Classification of major species in the sericite–Artemisia desert grassland using hyperspectral images and spectral feature identification

Sample plot establishment

In the study area, a 500 m long transect was established every 150 m, with a total of six parallel transects. Five 30 m × 30 m transects with a spacing of 30 m were arranged along each transect. Five transects were arranged on two diagonal lines of the transects with a 5-point sampling method as the data acquisition points. According to the field of view of the instrument, the area of the transect was set as 0.5 m × 0.6 m, with a total of 150 transects (Fig. 2).

Hyperspectral image acquisition

An SOC710VP imaging spectrometer was used to collect spectral data of plant communities in the field (spectral range, 400∼1,000 nm; resolution, 4.68 nm; number of bands, 128; and imaging resolution, 1,392 × 1,040 pixels), and the spectral measurement time ranged from 11:00 to 16:00 in April, June and September 2018 (solar altitude angle > 45°). The measurements were taken on sunny, cloudless, and windless or less windy (wind < 3 m/h) days. During measurement, the lens was vertically downwards, and the vertical height from the plant canopy was approximately 1.0 m. Based on the analysis of the spectral characteristics of the three main plants, the other plants distributed sporadically in the quadrat were extracted, and the community canopy spectrum and bare-ground hyperspectral images were collected in each quadrat. A total of 450 images were collected. In addition, the coverage and biomass of each species in the quadrat were measured to provide a reference for the extraction of spectral data.

Extraction of spectral data

The spectral data of pure pixels of S. transiliense, C. arenarius, P. sibirica and bare ground were extracted from the images of plant communities in each measurement period by using SRAnal 710 software of the imaging spectrometer. A total of 150 quadrat images were obtained in each period, yielding 150 reflectance samples of three types of plants and bare land from which the damaged samples were removed. To ensure the consistency of the reflectance samples, 90 reflectance samples of each type were ultimately classified: 360 reflectance samples in April (spring), 273 reflectance samples in total due to the scarcity of P. sibirica in June (summer), and 184 samples in total due to the scarcity of C. arenarius in September (autumn).

First-order differential processing

The first-order differential can eliminate the influence of partial atmospheric and soil backgrounds, highlight the characteristics of the vegetation spectrum, and be beneficial to the extraction of vegetation information such as vegetation indices and the leaf area index.

The first-order differential method was selected to process the spectrum, with the following formula: (1)${\displaystyle FR\left({\lambda }_i\right)=\frac{R\left({\lambda }_{i+1}\right)-R\left({\lambda }_{i-1}\right)}{2\Delta \lambda }}$ where FR(λ_i) is the first-order differential spectrum of wavelength λ_i; R(λ_i+1) is the original spectral reflectance of wavelength i+1; R(λ_i−1) is the original spectral reflectance of wavelength i-1; and Δλ is the wavelength difference between wavelength i and wavelength i+1.

Selection of spectral parameters

To explore the spectral response characteristics of the main species and bare land of the S. transiliense desert grassland, 10 common spectral characteristic parameters were selected, including eight location parameters, two area parameters and 11 vegetation indices, including seven wide-band vegetation indices and four narrow-band vegetation indices, as shown in Tables 2 and 3.

Fisher discriminant analysis is used to establish a linear discriminant function based on the original information of samples of known categories or the feature space information that can represent the information of samples of known categories, project the multidimensional spatial space to one-dimensional space, classify and discriminate the samples of unknown categories through the discriminant criteria, and then verify the classification accuracy of the sample classification information (Huang et al., 2019). (2)${\displaystyle S_{\omega }=\sum _{i=1}^C\sum _{X\in X_i}\left(x-m_i\right){\left(x-m_i\right)}^T,i=1,\dots ,C}$ (3)${\displaystyle S_b=\sum _{i=1}^C\left(x-m_i\right){\left(x-m_i\right)}^T}$ (4)${\displaystyle \overline{s_{\omega }}=\sum _{i=1}^C\sum _{y\in Y_i}{\left(y-\overline{m_i}\right)}^2,i=1,\dots ,C}$ (5)${\displaystyle J_F\left(\omega \right)=\frac{{\omega }^T{S}_b\omega }{{\omega }^T{S}_{\omega }\omega }}$ (6)${\displaystyle L\left(\omega ,\lambda \right)={\omega }^T{S}_b\omega -\lambda \left({\omega }^T{S}_{\omega }\omega -c\right)}$ (7)${\displaystyle S_b{\omega }^{\ast }=\lambda S_{\omega }{\omega }^{\ast }}$ where m_i is the mean vector of each class of samples in the original high-dimensional space; $\overline{m_i}$ is the mean vector of each class of samples in the one-dimensional space Y after projection; m is the mean vector of all of the samples; X is the sample point vector; S_ω is the discrete matrix within the sample class; S_b is the discrete matrix between sample classes; $\overline{s_{\omega }}$ is the pooled within-class scatter matrix; J_F(ω) is the Fisher criterion function; L(ω, λ) is the Lagrange function; λ is the Lagrange multiplier; and ω^∗ is the maximum value, which is the best projection method.

Screening of sensitive spectral parameters

Variance analysis and a discriminant model were used to screen the eight location parameters, two area parameters and 12 vegetation indices, and the, contributions of these parameters to the identification of the four objects were determined. The P value between two identified objects was determined by variance analysis. The smaller the P value is, the more significant the difference is. All 10 spectral characteristic parameters and 12 vegetation index parameters were taken as input parameters to obtain the initial discrimination accuracy. The characteristic parameters were removed one by one according to the P value. If the discrimination accuracy of the model after removing the parameters was greater than or equal to the initial accuracy, the parameter was removed; otherwise, the parameter was retained in the discrimination model (Fig. 3).

Data analysis

The spectral reflectance values of different recognition objects were extracted using SRAnal710 and imported into Excel 2016 to remove abnormal data. First-order differential processing was performed, and feature parameters and vegetation indices were calculated. Principal component analysis was performed in SPSS 20.0 software to screen feature bands, and significance tests were conducted on the feature parameters of each identified object and vegetation index. Fisher discriminant analysis was used to classify samples and verify accuracy. SigmaPlot 14.0 software was used for plotting.

Discriminant parameter analysis

In April, there were no significant differences in the characteristic parameters among the identified objects, which were L_b, L_re, L_g and A_re. For June, the parameters were L_b, L_re, L_g, M_g, M_r, M_re and A_b, and for September, they were L_b, L_re, M_b, L_g, M_g, M_r, M_re and A_b (Table 6).

In April, there were no significant differences in the vegetation indices among the identified objects, which were the DVI, NDGI, RI, PRI, PSRI, and VOG. For June, the measured indices were the PRI and RENDVI, and for September, they were the RVI, BNDVI, GNDVI and PRI (Table 7).

Discriminant parameter screening

Since the average P value can indirectly reflect the size of the difference, the parameters that did not have very significant differences were eliminated one by one from small to large according to the P value (P < 0.01), and the discrimination accuracy of the four recognition objects and the total discrimination accuracy of the test set were calculated. The group with more eliminated parameters and an accuracy no less than that of the noneliminated parameters was taken as the optimal set of parameters. The eliminated parameters and retained parameters of sericite, S. transiliense, C. arenarius, P. sibirica and bare land in April, June and September are shown in Tables 8 and 9.

Characteristic parameter discrimination

In April, 1,200 samples of the three plants and bare land were divided into a training set and a test set at a ratio of 7:3. Taking the selected spectral characteristic parameters as input variables, a discrimination model was established, and the recognition effect is shown in Fig. 4. Discriminant analysis of S. transiliense (A), C. arenarius (B), P. sibirica (C) and bare land (D) was performed. When classifying S. transiliense and the other three objects, S. transiliense was incorrectly identified 17 times, with an accuracy of 81.11%. The other objects were incorrectly identified as S. transiliense 102 times, with an accuracy of 62.22%, yielding a total accuracy of 71.67%. When classifying C. arenarius with the other three objects, C. arenarius was incorrectly identified six times, with an accuracy of 95.56%, and other objects were incorrectly identified as C. arenarius 51 times, with an accuracy of 79.26%, yielding a total accuracy of 87.41%. When classifying P. sibirica and the other three objects, the other objects were incorrectly identified as P. sibirica 52 times, with an accuracy of 82.96%, yielding a total accuracy of 85.93%. When classifying the bare land and the three plants, the bare land and all plants were classified into the correct category with an accuracy of 100%.

In June, 910 samples of the three plants and bare land were divided into a training set and a test set at a ratio of 7:3. The test set samples of S. transiliense, C. arenarius, P. sibirica and bare land were classified. When classifying S. transiliense and the other three objects, S. transiliense was incorrectly identified four times, with an accuracy of 95.56%. The other objects were incorrectly identified as S. transiliense three times, with an accuracy of 98.36%, yielding a total accuracy of 96.96%. When classifying C. arenarius with the other three objects, C. arenarius was incorrectly identified 15 times, with an accuracy of 83.33%, and the other objects were incorrectly identified 32 times as C. arenarius, with an accuracy of 82.51%, yielding a total accuracy of 82.92%. When classifying P. sibirica and the other three objects, P. sibirica was incorrectly identified once, with an accuracy of 66.66%, and the other objects were incorrectly identified three times as P. sibirica, with an accuracy of 98.89%, yielding a total accuracy of 82.78%. When classifying the bare land and the three plants, the bare land and all plants were placed in the correct category, and the other objects were incorrectly identified as bare land 32 times, with an accuracy of 82.51%, yielding a total accuracy of 91.26% (Fig. 5).

A total of 614 samples of three plants and bare land were divided into a training set and a test set at a ratio of 7:3. When classifying S. transiliense and the other three objects, S. transiliense was incorrectly identified zero times, with an accuracy of 100%, and the other objects were incorrectly identified as S. transiliense zero times, with an accuracy of 100%, yielding a total accuracy of 100%. When classifying C. arenarius with other objects, C. arenarius was incorrectly identified once, with an accuracy of 75%, and the other objects were incorrectly identified as C. arenarius 56 times, with an accuracy of 68.89%, yielding a total accuracy of 71.95%. When classifying the bare land and the other two plants, all the bare land was placed in the correct category, and the other objects were incorrectly identified as bare land four times, with an accuracy of 95.74%, yielding a total accuracy of 97.37% (Fig. 6).

Vegetation index model and accuracy evaluation

In April, 1,200 samples of three plants and bare land were divided into a training set and a test set at a ratio of 7:3. The samples in the test set were classified as S. transiliense, C. arenarius, P. sibirica and bare ground. When classifying S. transiliense and the other three objects, S. transiliense was incorrectly identified 10 times, with an accuracy of 88.89%. The other objects were incorrectly identified as S. transiliense 57 times, with an accuracy of 78.89%, yielding a total accuracy of 83.89%. When classifying C. arenarius with the other three objects, C. arenarius was incorrectly identified seven times, with an accuracy of 92.22%, and the other objects were incorrectly identified as C. arenarius 46 times, with an accuracy of 82.96%, yielding a total accuracy of 86.30%. When classifying P. sibirica and the other three objects, P. sibirica was incorrectly identified three times, with an accuracy of 96.67%. The other objects were incorrectly identified as P. sibirica 38 times, with an accuracy of 85.93%, yielding a total accuracy of 91.30%. When classifying the bare land and the three plants, the bare land and all plants were placed in the correct category, with an accuracy of 100% (Fig. 7).

In June, 910 samples of three plants and bare land were divided into a training set and a test set at a ratio of 7:3. When classifying S. transiliense and the other three objects, S. transiliense was incorrectly identified zero times, with an accuracy of 100%. When classifying C. arenarius with the other three objects, C. arenarius was incorrectly identified as S. transiliense four times, with an accuracy of 95.56%, and the other objects were incorrectly identified as C. arenarius eight times, with an accuracy of 95.63%, yielding a total accuracy of 95.60%. When classifying P. sibirica and the other three objects, P. sibirica was incorrectly identified zero times, with an accuracy of 100%. Other objects were incorrectly identified as P. sibirica zero times, with an accuracy of 100%, yielding a total accuracy of 100%. When classifying the bare land and the three plants, the bare land and all plants were placed in the correct category, and the other objects were incorrectly identified as bare land 15 times, with an accuracy of 91.80%, yielding a total accuracy of 95.90% (Fig. 8).

In September, 614 samples of three plants and bare land were divided into a training set and a test set at a ratio of 7:3. When classifying S. transiliense and the other three objects, S. transiliense was incorrectly identified zero times, with an accuracy of 100%, and the other objects were incorrectly identified three times as S. transiliense, with an accuracy of 96.81%. The total accuracy was 98.41%. When classifying C. arenarius and other objects, C. arenarius was incorrectly identified once, with an accuracy of 75%, and the other objects were incorrectly identified as S. transiliense 25 times, with an accuracy of 86.11%, yielding a total accuracy of 80.56%. When classifying bare land and the other two plants, bare land was incorrectly identified five times, with an accuracy of 94.44%, and the other objects were incorrectly identified three times as bare land, with an accuracy of 96.81%, yielding a total accuracy of 95.63% (Fig. 9).