In the process of hyperspectral image(HI) filtering by Savitzky-Golay(SG) filter, the spatial information of HI is ignored which lead to the accuracy of recognition model of wheat HI dataset can only reach 87.088 9%. This paper proposes a method, TSG filter, that combines space-spectrum information of HI. By this way, the accuracy of the model is improved by 12.066 7% compared with SG filter. The algorithm expands one-dimensional SG convolution core into two-dimensional SG convolution core in four directions. Then the HI data can be quickly convoluted using convolution theorem and fast Fourier transform, so that the space-spectrum noise of HI can quickly filter. When the TSG filter core coefficient m=2~4, n=3~5, the SNR of the HI is increased by more than 10%, the PSNR is higher than 30 dB, and the SSIM is greater than 96%, which means TSG filter maintains the original HI characteristics well and improve the SNR significantly. After TSG filtering (m=3, n=4) or SG filtering (m=7, n=3), by comparing with the gray image and spectrogram of Pavia University HI, it can be seen that after TSG filtering, the band noise of the image is suppressed, the peak height of the characteristic peaks is increased by 31.68%, and the average intensity of the characteristic band is increased by 41.83%, while after SG filtering the band noise of the image is still clear and the relative peak height of the characteristic peak is up to 13.40%. The model of wheat HI recognition model based on TSG-PCA-SVM algorithm is constructed. The training set contains 500 sample points and the test set contains 4 500 sample points. The accuracy of this model is as high as 99.155 6% and the kappa coefficient is 0.983 613 while predicting the test set. The total accuracy predicting all 116 880 samples in the wheat HI data set is as high as 99.206 0%, which means the classification model has high accuracy and good consistency, and the classification accuracy is significantly improved compared with SG filtering that only reaches 87.088 9%. In conclusion, this paper provides a new idea for HI filtering and provides a reference for the construction of a hyperspectral identification system of wheat Hi dataset.