Identification of plant species in an alpine steppe of Northern Tibet using close-range hyperspectral imagery

Abstract

The identification of plant species in alpine steppes of Northern Tibet is of great significance for revealing community structures and for monitoring vegetation degradation and restoration from remote sensing images. Plants in the alpine steppe of Northern Tibet are short, sparse, and highly heterogeneous in spatial distribution. This peculiarity makes the plant species identification here much more difficult than the identification of plants with high spatial homogeneity. We aimed to explore the potential of close-range hyperspectral imaging for plant species identification in alpine steppe under field conditions. Specifically, we assessed which spectral bands are effective and which classification methods are suitable for plant species identification. A close-range hyperspectral image of grassland in Nagqu, Tibet were acquired in August 2018. Four methods, including derivatives, continuum removal, spectral indices, and principal components were used to enhance the differences in spectral characteristics between plant species. Then, two band selection methods, including Mahalanobis distance and variable importance evaluations based on a random forest (RF) were used to reduce dimensionality and select indicators beneficial for identifying grass species. Four datasets were constructed based on those indicators and were used as the input data for four classifiers, support vector machine (SVM), RF, artificial neural network (ANN), and spectral angle mapper (SAM). We found that (1) bands selected using Mahalanobis distance and variable importance evaluation method showed that the red bands, red edge bands, and spectral indices were important for plant species identification; (2) among the four classifiers, the ANN classifier had the highest overall classification accuracy on Dataset 3 of the reflectance images, which was 94.73%, and the Kappa coefficient was 0.93; (3) the machine learning algorithms RF and ANN performed well for identifying plant species, with an overall accuracy more than 91.59% and kappa coefficient above 0.89. These results suggest that close-range hyperspectral image and machine-learning classifiers, such as RF and ANN, can be used to effectively identify plant species in alpine steppe.