Classification of grassland conditions using a hyperspectral camera and deep learning

Grassland is the most widespread vegetation type in terrestrial ecosystem, but it has been threatened by degradation in recent years. Developing an operational species detection model is necessary for achieving grassland monitoring and making management plans. Therefore, this study aims to quickly and accurately classify grassland species based on hyperspectral imaging (HSI) technology and deep learning algorithms. In the present study, 16,200 hyperspectral data are collected from grassland samples over a period of 3 years using a hyperspectral imager, with a wavelength range of 400-1000 nm. Second, the hyperspectral data are preprocessed by the multiple scatter correction and mean normalization, improving the quality of input data and thereby enhancing modelling capabilities. Finally, four models are established, including temporal convolutional neural network (TempCNN), recurrent neural network with long short-term memory (LSTM-RNN), Transformer, and support vector machines (SVM). The results show that the preprocessed data have stronger modelling ability than the original data, and the classification performance of the model is ranked in descending order as Transformer, LSTM-RNN, TempCNN, and SVM. Among them, the classification performance of Medicago sativa L. in the TempCNN is superior to other combinations. The LSTM-RNN achieved accuracy of 1 for Agropyron cristatum var. pectinatum and Leymus chinensis (Trin.) Tzvel. The accuracy of both the Transformer and SVM for Leymus chinensis (Trin.) Tzvel. and Lotus corniculatus L. is 1. The results indicate the effectiveness and robustness of the proposed hyperspectral imaging technology combined with deep learning model, which has well classification performance for specific forage varieties.