Experiment And Optimization Of Soil Organic Matter Detection Method Based On Machine Olfaction Technology

Soil organic matter is an important constituent of the soil solid phase system.It is the most important source of organic carbon on earth and also has a large impact on the quality of crop growth,developmental processes and final yield,and improper soil management practices can easily result in lower organic matter content.Therefore,measuring its content change is of great significance for ecological protection and crop production.Chemical detection methods are limited by many aspects such as detection time,operational complexity,and professional personnel,and do not have the possibility of in-situ detection.Soil structure,moisture,and iron oxide can interfere with the accuracy of spectral detection.The analytical instruments of thermal decomposition gas chromatography mass spectrometry have shortcomings such as large volume and high cost.Currently,the research and development of machine olfactory systems in soil organic matter has achieved better results compared to traditional methods,with the main advantages of high accuracy,low cost,and fast detection speed.However,the existing research sensors have high power consumption and relatively large volume;The optimal sensor array for soil organic matter detection has not been selected,and redundant sensors may exist;Lack of suitable pattern recognition algorithms,etc.Therefore,it may cause adverse consequences such as low system recognition accuracy.In order to improve the prediction accuracy of soil organic matter content detection based on machine olfactory technology,the following research work has been mainly carried out in this paper:(1)Based on the machine olfaction technology,a soil organic matter detection system with MEMS metal oxide semiconductor gas sensor as the core has been developed independently.Two major parts,thermal cracking device and machine olfactory detection device,together form the hardware structure of the system.The thermal cracking device mainly consists of a miniature tube muffle furnace,vacuum flange,rubber hose,quartz tube,etc.,which is used to excite and increase the concentration of soil cracking gas.The machine olfactory detection device is mainly composed of gas-sensitive sensor array,gas pump,signal processing module,NI data acquisition card,laptop computer and other parts.The software part was developed with LabVIEW as the acquisition software,which has the function of real-time display and storage of the gas response information collected by NI data acquisition card.Soil was randomly selected to test the detection system pre-experiments and explored the test process system working parameters and processes.The test results show that the array composed of 10 different types of MEMS gas sensors can realize the work of collecting the information of soil cracking gas,and lay the foundation for the system to accurately detect the organic matter content of the soil sample set.(2)The effects of different array optimization methods on the establishment of soil organic matter prediction models were investigated.The detection of soil sample set information was completed by this system,and the response curves used to characterize the sample set information were obtained.The first-order integrated value,first-order differential value,relative change value,relative steady-state mean value,mean value,8th second transient value and maximum value of the response curve were extracted,and a 112×10×7 soil fission gas olfactory feature space was constructed.Genetic algorithm(GA),RReliefF algorithm and Boruta algorithm in feature selection algorithm were used to optimize the original array,and the optimized results were constructed as extreme learning machine model(ELM)with regression prediction index as the evaluation criterion.The results show that all the array optimization methods used can eliminate a certain number of sensors and reduce the number of features,and at the same time can improve the prediction performance of the model,and the optimization effects of the three algorithms are in the following ascending order:GA<RReliefF<Boruta.after using Boruta algorithm to optimize the original array,the R~2,RMSE,and RPD in the training set changed by 15.94%,27.59%,and31.29%,respectively.27.59%,31.29%,and the decision coefficient,RMSE,and RPD in the training set changed by 16.83%,23.84%,and 38.12%,respectively,with the most excellent optimization effect.(3)Linear feature dimensionality reduction(PCA)and nonlinear dimensionality reduction(KPCA)methods were used to reduce the olfactory feature space,respectively.The transformed8-dimensional new features namely contain most of the information of the original features in 20dimensions,while reducing the correlation between features.The sum of the first 8 kernel principal components of KPCA was 98.92%,and its prediction set R~2=0.8434 was regressed using SVM model.soil organic matter inversion models of LOOCV-PLSR,SSA-RF,and PSO-LSTM were constructed.For the main key parameters affecting the performance of different models,the optimal number of potential variables in the PLSR model was determined using the leave-one-out cross-validation method;the appropriate number of decision trees and the depth of trees in the RF model was found using SSA;the optimal parameters of the learning rate and the number of implicit layer cells in the LSTM model were obtained using the PSO pair.Among them,the PSO-LSTM model has excellent prediction performance with the prediction sets R~2,RMSE,and RPD of 0.9288,0.7400,and 3.7282,respectively,which is expected to be helpful for accurate detection of soil organic matter content.