【作者】 刘鹤；

【导师】 黄东岩；

【作者基本信息】 吉林大学 ， 农业机械化工程， 2023， 博士

【摘要】 土壤是植物生长的重要载体,而土壤中的营养成分对于植物的生长、发育和产量有着直接且关键性的作用。其中,土壤有机质、全氮、有效钾和速效磷的含量是土壤肥力和养分的重要指标。因此,快速、高效地获取土壤主要养分含量是保持土壤肥力、维持良好耕地质量的基础。然而,现有的化学检测方法对土壤有着较强的破坏力,缺乏时效性和简易性。本文旨在融合工学、农学、生物学等多学科知识,运用热裂解技术、机器嗅觉技术和电子信息工程等交叉学科技术,为土壤主要养分含量快速、灵敏、低成本的检测提供新研究思路与新方法。突破了传统检测方法的“成本高、检测周期长”等检测难题,促进土壤养分检测设备的智能化、精准化发展。具体研究内容和研究结果如下:(1)基于热裂解技术和机器嗅觉技术的土壤主要养分含量检测装置设计。首先,利用热裂解技术搭建热裂解室实现对土壤样本在400℃时的高温、快速裂解。然后,采用机器嗅觉技术,将土壤裂解后的气体通入由20种不同类型的半导体气体传感器组成的传感器阵列并检测相关气体浓度变化。最后,设计了由信号处理电路、采集卡以及Lab VIEW软件开发的上位机组成的数据采集系统,实现了传感器响应数据的传输、采集、显示和储存等。对搭建的检测装置进行了响应测试试验和土壤全氮分类试验,所有传感器在通电一段时间后会保持稳定输出状态,并在通入土壤高温裂解气体后产生独特且不同的剧烈响应,准确分类了土壤全氮的含量等级,为土壤主要养分含量检测方法的建立奠定了试验基础。本文创新性地设计出一个集多种技术为一体的土壤主要养分含量检测装置,该装置具有灵敏度高、操作简便、成本低等优点。(2)构建土壤主要养分原始嗅觉特征空间并提出基于相关系数和互信息的传感器阵列优化方法。根据传感器响应数据特点,提取了经卷积函数平滑去噪后响应曲线的第7秒瞬态值、一阶平均微分系数、最大值和相对稳定状态均值构建了112×80(112个样本×20个传感器×4个特征)维原始嗅觉特征空间并完成了所有数据的均值方差归一化处理,归一化后的土壤主要养分含量数据大体成正态分布。采用相关系数分析传感器相关性,利用互信息测试相关程度极高的传感器组合,剔除了重复或冗余的传感器。其中,土壤有机质、全氮、有效钾和速效磷经传感器阵列优化后覆盖的传感器个数相较原始传感器阵列分别减少了2、1、7和5个,简化传感器阵列规模的同时降低了传感器之间的交叉敏感性,为后续嗅觉特征空间优化、预测模型的建立奠定基础。(3)土壤主要养分嗅觉特征空间优化方法对比研究。采用2种特征选择算法(RF、Boruta)和2种特征提取算法(PCA、SPCA)分别对不同土壤主要养分经传感器阵列优化后的特征空间完成优化,以决定系数(R~2)为评价指标建立长短期记忆(LSTM)预测模型。结果表明:特征提取算法PCA、SPCA可以很好地减少特征空间的复杂度和规模,降维效果显著,但R~2提升幅度较小;特征选择算法中,Boruta算法相较于RF简单的特征排序,考虑到了特征的完整性质,使R~2提升幅度最大。其中,经过Boruta特征选择后的嗅觉特征空间相较于经过传感器阵列优化的嗅觉特征空间,土壤有机质、全氮、有效钾和速效磷的R~2分别提升至0.8231、0.8067、0.7149和0.6848,特征数分别减少了41、52、33和43个,传感器个数分别减少了3、6、3和4个,R~2分别提升了10.23%、9.38%、11.04%和15.96%。减少了嗅觉特征空间的复杂度,以更少的传感器和特征维度保留更广泛的嗅觉空间信息。(4)基于机器学习的土壤主要养分含量预测模型的建立与优化。采用随机森林(RF)、支持向量回归(SVR)、偏最小二乘回归(PLSR)和最小二乘支持向量机(LSSVM)算法分别对土壤有机质、全氮、有效钾和速效磷经Boruta特征选择后得到的嗅觉特征空间建立预测模型,以R~2、均方根误差(RMSE)、平均绝对误差值(MAE)和性能偏差比(RPD)为模型评价指标,其中,LSSVM模型的预测性能最佳。引入了麻雀搜索算法(SSA)、粒子群算法(PSO)、鲸鱼优化算法(WOA)对LSSVM模型的关键参数进行优化,其中,3种优化算法中PSO-LSSVM预测值与真实值的拟合程度最高,模型的评价指标和模型性能最佳。土壤有机质、全氮、有效钾和速效磷的PSO-LSSVM模型测试集评价指标R_v~2分别达到了0.9420、0.9400、0.8028和0.60119;RMSE_V分别达到了0.6928、0.0404、22.1000和4.6173;RPD_V分别达到了4.1415、3.9081、2.0554和1.5476;MAE_V分别达到了0.5631、0.0331、16.8424和3.6659;模型性能等级分别达到了“极佳”、“极佳”、“合格”和“较差”。PSO-LSSVM模型为土壤有机质、全氮含量预测提供了一种较为可靠的关系模型,有效钾、速效磷含量检测精度、模型性能等级虽不及土壤有机质、全氮,但也高于原始嗅觉特征空间性能,基本实现了有效钾和速效磷的近似定量预测,为后续研究提供参考方法。综上所述,本文搭建了一种基于热裂解和机器嗅觉的土壤主要养分含量检测系统,并对其传感器阵列、嗅觉特征空间及预测模型进行了优化,基本可以实现土壤主要养分含量的定量分析和预测,克服了化学方法“成本高、检测周期长”的检测难题,突破了光谱分析易受水分、氧化铁等影响从而导致精度低的瓶颈。为土壤主要养分含量检测提供了一种快速、高效、经济的新方法。更多还原

【Abstract】 Soil is an important carrier for plant growth,and the nutrients in soil play a direct and crucial role in plant growth,development,and yield.Among them,the content of soil organic matter,total nitrogen,available potassium and available phosphorus are important indicators of soil fertility and nutrients.Therefore,obtaining the main nutrient content of soil quickly and efficiently is the foundation for maintaining soil fertility and maintaining good farmland quality.However,existing chemical detection methods have strong destructive power on soil and lack timeliness and simplicity.The purpose of this paper is to provide new research ideas and methods for rapid,sensitive,and low-cost of soil nutrient content by integrating engineering,agronomy,biology and other multi-disciplinary knowledge,and using interdisciplinary technologies such as pyrolysis technology,machine olfaction technology and electronic information engineering.Breaking through the challenges of traditional detection methods such as high cost and long detection cycle,promoting the intelligent and precise development of soil nutrient detection equipment.The main research content and results are as follows:(1)Based on the pyrolysis technology and machine olfaction technology,the design of soil nutrient content detection device.First,the pyrolysis technology was used to build a pyrolysis chamber to achieve high temperature and rapid cracking of soil samples at 400℃.Then,using machine olfactory technology,the gas from soil cracking is introduced into a sensor array composed of 20 different types of semiconductor gas sensors and the relevant gas concentration changes are detected.Finally,a data acquisition system consisting of a signal processing circuit,a collection card,and an upper computer developed with Lab VIEW software was designed to achieve the transmission,collection,display,and storage of sensor response data.Response testing and soil total nitrogen classification tests were conducted on the constructed detection device.All sensors maintained a stable output state after being powered on for a period of time,and produced unique and different intense responses after introducing soil high-temperature pyrolysis gas.The accurate classification of soil total nitrogen content levels laid the experimental foundation for the establishment of soil main nutrient content detection methods.This article innovatively designs a soil main nutrient content detection device that integrates multiple technologies.The device has the advantages of high sensitivity,simple operation,and low cost.(2)The original olfactory feature space of main soil nutrients was constructed and the optimization method of sensor array based on correlation coefficient and Mutual information was proposed.Based on the characteristics of sensor response data,the 7th second transient value,first-order average differential coefficient,maximum value,and relative stable state mean of the response curve smoothed and denoised by convolutional function were extracted,and 112×80(112 samples×20 sensors×4 features)were constructed dimension original olfactory feature space and completed the normalization of mean and variance of all data,and the normalized soil nutrient content data is generally Normal distribution.The correlation coefficient is used to analyze the sensor correlation,and the Mutual information is used to test the sensor combination with high degree of correlation,and the duplicate or redundant sensors are eliminated.Among them,the number of sensors covered by Soil organic matter,total nitrogen,available potassium and available phosphorus optimized by the sensor array is reduced by 2,1,7 and 5 compared with the original sensor array,which simplifies the scale of the sensor array and reduces the cross sensitivity between sensors,laying the foundation for the subsequent optimization of olfactory feature space and the establishment of prediction model.(3)Comparative study on spatial optimization methods for olfactory characteristics of major soil nutrients.Two feature selection algorithms(RF,Boruta)and two feature extraction algorithms(PCA,SPCA)were used to optimize the feature space of different soil nutrients after sensor array optimization,long short-term memory(LSTM)prediction model was established with coefficient of determination(R~2)as evaluation index.The results show that feature extraction algorithms PCA and SPCA can effectively reduce the complexity and scale of the feature space,with significant dimensionality reduction effects,but a small increase in R~2;In the Feature selection algorithm,Boruta algorithm,compared with the simple feature sorting of RF,takes into account the integrity of features,which maximizes R~2improvement.Among them,compared with the olfactory feature space optimized by sensor array,the R~2of soil organic matter,total nitrogen,available potassium and available phosphorus in the olfactory feature space after Boruta feature selection increased to 0.8231,0.8067,0.7149 and 0.6848,respectively,and the number of features decreased by 41,52,33 and 43,respectively.The number of sensors is reduced by 3,6,3 and 4,respectively and the R~2increased by 10.23%,9.38%,11.04%and15.96%,respectively,reducing the complexity of olfactory feature space,preserve broader olfactory spatial information with fewer sensors and feature dimensions.(4)Establishment and optimization of a prediction model for soil main nutrient content based on machine learning.Random forest(RF),support vector regression(SVR),partial least squares regression(PLSR)and least squares support vector machine(LSSVM)algorithms were used to establish prediction models for the olfactory feature space obtained from soil organic matter,total nitrogen,available potassium and available phosphorus after Boruta feature selection,coefficient of determination(R~2),root-mean-square deviation(RMSE),mean absolute error(MAE)and performance deviation ratio(RPD)were used as model evaluation indicators,among which LSSVM model had the best prediction performance.The sparrow search algorithm(SSA),particle swarm optimization(PSO),and whale optimization algorithm(WOA)were introduced to optimize the key parameters of the LSSVM model.Among the three optimization algorithms,PSO-LSSVM predicted values had the highest fitting degree with the true values,and the evaluation indicators and performance of the model were the best.The evaluation index R_v~2of PSO-LSSVM model test set for Soil organic matter,total nitrogen,available potassium and available phosphorus reached 0.9420,0.9400,0.8028 and 0.60119 respectively;The RMSE_Vreached 0.6928,0.0404,22.1000,and 4.6173,respectively;RPD_Vreached 4.1415,3.9081,2.0554,and 1.5476 respectively;MAE_Vreached 0.5631,0.0331,16.8424,and 3.6659respectively;the performance level of the model reached"excellent","excellent","qualified"and"poor"respectively.PSO-LSSVM model provides a more reliable Relational model for the prediction of Soil organic matter and total nitrogen content.Although the detection accuracy and model performance level of available potassium and available phosphorus content are lower than those of Soil organic matter and total nitrogen,they are also higher than the original olfactory feature space performance.It basically realizes the approximate quantitative prediction of available potassium and available phosphorus,providing a reference method for subsequent research.In conclusion,this paper has built a soil nutrient content detection system based on pyrolysis and machine olfaction,and optimized its sensor array,olfactory feature space and prediction model,which can basically achieve quantitative analysis and prediction of soil nutrient content,overcome the detection problem of"high cost and long detection cycle"of chemical methods,and break through the spectral analysis vulnerable to water The bottleneck of low accuracy is caused by the influence of iron oxide and other factors.This provides a fast,efficient,and economical new method for detecting the main nutrient content in soil.更多还原