A segmented Transformer feature extraction network is proposed to address the challenges posed by near infrared spectroscopy, including high dimensionality, susceptibility to noise interference, and high spectral similarity among samples from different provinces. This network is applied to tobacco leaf origin identification to enhance classification accuracy. First, based on the one-dimensional structure of near infrared spectroscopy, an embedding layer is designed to compress features using one-dimensional convolution. Second, the data is divided into three parts along the spectral dimension using sliding windows. The Transformer architecture is improved to extract spectral features, addressing the issue of computational inefficiency caused by a large number of spectral bands. Last, to adapt to the characteristics of spectral data, a regression head with multiple layers of one-dimensional convolution is designed to predict the origin of the samples. To validate the effectiveness of the proposed algorithm, several comparative experiments are conducted, comparing classification accuracy, precision, recall, and other metrics with other algorithms. The superiority of each structure in the model is verified. The experimental results demonstrate that the proposed model effectively utilizes the spectral structure for feature extraction and noise suppression, successfully accomplishing the task of tobacco leaf origin identification.