Smoker's melanosis (SM) is a way to determine if someone is a smoker or not from the tongue area. Smoker's melanosis changes the color of melanin pigmentation on the oral mucosa into brown or black. The method is usually performed by TCM experts or doctors with TCM methods. However, its detection is noninvasive and takes much time due to the long procedure. This research aims to make Smoker's tongue detection systems using noninvasive and nondestructive methods by implementing Deep Learning algorithms. The study used a hyperspectral camera with a VNIR wavelength to record a person's tongue image and process it into information used for this system. The algorithm implementation was performed on five different datasets based on the ROI retrieval area on the tongue of smokers and nonsmokers. This research focuses on Deep Learning algorithms implementation. Convolutional Neural Network (CNN) algorithm used has two types of architecture Autoencoder and Proposed Architecture. Both architectures are run by optimization algorithms such as SGDM, Adam, and RMSProp. In addition to comparing the two CNN architectures, the research examines the PCA-SVM to see the performance of Deep Learning and Machine Learning to be implemented in this research data. Proposed Architecture achieved an accuracy of 95% in SGDM and PCA-SVM optimization algorithms used to make the accuracy of 81%. These results suggest that smoking tongue detection systems can work better with Deep Learning scanning.