The extension of poisonous plants in alpine meadows is one of the main problems of the grassland ecosystem in the Qinghai-Tibet Plateau. The classification technology of poisonous plants in alpine meadows is of great significance for timely monitoring, scientific preventing and controlling changes in grassland communities. In recent years, poisonous plants species and harmful areas have increased rapidly. Traditional manual field surveys were time-consuming and laborious, and poorly represented the survey results. At the same time, poisonous plants have certain differences in geographical distribution, so it is not easy to conduct large-scale investigations by the workforce. Hyperspectral remote sensing technology has great advantages in the fine classification of poisonous plants due to its high resolution, multiple bands, integration of maps, and so on, which can meet the needs of fast, accurate, and large-scale acquisition of poisonous plants. Some scholars have carried out studies on the spectral reflectance characteristics of grassland plants, which proved that the spectral reflectance characteristics of plants could effectively distinguish their species. On the contrary, there are few reports on the selection of spectral reflectance characteristics variables of poisonous plants and the construction of a predictive classification model based on the spectral characteristics of poisonous plants. In this study, 11 kinds of main poisonous plants field spectrum data on alpine meadows, including Oxytropis ochrocephala, O latibracteata, Astragalus polycladus, Saussurea hieracioides, Ligularia virgaurea, Anaphalis lactea, Cirsium souliei Stellera chamaejasme Elsholtzia Densa Aconitum gymnandrum, and Pedicularis cheilanrthifolia (in Tianzhu County and Maqu County, Gansu Province) were collect by using the SOC710VP near-infrared hyperspectral imager. The Savitzky-Golay convolution smoothing algorithm (SG) was applied to denoise the original spectral values, the first-order differential derivative (FD) was used to carry out spectral feature analysis, and the canonical discriminant analysis (CDA) was performed to sort the absolute values of the standardized score coefficients of 16 selected spectral feature variables. Then from the size of large to small, they were added to 5 algorithms, namely random forest (RF), support vector machine-radial kernel function (SVM-RBF) k-nearest neighbor classification (KNN), naive bayes (NB), and decision tree (CART) to construct classification models and screen the best feature variables, and the confusion matrix was used to evaluate the classification effects. The results showed that: (1) The overall classification accuracy of canonical discriminant analysis (CDA) for 16 spectral characteristic variables was 92. 34%, R-2 = 0. 89; (2) The best classification spectral characteristic variables were selected as green peak amplitude (Mg), blue edge area (Ab), red edge amplitude (Mre), red edge area (Are), red edge position (Lre), NDVI2, and RVI1; (3) The selected 7 spectral characteristic variables were used to classify poisonous plants, and then the SVM-RBF has the best classification effects, with an accuracy of 96.45%.
