Original papersDiscrimination among tea plants either with different invasive severities or different invasive times using MOS electronic nose combined with a new feature extraction method
Graphical abstract
Introduction
Tea with great flavor and high content of beneficial substances (Suzuki et al., 2016), is the most widely consumed beverage aside from water (Vernarelli and Lambert, 2013). Especially in China, tea has been used as a daily beverage and crude medicine for thousands of years (Yen and Chen, 1995). Tea plant (Camellia sinensis) is grown in about 60 countries (Ma et al., 2016) and an important crop in many countries, such as China, Japan, Kenya and India (Dong et al., 2011, Ramya et al., 2013), etc. The tender leaves tea plant are the raw materials of tea (Saravanakumar et al., 2007). The health of tea plant is therefore a crucial factor for producing high yield and high quality of tea. However, many pests attack tea leaves during tea plant growth, causing 5–55% losses in tea production and 0.5 billion–1 billion dollars in economic losses (Hazarika et al., 2009). Tea geometrid (Ectropis obliqua, a chewing insect) is one of the most common pests of tea plant in China and causes severe damage by feeding on tea leaves (Ma et al., 2016, Wang et al., 2016). In this work, the damage caused by Ectropis obliqua was studied. Invasive severity and invasive time are two important parameters for pest damage detection. Hence, in this study, two aspects, including tea plants with different invasive severities and with different invasive times, were discussed for giving a comprehensive information of pest damage.
A lot of researches about the reaction of plant to pests have been published and the results showed that volatile organic compounds (VOCs) emitted by plant change when the plant is attacked by pests (Henderson et al., 2010, Holopainen and Gershenzon, 2010, Snoeren et al., 2007, Yu et al., 2008b). Some researches also showed response changes of tea plant when exposed to multiple stresses (Fu et al., 2015, Mei et al., 2016), and VOCs vary both quantitatively and qualitatively with infestation duration and herbivore density (Cai et al., 2014). Electronic nose (E-nose) (Kiani et al., 2016, Markom et al., 2009), which is a nondestructive technique that mimics the human olfactory system detecting based on volatiles emitted by samples, is therefore possible to detect pest damage (Lampson et al., 2014). Besides, it is an instrument embedded with a chemical sensor array of partial specificity cooperating with appropriate pattern recognition algorithms used for detection of chemicals (Gardner and Bartlett, 1999). Gas Chromatography-Mass Spectrometer (GC-MS), which is a precise detection technique and able to determine the constituents of volatiles and their contents (Ivanova et al., 2013), was also employed and combined with E-nose for giving a more reliable results.
Metal oxide semiconductor (MOS) gas sensors, which have the advantages of cross-sensitivity, broad spectrum response and low-cost, have been widely used in E-nose application (Zhang et al., 2014). The electrical signal of E-nose based on MOS sensors consists of a few hundred measured values, which are too many to be signals as input of analytical methods and exist data redundancy. Feature extraction, as a process of finding a small set of values that largely represented the whole signals, is constantly required (Carmel et al., 2003). Furthermore, feature extraction method is one of the key points of performance improvement of E-nose systems because feature extraction is the first step of the sensor signal processing (Gardner et al., 2005). The feature extraction methods from original response curves, such as maximum value, average value, maximum slope, maximum variance, standard deviation and root mean square (Distante et al., 2002, Yu et al., 2008a), are commonly applied for E-nose because they are visual, intuitive and fast to compute. In particular, maximum value and average value are frequently-used features (Dai et al., 2015).
However, the feature extraction methods from original response curves leave some potential information carried in the signals unutilized. Although these methods have been used for many simple applications successfully, it is widely accepted that more sophisticated features are required when turning to more demanding tasks (Hong et al., 2014). Feature extraction methods based on curve fitting extract features depending on the whole dataset of E-nose and largely overcome this shortcoming. In curve fitting, the critical and complex problem is how to select and envision the specific form of the unknown function (Yan et al., 2015). The most commonly used models (Yan et al., 2015), such as polynomial functions, exponential functions, and Gaussian functions, are all nonlinear, which makes the fitting process complicated and long. A feature extraction method based on a new curve fitting model, a piecewise function, is introduced in this study. This model transforms the nonlinear data into linear first, and then linear equation is employed to fit the linear data and calculate the unknown parameters as features. Hence, this feature extraction method is simple and quick to compute.
In this work, E-nose was employed, and GC-MS was used as an auxiliary technique for proving the potential of E-nose in detecting tea plants either with different invasive severities or different invasive times. A new feature extraction method based on a piecewise function for MOS sensor response was proposed for detecting tea plants either with different invasive severities or different invasive times and its performance was compared with those of feature extraction methods based on the three commonly applied models-polynomial function, exponential function, and Gaussian function. Curve fitting figure and root mean squared error (RMSE) were introduced to evaluate the fitting performance of four curve fitting models, while correct classification rate obtained by multi-layered perceptron (MLP) was calculated for assessing classification performance as well as proving the advantage of E-nose in tea plant detection. The main objectives of this research are: (1) to prove the ability of E-nose in detecting tea plants either with different invasive severities or with different invasive times, (2) to propose a new feature extraction method for E-nose data analysis, (3) to explore if the new proposed method would outperform the feature extraction method based on conventional widely applied models.
Section snippets
Experimental sample and design
This study was carried out using the tea plant cultivar “clone Longjing43”, which was 20–30 cm high and had 10 leaves roughly for each plant, and pest (Ectropis obliqua) at the 3th larval stage was provided by Tea Research Institute, Chinese Academy of Agricultural Sciences. Each tea plant was cultivated in tea plantation (at field scale) in Hangzhou, China, and watered weekly. They were taken care of carefully making sure that each tea plant was healthy and undamaged. And the chosen tea plants
Results and discussion
GC-MS was employed first to detect tea plants with pest damage and for proving the difference of VOCs emitted by tea plants with either different invasive severities or different invasive times. E-nose was also applied to detect tea plants either with different invasive severities or different invasive times. The E-nose datasets were analyzed by feature extraction, feature selection and classification algorithm. Furthermore, four different curving fitting feature extraction methods, including
Conclusion
In this study, E-nose and GC-MS were employed to detect two aspects of tea plants with pest damage, including tea plants with different invasive severities and with different invasive times. The results of GC-MS proved the difference of VOCs emitted by tea plants with either different invasive severities or different invasive times and the possibility of E-nose in detecting them. For E-nose data analysis, feature extraction methods based on four curve fitting models were compared for choosing
Acknowledgements
The authors acknowledge the financial support of the Chinese National Foundation of Nature and Science through Project 31370555 and 31670654.
References (37)
- et al.
Power system voltage stability monitoring using artificial neural networks with a reduced set of inputs
Int. J. Electr. Power Energy Syst.
(2014) - et al.
A feature extraction method for chemical sensors in electronic noses
Sens. Actuat. B: Chem.
(2003) - et al.
Longjing tea quality classification by fusion of features collected from E-nose
Chemometr. Intell. Lab. Syst.
(2015) - et al.
On the study of feature extraction methods for an electronic nose
Sens. Actuat. B: Chem.
(2002) - et al.
Determination of (4-methylcyclohexyl) methanol isomers by heated purge-and-trap GC/MS in water samples from the 2014 Elk River, West Virginia, chemical spill
Chemosphere
(2015) - et al.
Enhancing electronic nose performance by sensor selection using a new integer-based genetic algorithm approach
Sens. Actuat. B: Chem.
(2005) - et al.
Detecting stink bugs/damage in cotton utilizing a portable electronic nose
Comput. Electron. Agric.
(2010) - et al.
Multiple stress factors and the emission of plant VOCs
Trends Plant Sci.
(2010) - et al.
Comparison of spectral clustering, K-clustering and hierarchical clustering on e-nose datasets: application to the recognition of material freshness, adulteration levels and pretreatment approaches for tomato juices
Chemometr. Intell. Lab. Syst.
(2014) - et al.
Application of electronic nose systems for assessing quality of medicinal and aromatic plant products: a review
J. Appl. Res. Med. Aromatic Plants
(2016)