Abstract: Atractylodes lancea（Thunb.）DC. and Atractylodes japonica Koidz. Ex Kitam. are highly similar in appearance, composition, and other aspects．Traditional identification methods based on morphological or chemical indicators have low classification accuracy under conditions of small sample sizes or non-destructive testing．A deep learning classification network called SS-FusionNet was proposed, which integrated spectral and image information for high-precision classification of Atractylodes lancea（Thunb.）DC. and Atractylodes japonica Koidz. Ex Kitam. slices under hyperspectral image and small sample conditions．Atractylodes lancea（Thunb.）DC. and Atractylodes japonica Koidz. Ex Kitam. slices sample data were collected by hyperspectral imaging system．An autoencoder network was pre-trained using unlabeled hyperspectral data to enable encoder module to extract image features from spectral data．Spectral features were deeply fused with image features, and classification was performed by combining upsampling convolution module．Experimental results showed that under small sample conditions, SS-FusionNet achieved a classification accuracy of 92.7%, which was 7.5 percentage points higher than 85.2% classification accuracy of support vector machines and 6.1 percentage points higher than 86.6% accuracy of convolutional neural networks．A new ideas and methods was procided for in-depth identification research on traditional Chinese medicine species．