Based on multispectral digital image texture feature, a new rapid and nondestructive method for discriminating tea categories was put forward. The new method combines the advantages of DCT (discrete cosine transform) and least squares-support vector machine (LS-SVM). In the present study, the images for each sample were captured using a red (R) waveband, near infrared (NIR) waveband and green (G) waveband multispectral digital imager. The three wavebands of image can be combined into one image, which contains more information than images captured by ordinary digital cameras, and the NIR image can catch more information than visible spectrum. Three images for one sample can be obtained simultaneously. Eighty filters were designed based on DCT. One hundred twenty images (twenty for each category) were used for calibration set and one hundred twenty mages (twenty for each category) were used as the prediction. Finally, tea category was classified by LS-SVM. The classification rate using Sd of NIR image was only 73.33%, while it reached 100% using 8 filtered images. The overall results show that the technique combining DCT and SVM can be efficiently utilized for texture recognition of multispectral image, and it also is an effective and simple discrimination way for the tea categories. The whole process is simple and easy to operate, and can be transferred to the industrial world for on-line application.