Hyperspectral Classification of Grasslands for Sustainable Management Using Feature Fusion GRCNet

Grasslands play a crucial role in ecosystems, influencing key ecological functions such as biodiversity, climate regulation, and soil and water conservation. With the impacts of environmental changes and human activities, the functional status and health of grasslands are facing challenges. Therefore, efficient identification of grassland status is of great significance for the sustainable management, ecological protection, and restoration of grassland resources. To support the sustainable development of grasslands, the GRCNet network model is proposed. Grassland sweep photography was performed via a UAV-mounted hyperspectral imager to establish a 13-category grassland hyperspectral dataset. Then, Gaussian filter and principal component analysis (PCA) were used for noise reduction and dimensionality reduction in the hyperspectral images, and the GRCNet network model, which mainly consists of the GCA module, the RDC module, and the VIT-Base module, was established for the classification task. The experiments use the average accuracy, overall accuracy, F1 score, Kappa coefficient, and running time as the performance indexes, and the eight methods are compared with GRCNet. The results showed that the GRCNet network performed the best with AA reaching 92.84%, OA reaching 94.59%, F1 score reaching 97.22, and Kappa coefficient reaching 0.93. The GRCNet method is 10–20% more accurate than other methods. Comparative tests were also conducted using two public datasets, and GRCNet performed the best with a 2–20% improvement in accuracy. The results demonstrate the effectiveness of the GRCNet network model in hyperspectral grass classification tasks, which can be used as an efficient solution to the current problem of low hyperspectral accuracy and poor stability.