator.Use the convolution operator to learn the correlation between different bands and the nonlinear mapping relationship between each band and the estimated value of nitrogen content.The effectiveness of the model was verified through tests,and the nitrogen sensitive bands in the near-infrared spectrum data of the rubber blades and the nitrogen sensitive areas of the blades were located.3.Proposed a leaf spectrum analysis model based on the attention mechanism,and uses the nonlinearity of the activation function in the neural network to update the weight of the data split between the space dimension and the channel dimension.Explore the correlation between different regions,different wavebands and leaf nitrogen content,reduce data preprocessing steps,use complete sampling data to estimate leaf nitrogen content,so as to improve the robustness of the non-destructive diagnosis process of rubber tree leaf nitrogen content.Based on the attention mechanism,the global data is mapped in different spaces,and the global data is sparsely weighted through the query mechanism in the algorithm.