Hyperspectral classification of main features of Seriphidium transiliense desert grassland based on different machine learning algorithms

Machine learning algorithms are widely used in the field of spectral classification. Different algorithm models directly affect the classification effect of ground objects. In this study, three main features of Seriphidium transiliense, Ceratocarpus arenarius, and soil in Seriphidium transiliense desert grassland were used as classification objects. Hyperspectral data of grassland vegetation communities were collected during the vigorous growth period of vegetation. The differences in spectral reflectance of different features were analyzed, and the characteristic bands were selected and substituted into the best index factor (OIF) to synthesize false color images. Three different machine learning algorithms, such as random forest (RF), support vector machines (SVM), and artificial neural network (ANN), were used to establish classification models. The results revealed that : 1) The spectral reflectance of plants showed an inverted ' U ' trend in the visible light band and began to rise sharply in the near-infrared band, indicating a 'red edge' phenomenon. The change trend of soil spectral reflectance was relatively stable but gradually increased with an increase in wavelength. 2) The best classification band combination calculated using OIF was: 499.69, 535.78, 633.28 nm, and the OIF value was 0.10. 3) The overall classification accuracy of the three different machine learning algorithms was greater than 90%. The random forest algorithm classification model had the highest accuracy, with an overall accuracy of 97.54% and a Kappa coefficient of 0.95. The classification accuracies of the three types of ground objects under different algorithms were as follows: soil > Seriphidium transiliense > Chenopodium album. In general, the random forest algorithm had the best classification effect on the main features of desert grassland.