Application of electronic nose in detection of cotton bollworm infestation at an early stage

Cotton bollborm is one of the main pests of cotton. Cotton is under threat of yield loss and poor quality because of the cotton bollworm. However, cotton bolworms tend to hide in the cotton plants so that there are limitations for conventional detection methods, such as acoustic signal method, image recognition method and spectral imaging technology. A lot of researches have shown that volatile organic compounds (VOCs) released by plants will change when they are attacked by pests. So it is possible to get the cotton bollworm damage information by detecting the volatiles. Currently, gas chromatograph-mass spectrometer (GC-MS) can accurately detect the composition and content of volatile matter. However, this method has some disadvantages in practical application, such as time-consuming, high cost and inconvenience. The electronic nose is composed of sensor array, which is an instrument to analyze, identify and detect most of the volatiles. In this study, electronic nose was used to detect the cotton plants infested with cotton bollworm of different amounts at an early stage. The volatile organic compounds (VOCs) in cotton were analyzed by GC-MS. The plant height of cotton used in the study was 50-70 cm, and the boll period was about 12 weeks. Cotton bollworms used in the study were at second-instar. The VOCs emitted by the undamaged and damaged cotton plants detected by GC-MS were different, which indicated that electronic nose had potential in the application of cotton bollworm detection. The curve of electronic nose sensor was obtained for cotton plants infected by different numbers of cotton bollworm. Then five kinds of feature parameters were extracted from the curves of electronic nose sensors : stable value, area value, mean differential value, wavelet energy value and the coefficients of the fitted quadratic polynomial function. Feature parameters were selected based on three kinds of neural network methods: multilayer perceptron neural network (MLPNN), radial basis neural network (RBFNN) and extreme learning machine (ELM). Then stable value, area value and mean differential value were selected because of their better classification performance among the five kinds of feature parameters. Multiple-features were combinations of single-features. The classification analysis was carried out based on multiple-features and three kinds of neural network methods. And support vector machine regression (SVR) models were established based on single-features and multiple-features, respectively. The results showed that the classification performance of multiple-features was better than that of single-features. The classification performance was best based on "stable value and mean differential value" features and ELM. The classification accuracy of training set and test set based on "stable value and mean differential value" features were both 100%. The regression models based on multiple-features were better than that based on single-features. And the regression model was the best based on "area value and mean differential value" features. The coefficient of determination (R2) and root mean square error (RMSE) of the regression model based on the training set of "area value and mean differential value" were 0.994 0 and 0.086 0. The R2 and RMSE of the regression model based on the test set of "area value and mean differential value" were 0.923 0 and 0.370 9. The results show that feature election and multiple-features can improve the classification performance of the electronic nose for infested cotton plants. It can be concluded that electronic nose has strong potential for the application of detection of cotton plants infested with cotton bollworm at an early stage.