Simultaneous Determination Of Nitrogen, Phosphorus, Total Organic Carbon, And Multiple Cations In Soil Solutions Based On Machine Learning

With the development of society,the problem of soil pollution is becoming more and more serious.Real-time monitoring of soil indicators is very important for environmental workers.Traditional detection methods can only test one type of indicator individually,but cannot measure multiple indicators at the same time.At the same time,it is difficult for traditional methods to measure the indicators of complex chemical systems such as domestic sewage,river water and soil solution,which require complex pretreatment and separation technologies.The search for convenient and rapid methods for simultaneous determination of complex chemical systems is of interest to many researchers.In this study,a new method for simultaneous determination of multiple indices in complex chemical systems was proposed.The concentration of each indicator in soil solution can well reflect the change of soil properties.That is why soil solutions have become of interest to most researchers.In this paper,soil solution as the research object,using spectroscopic method and machine learning technology to simultaneously determine a variety of indicators in soil solution.Taking the establishment of soil indicator rapid detection system in Chang 'an District of Xi'an as an example,the experimental algorithm was constructed.This method is the basic stage for the establishment of environmental regional monitoring network,which can lay the foundation for building an intelligent soil pollution detection technology for the study area.After years of exploration,the research group has made great progress in both the stability of new spectral instruments and the selection of deep learning procedures.It has stepped from the previous exploration stage of laboratory standard solutions to the detection of real samples.Firstly,by collecting soil samples and obtaining soil solutions,the concentrations of ammonia nitrogen,total nitrogen,K+,Ca2+,Na+,Mg2+,total phosphorus and total organic carbon in each soil solution were measured by large-scale instrument.The composite chromogenic agent was then allowed to react with the soil solution in the sample cell of the spectrometer for chromogenic reaction.Based on the high throughput technology,144,000 experimental characteristic images were obtained by using a new spectroscopic instrument.The characteristic image was used to represent the concentration change of each indicator in soil solution.The feature images and the corresponding concentration labels were entered into the program respectively,and the model was established by supervised learning.Machine learning is used to capture the subtle variations between these feature images.The performance of the model is measured by the fitting effect of the test set.The "gray-scale contour plot" and"gray-scale curve plot" are plotted using the feature images to analyze the similarity and difference of the feature images from the plots.The experimental repeatability and feasibility of the research method are verified.The main conclusions are as follows:(1)A method for presenting information inside complex chemical systems(soil solutions)was established.A new spectroscopic instrument was used to record the change of indicator concentration in soil solution.The change of indicator concentration was presented by the color of chromogenic agent and soil solution.As a carrier of soil solution sample information,feature image is also an important part of experimental data.(2)The screening of the composite chromogenic agent was completed.The selection of chromogenic agent is the most important part in the experiment.By conducting a hole plate experiment with a single chromogenic agent and soil solution sample(the indicator corresponds to the chemical solution),the color difference between the color development result image and the reference image is calculated.The larger the color difference the more obvious the color response and the more favorable the final experimental results.The most suitable single chromogenic agents were selected from the histogram of color difference value:Neutral red,Chromium azure,Bromothymol blue,Bromocresol purple,Bromophenol blue,Methyl orange,Bromopotassium phenol green,Calcium reagent sodium carboxylate,Potassium phenol red.The final compound chromogenic agent is prepared by single chromogenic agent.(3)Three networks(Inception V1,ResNet-50,and SqueezeNet V1.1)were selected to build the models,and the effectiveness of test set fitting was used to measure the advantages and disadvantages of the models established by the three networks.The models built by Inception V1 and ResNet-50 were more accurate,and the R2 of the eight indicators corresponding to the test set reached 99.00%.Because the Inception V1 procedure is more stable,the Inception V1 network was used to study other influences on machine learning.When the data density of the feature image was changed,it was found that when the feature image was 25 px×25 px,the R2 of the test set corresponding to machine learning reached 98.79%,indicating that the feature image had holographic properties.The excellent performance of the model on the test set shows the reliability of the research method.(4)The feasibility of the experimental method was verified by analyzing the feature images.The "grayscale-curve","grayscale-contour" and "three-dimensional spatial image" were drawn to analyze the differences between the feature images.The changes of gray values in the same position of different feature images can be seen from the "grayscale-curve ".The "grayscale-contour plot" shows the variability of the feature images by color differences.The "3D spatial image" is obtained by drawing the difference images between the feature images,which reflects the difference information between the feature images in a three-dimensional form.The repeatability of the experiment was verified by the gray value analysis of the parallel sample's characteristic images.The similarity of feature images of parallel samples can reach more than 98.00%.(5)In this paper,a machine learning strategy for the analysis of multiple chemical indicators in soil solutions based on color spectral images is proposed.The results show that a well-trained machine learning model can accurately predict multiple chemical indicators simultaneously,which is a reference value for the quantitative analysis of soil solutions and other complex chemical systems.