【作者】 李芳；

【导师】 由天艳；

【作者基本信息】 江苏大学 ， 农业工程， 2024， 博士

【摘要】 土壤重金属污染会对农作物、环境和人体健康造成严重危害,如降低土壤生产力、破坏土壤生态功能、影响农产品安全和质量、引发各种慢性或急性中毒和疾病等。土壤重金属检测能够及早发现和评估土壤中重金属的污染程度,了解土壤质量和环境状况,为推动土壤重金属污染治理提供科学依据和数据支持。常用的土壤重金属分析方法包括原子光谱法、电感耦合等离子体质谱法、电化学法、X射线荧光(X-ray fluorescence,XRF)光谱法和红外光谱法等。原子光谱法、电感耦合等离子体质谱法准确度高、精密度好,但前处理复杂、检测设备昂贵且操作复杂;电化学法仪器简单,但同样存在样品前处理复杂的问题;XRF光谱法和红外光谱法前处理简单、仪器便携、可无损检测,是土壤重金属现场筛查的理想工具,然而其预测结果的精准度有待提高。为了快速获得土壤重金属含量,提高现场分析结果的灵敏度和可靠性,建立高性能的分析模型并配套开发相关软件是很有必要的。基于此,本论文从提高土壤重金属定量分析模型的准确度入手,引入神经网络、深度学习等算法,结合光谱融合技术,建立了一系列定量分析模型,以实现土壤重金属的灵敏检测,检测的土壤重金属包括:砷(As)、镉(Cd)、铬(Cr)、铜(Cu)、镍(Ni)、铅(Pb)和锌(Zn);同时构建了光谱传递模型以保障模型和数据的通用性;最终,开发了土壤重金属光谱定量分析软件,成功应用于农田土壤重金属的现场、快速、准确检测。主要研究内容如下:(1)基于多层感知机和弹性网络回归的XRF定量模型测定土壤重金属:为了提高土壤重金属快速检测结果的准确度,考察不同建模算法的效果。在重金属元素特征峰能量范围内筛选特征,提高建模效率,更全面地利用XRF激发的特征信息;探讨了每种元素最佳预处理方法;经模型结构设计、参数优化后建模。选用多层感知机(MLP)和弹性网络回归(ENR)建模,并与不同方法对比,结果表明基于ENR建立的模型性能最优。基于ENR的XRF土壤重金属定量模型具有良好的预测性能,为XRF检测土壤重金属提供了一种新的预测模型。(2)基于SPA特征提取的XRF、NIR(near-infrared)、MIR(mid-infrared)光谱融合定量模型测定土壤重金属:为了进一步提高模型预测精度,基于工作(1)优选的ENR算法,引入近红外(NIR)和中红外(MIR)光谱及数据融合技术,建立融合模型。对比PCA、GA、SFLA、SPA和CARS五种特征提取算法,SPA特征提取结果最好。分别建立单光谱、数据级融合和特征级融合模型,最终模型预测准确度:特征级融合>数据级融合>单光谱。基于SPA提取的XRF-MIR特征级融合模型预测准确度最高,为估算土壤重金属元素含量提供了一种新的建模方案。(3)基于SSA优化CNN的XRF-MIR定量模型灵敏检测土壤重金属:为了提高预测结果的灵敏度,增加模型的实用性,针对GB 15618-2018中规定的风险筛选值以下含量样品建模。利用工作(2)中基于SPA的XRF-MIR光谱融合策略,引入卷积神经网络(CNN)算法建模。设计优化了各元素的网络结构和初始参数,并采用麻雀搜索算法(SSA)对网络的批大小、学习率和正则化系数自动寻优。最终建立的SSA-CNN模型预测模型灵敏度高、可靠性强,可用于土壤重金属灵敏分析,为土壤重金属污染状况监测和评估提供了光谱融合新方法。(4)基于SSA优化CNN的光谱传递模型构建:为了增加光谱数据和模型的普适性,提高数据利用率和工作效率,建立基于SSA-CNN的光谱传递模型。对比DS、PDS、PA和SSA-CNN四种光谱传递算法,最佳方法为SSA-CNN。设计CNN网络结构,确定网络参数,并利用SSA自动寻优后,建立光谱传递模型。分别利用SSA-CNN光谱传递后矩阵和XRF主机光谱矩阵建模,对两组模型预测结果进行准确性评估和Bland-Altman一致性分析,结果显示所有预测数据点均落在95%置信区间内,预测结果间一致性程度高。基于SSA-CNN的光谱传递模型具有良好的准确性、实用性,为构建光谱传递模型提供了新思路。(5)土壤重金属光谱分析软件设计及在农田土壤重金属现场检测中的应用:为了方便模型推广、现场快速分析数据,开发了土壤重金属光谱定量分析系统,将光谱处理和数据分析一体化,并可视化呈现结果。单个样品的“采样-检测-预测”过程可以在20 min内完成。将该软件结合便携式分析仪器,在实际农田土壤重金属的检测中进行了应用验证,各元素的相对误差:Cd>20%,As接近10%,其余元素均小于10%,表明除Cd外,软件具有良好的现场预测性能,解决了土壤重金属检测中现场分析数据困难和准确度低的难题。开发的软件可以同时处理XRF和MIR两种光谱信号,填补了相应领域中的软件空白。更多还原

【Abstract】 Heavy metal contamination of soil can cause serious harm to crops,environment and human health,such as reducing soil productivity,destroying soil ecological functions,affecting the safety and quality of agricultural products,and triggering a variety of chronic or acute poisoning and diseases.Soil heavy metal measurement can detect and assess the degree of heavy metal contamination in soil at an early stage,understand the soil quality and environmental conditions,and provide scientific basis and data support for decision-making on soil pollution control and environmental protection.Commonly used methods for analyzing heavy metals in soil include atomic spectrometry,inductively coupled plasma mass spectrometry(ICP-MS),electrochemical method,X-ray fluorescence spectrometry and infrared spectrometry.Atomic spectrometry and ICP-MS are accurate and precise,but the pretreatment is complicated,and the detection equipment is expensive and complicated to operate.The apparatus of electrochemical method is simple to use,but also has the problem of complicated sample pretreatment.X-ray fluorescence spectroscopy and infrared spectroscopy are ideal tools for in-situ screening of soil heavy metals due to their simple preprocessing,portable instruments and non-destructive testing.However,the accuracy of their prediction results needs to be improved.In order to obtain the soil heavy metal contents quickly and improve the sensitivity and reliability of the field analysis results,it is necessary to establish high-performance analytical models and develop supporting software.Based on this,this paper aims to improve the accuracy of soil heavy metal quantitative analysis models by introducing algorithms such as neural networks and deep learning,combined with spectral data fusion technology,to establish a series of quantitative analysis models and achieve sensitive detection of soil heavy metals.The detected soil heavy metals include arsenic(As),cadmium(Cd),chromium(Cr),copper(Cu),nickel(Ni),lead(Pb),and zinc(Zn).The spectral transfer models were also constructed to guarantee the universality of the models and spectral data.Finally,the quantitative analysis software of soil heavy metal was developed,which was successfully applied to the on-site and rapid detection of heavy metals in farmland soil.The specific work of this thesis included the following five parts:(1)In order to improve the accuracy of the rapid detection results of soil heavy metals,multilayer perceptron(MLP)and elastic network regression(ENR)were selected for modeling separately to investigate the effects of different methods on the prediction results.Feature selection within the characteristic peak energy ranges of heavy metal elements improves the modeling efficiency and more comprehensively utilizes the characteristic information of XRF excitation.The optimal preprocessing method for each element was explored for different heavy metals.After the model structure design and parameter optimization,modeling was carried out.The results indicated that the ENR model outperformed the MLP model.The ENR-based soil heavy metals quantitative models obtained by XRF have good prediction performance,which provides a new prediction model for the detection of heavy metals in soil by XRF.(2)In order to further improve the prediction accuracy of the quantitative model,based on the ENR algorithm,NIR(near-infrared)and MIR(mid-infrared)infrared spectral signals and spectral data fusion technology were introduced to establish spectral fusion models.Five methods,PCA,GA,SFLA,SPA and CARS,were used for feature extraction,and the results showed that the analytical model built based on SPA feature extraction had the best performance.The single sensor models,data-level fusion models and feature-level fusion models were established.Comparing various evaluation indicators,the accuracy of model prediction results was found to be:feature-level fusion>data-level fusion>single sensor models.The XRF-MIR fusion models based on SPA feature extraction showed the best prediction accuracy,which provides a new modeling scheme for estimating metal elements in soil.(3)In order to improve the sensitivity of quantitative analysis results and enhance the practicality of the models in actual soil screening work,quantitative prediction models were established for samples below the risk screening value specified in GB 15618-2018.The XRF-MIR feature-level fusion methods based on SPA feature extraction were utilized,and quantitative analysis models were established by introducing the CNN algorithm.After data preprocessing,feature extraction,and dataset division,the CNN network structure,convolutional kernel size,initial learning rate,and descent gradient were designed and optimized for each metal element,and the SSA intelligent optimization algorithm was used to automatically seek the optimization of the batch size,learning rate,and regularization coefficient of the network.The SSA-CNN-based spectral fusion quantitative analysis model is highly sensitive and can be used for rapid and accurate quantitative analysis of soil heavy metals,which provides a new spectral fusion method for monitoring and evaluating soil heavy metal pollution.(4)In order to increase the universality of the spectral data and the constructed models,and to improve the data utilization and work efficiency,a spectral transfer model based on SSA-CNN was established.Four different spectral transfer algorithms were compared,the best spectral transfer method was preferred to be SSA-CNN.The CNN network structure was designed,the network parameters were determined,and the spectral transfer model was built after automatic optimization using SSA for batch size,learning rate and regularization coefficient.Quantitative analysis models were established using the SSA-CNN spectral transfer posterior matrix and the master XRF spectral matrix,respectively.The accuracy of the prediction results was evaluated and the Bland--Altman consistency analysis was performed.The results showed that all predicted data points fell within the 95%confidence intervals,which indicates a high degree of consistency between the two sets of prediction results.The spectral transfer model based on SSA-CNN has good accuracy and practicability,which provides a new idea for constructing spectral transfer model.(5)In order to facilitate the promotion of the established models,the rapid on-site analysis of data,the data processing means and modeling methods involved in this thesis were synthesized to develop a quantitative analysis software for soil heavy metals,and presented the relevant results in a visual way.The final software had six functional modules:main program,spectral preprocessing,feature extraction,quantitative analysis,model transfer,and content prediction.Simply click the corresponding button and select the appropriate method to achieve prediction.The"sampling-detection-prediction"process for a single sample can be completed in 20 minutes.The software combined with portable analytical instruments,were applied in the detection of soil heavy metals in actual farmland.Relative errors for each element:Cd>20%,As about 10%,Cr,Cu,Ni,Pb,Zn were less than 10%.The results showed that the software had high completion and good predictive performance,which solved the problems of difficult on-site data analysis and low accuracy in soil heavy metals detection.The software developed can process both XRF and MIR spectral signals and be used for on-site measurement,filling the software gap in the corresponding field.更多还原