基于多传感器人工嗅觉系统的土壤有机质含量检测方法

Detection Method of Soil Organic Matter Based on Multi-sensor Artificial Olfactory System

摘要:

为了实现对土壤有机质含量的快速、方便、准确测量，本文提出了一种基于多传感器人工嗅觉系统的土壤有机质含量检测方法。选取10个不同型号的氧化物半导体式气体传感器组成传感器阵列，并采用不同浓度的硫化氢、氨气和甲烷等标准气体对传感器阵列进行了响应测试，从响应曲线可以看出，传感器阵列对不同浓度、种类的标准气体皆有响应且响应结果不同，随着标准气体浓度的增大传感器阵列的响应曲线也随之上升，表明传感器阵列具有较高的特异性和一定的交叉敏感性。提取每个传感器土壤气体响应曲线上的响应面积、最大值、平均微分系数、方差、平均值和最大梯度6个特征构建人工嗅觉特征空间。采用偏最小二乘法回归（PLSR）、支持向量机回归（SVR）和BP神经网络（BPNN）算法建立人工嗅觉特征空间与土壤有机质含量关系的预测模型，使用决定系数（R2）、均方根误差（RMSE）和绝对平均误差（MAE）评估预测模型的性能。试验结果表明，PLSR、BPNN、SVR测试集的R2分别为0.80878、0.87179和0.91957，RMSE分别为3.6784、3.1614、2.4254g/kg，MAE分别为3.1079、2.4154、2.1389g/kg。SVR算法建立的模型R2最高，RMSE、MAE最小，比PLSR、BPNN具有更好的预测性能，可用于土壤有机质含量的测量。

Abstract:

In order to achieve a rapid, convenient and accurate measurement of the soil organic matter (SOM) content, a multi-sensor array based on artificial olfactory technology was designed to detect SOM content. Totally ten different types of oxide semiconductor gas sensors were selected to form the sensor array, and different concentrations of hydrogen sulfide, ammonia and methane standard gases were used to test the response of the sensor array. It can be seen from the response curve that the sensor array responded to different concentrations and types of standard gases, and the response results were not the same. With the increase of the standard gas concentration, the response curve of the sensor array was also increased. The test results showed that the sensor array had high specificity and certain cross-sensitivity. And from the soil gas response curve of each sensor, six features (response area value, maximum value, average differential coefficient, variance, average value and maximum gradient) were extracted to construct an artificial olfactory feature space (AOFS). Then, partial least square regression (PLSR), back propagation neural network (BPNN) and support vector machine regression (SVR) algorithms were used to establish the prediction model of AOFS and SOM content, and the coefficient of determination (R2), root mean square error (RMSE) and absolute average error (MAE) were used to evaluate the performance of the prediction model. The test results showed that the R2 of the PLSR, BPNN, and SVR test sets were 0.80878, 0.87179 and 0.91957, the RMSE were 3.6784g/kg, 3.1614g/kg and 2.4254g/kg, and the MAE were 3.1079g/kg, 2.4154g/kg and 2.1389g/kg, respectively. The model established by the SVR algorithm had the highest R2, the smallest RMSE and MAE. It had higher predictive performance than PLSR and BPNN, and can be used for the measurement of SOM content.