Research On Discrimination Of Varieties Of Invasive Weeds Based On Visible And Near-Infrared Spectroscopy

Biological invasion is recognized as one of the greatest threats to the ecological environment worldwide. And in China, the majority of the invasive species are invasive plants. Generally speaking, there are many species of invasive and native plants with similar morphological characteristics. However, commonly used methods for identifying these plants are to some extent time-consuming,costly and experience-demanding. Therefore, developing a convenient and accurate method for identifying these invasive plants is necessary, which is very helpful for the early prediction and prevention of the invasive plants.In this paper, identifications of seven varieties of invasive weeds were studied based on Vis/NIR spectroscopy technique and pattern recognition methods (BPNN and SVM). The main contents were listed as follows:(1) Four varieties of invasive weeds, Conyza canadensis (L.)Cronq, Erigeron annuus (L.)Pers., Erigeron philadelphicus L. and Conyza bonariensis(L.)Cronq, were employed. And 60 X4 samples were obtained. For each variety of weed,40 samples were randomly selected as the training set, and the remaining 20 samples were used as the prediction set. Ten types of spectral data pretreatment methods were used to pretreat the spectral data. Based on principal component analysis, two types of pattern recognition methods (BPNN and SVM) were applied to build recognition models for identifying these invasive weeds.20 recognition models (10 BP models and 10 SVM models) were built and their discrimination accuracies were evaluated. And the results showed that:1) Recognition ratio of 97.5% was achieved by 5 SVM models combined with the spectral data pretreatment methods of MSC,SNV,moving average smoothing(15 points),normalization and SNV+Detrending respectively; 2) The BP model combined with the pretreatment method of normalization displayed the best performance of the 10 BP models, and its recognition ratio was 93.75%; 3) The classification accuracies of the 10 SVM models were higher than that of the 10 BP models when using the same input variables. In order to simplify the recognition models and to improve the efficiency of developing them,5 groups of characteristic wavelengths were selected to build 6 recognition models (5 SVM models and 1 BP model). And the best recognition ratio of the 5 SVM models was 96.25% and 85% was achieved by the BP model. These results showed that it was feasible to use characteristic wavelengths as input variables to build recognition models, which was an effective approach to simplify the recognition models.(2) Another three varieties of invasive weeds (Amaranthus retroflexus L.,Amaranthus tricolor L. and Amaranthus viridis L.) were employed for identification. And 50×3 samples were obtained. For each variety of weed,30 samples were randomly selected as the training set, and the remaining 20 samples were used as the prediction set. Ten types of spectral data pretreatment methods were employed for pretreating the spectral data. Based on principal component analysis, two types of pattern recognition methods (BPNN and SVM) were applied to establish recognition models for identifying these invasive weeds. And the results showed that:1) The best recognition ratio of 98.33% was achieved by the SVM model combined with the first derivative spectra; 2) The BP model combined with the first derivative spectra displayed the best performance of the 10 BP models, and its recognition ratio was 90%; 3) Every one of the 10 SVM models gave a better performance than the BP model with the same input variables. Based on the laodings of the PCs,13 characteristic wavelengths were selected to build 2 recognition models (1 SVM model and 1 BP model). And the recognition ratio of the SVM model was 100%, and 93.33% was achieved by the BP model. These results showed that the recognition ratios of the models built by the characteristic wavelengths were higher than that of the models built by the full waveband, which indicated that the selected characteristic wavelengths reflected the main characteristics of the three varieties of weeds, and could improve the accuracies of the recognition models.