Abstract: Using a FieldSpec4 Hi-Res portable land cover spectrometer, four common shrub species in western Sichuan, including Salix cupularis、Potentilla fruticosa、Sibiraea angustata and Berberis diaphana Maxin were collected for field spectral data. The original spectral, first-order derivative spectral, and spectral de envelope lines data were analyzed to extract "green peak position", "green peak amplitude", "red valley position", "red valley amplitude", "red edge position", "red edge amplitude", and "red edge amplitude", which were used as inputs for neural network tree species recognition models to classify and recognize shrub species. The training accuracy of the neural network model was 93.0%, the validation accuracy was 83.3%, and the testing accuracy was 87.5%. The model had high accuracy and overall good classification results. The classification results showed that Salix cupularis and Potentilla fruticosa had the highest accuracy. The training accuracy of Berberis diaphana Maxin was high, while the testing accuracy was low, and the classification accuracy was unstable, mainly misclassified as Potentilla fruticosa; Sibiraea angustata was mainly misclassified as Potentilla fruticosa. Based on measured hyperspectral data, establishing a neural network tree species recognition model could effectively distinguish different shrub species, providing theoretical and technical support for hyperspectral remote sensing to carry out forest resource tree species classification and dynamic monitoring.