摘要
为了实现土壤有机质快速、准确的测量,提出了一种基于人工嗅觉的土壤有机质含量检测方法。首先,由不同温度控制的10个气体传感器所构成的阵列对土壤样品气体进行采集;然后,提取每个传感器响应曲线上的7个特征(包括最大值、最小值、平均值、平均微分系数、响应面积、第30秒的瞬态值和第60秒的瞬态值),构建嗅觉特征空间;对特征空间优化后,采用回归算法建立预测模型。为减小不同测定算法、异常样本以及冗余特征对模型预测性能的影响,在应用蒙特卡罗抽样(Monte Carlo sampling,MCS)法剔除异常样本的基础上,采用主成分分析(Principal component analysis,PCA)法对特征空间进行降维处理,评估了包括偏最小二乘法回归(Partial least square regression,PLSR)、支持向量机回归(Support vector machine regression,SVR)和BP神经网络(Back propagation neural network,BPNN)等3种建模方法对土壤有机质含量的预测性能,选用决定系数R2、均方根误差(RMSE)和预测偏差比(RPD)评价各模型的预测性能。测试集验证结果表明,PLSR、SVR和BPNN这3种模型的预测值和样本的观测值之间的R2分别为0.86、0.91和0.85,RMSE分别为2.49、2.05、2.68 g/kg,RPD分别为2.49、3.02和2.32。SVR模型的预测性能高于PLSR模型和BPNN模型,可对土壤有机质含量进行准确预测。
In order to measure soil organic matter content quickly and accurately,a method based on artificial olfactory was proposed.Firstly,the response curves of soil gas were collected by an array composed of 10 gas sensors controlled at different temperatures.And then seven features,including the maximum value,minimum value,mean value,mean differential coefficient value,response area value,transient value at the 30th second and transient value at the 60th second were extracted from each sensor response curves to build an olfactory feature space.Finally,the prediction model was established by using the regression algorithm.To reduce the influence of different regression algorithms,abnormal samples and redundancy characteristics on the prediction performance of the model,the Monte Carlo sampling(MCS)method was used to eliminate abnormal samples,and the principal component analysis(PCA)method was used to reduce the dimension of olfactory feature space.Moreover,three modeling methods,including partial least square regression(PLSR),support vector machine regression(SVR)and back propagation neural network(BPNN),were used to predict soil organic matter content.And the predictive performance of each model were evaluated by coefficient of determination(R2),root mean square error(RMSE)and ratio of prediction derivation(RPD).The results showed that the R2 values of PLSR,SVR and BPNN were 0.86,0.91 and 0.85,respectively;the RMSE values were 2.49 g/kg,2.05 g/kg and 2.68 g/kg,respectively;and the RPD values were 2.49,3.02 and 2.32,respectively.The prediction performance of SVR model was higher than that of PLSR model and BPNN model,which can accurately predict the organic matter content.The results can provide a reference method for the prediction of soil organic matter.
作者
朱龙图
李名伟
夏晓蒙
黄东岩
贾洪雷
ZHU Longtu;LI Mingwei;XIA Xiaomeng;HUANG Dongyan;JIA Honglei(College of Biological and Agricultural Engineering, Jilin University, Changchun 130022, China;Key Laboratory of Bionic Engineering, Ministry of Education, Jilin University, Changchun 130022, China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2020年第3期171-179,共9页
Transactions of the Chinese Society for Agricultural Machinery
基金
国家重点研发计划项目(2016YFD070030201)
吉林省科技计划项目(20190302116GX)。
关键词
土壤有机质
人工嗅觉系统
蒙特卡罗抽样
特征降维
预测模型
soil organic matter
artificial olfactory system
Monte Carlo sampling
feature dimension reduction
prediction model
