Lotus root starch is highly nutritious, and its production process is complex. Some unscrupulous businessmen, driven by profit, adulterate lotus root starch with cheaper common starch or mix common starch into lotus root starch. Traditional methods for authenticating lotus root starch are time-consuming, labor-intensive, and destructive. Hyperspectral imaging technology, with its advantages of rapid, non-destructive, and accurate, has been widely applied in food safety detection. Therefore, this study proposes a method for quickly distinguishing between lotus root starch and other common starches, as well as identifying adulterated lotus root starch, by combining hyperspectral imaging technology with deep learning. Hyperspectral images of pure lotus root starch, four types of common starches, and adulterated starches were collected in the wavelength range of 900~1 700 nm using hyperspectral imaging technology. Several regions of interest (ROI) were delineated in the hyperspectral images of pure lotus root starch and the four common starches, and the average reflectance of each ROI was calculated from the original spectral data to build classification models. Abnormal bands affected by noise at the beginning and end of the original spectra were removed, leaving 443 bands between 940 and 1 675 nm. Outliers in the spectral data were then eliminated using the Isolation Forest (IF) algorithm. To enhance model training efficiency, the Competitive Adaptive Reweighted Sampling (CARS), Bootstrapping Soft Shrinkage (BOSS), and Channel Attention Mechanism Module (CAMM) were employed to extract 45, 32, and 12 feature wavelengths from the 443 bands, respectively. Partial Least Squares Discriminant Analysis (PLS-DA) classification models were constructed based on the spectral data after feature wavelength extraction, with the CAMM-PLS-DA model showing the best recognition effect, achieving an accuracy of 95.25% in the test set. To determine the optimal classification model, PLS-DA, Support Vector Machine (SVM), and Convolutional Neural Network (CNN) classification models were established using spectral data with different numbers of feature wavelengths extracted by CAMM. The CAMM-CNN model exhibited the best classification performance, with a highest accuracy of 99.69% in the test set. To further verify the ability of the CAMM-CNN model to distinguish adulterated lotus root starch, the spectral data of all pixel points in the hyperspectral images of adulterated lotus root starch were input into the trained CAMM-CNN model for discrimination. Visualization images showed that the model successfully identified various types of common starches in the adulterated lotus root starch. The results indicate that the combination of hyperspectral imaging technology and deep learning methods can effectively be applied to the authentication of lotus root starch, providing a new detection approach to combat the adulteration of lotus root starch and ensure its safety.