In this study, Yuxi City of Yunnan Province, a typical tobacco-planting area in China, was selected as the study object, based on a dataset consisting of 156 pairs of soil-tobacco quality grades, soil key elements were identified through principal component analysis, and then three machine learning methods, namely the Back Propagation Neural Network (BPNN), Random Forest (RF) and Support Vector Machine (SVM) were employed to construct the prediction model of tobacco quality grade in order to achieve its spatial prediction and mapping. The results showed that based on 17 soil indicators, 11 specific indicators were identified as soil key elements, among these, clay content exhibited the highest contribution (accounting for 18.5%) to the variation in tobacco quality grades. The independent validation demonstrated that RF model achieved the highest accuracy (0.78) and Kappa coefficient (0.76) in the predictive performance, followed by SVM model, while BPNN model exhibited the least favorable outcomes. In terms of recall and precision, RF model demonstrated a descending level of accuracy in correctly categorizing tobacco quality grades, with the order of Level 5 > Level 1 > Level 2. Tobacco quality of Level 1 and 5 were predominantly distributed in the eastern part of Yuxi, with the easternmost Huanning County being the prime cultivation area for high-quality tobacco.