Abstract: In order to develop a more reliable evaluation method for tobacco-growing soil, soil samples were collected from a raw material site of China Tobacco Jiangsu Industrial Limited Corporation and the comprehensive state of the soil sample fertility was evaluated by analytic hierarchy process (AHP) and attribute reduction method in rough set. The evaluation results were verified by the tobacco leaf yield of the corresponding year. The results showed that: 1) The differences of soil fertility indices from 15 sampling sites were not the same. There were significant differences in 7 indices, such as the contents (mass fraction) of organic matter, alkali-hydrolyzed nitrogen, water-soluble chlorine, readily-available phosphorus, readily-available potassium, available sulfur and available magnesium. There were smaller differences in soil pH and contents of total potassium, available boron, available molybdenum and available calcium, while there were no significant differences in total nitrogen and total phosphorus. 2) All the indices screened by multiple comparisons and their initial weight coefficients determined by the AHP were contents of organic matter(0.229), alkali hydrolyzed nitrogen(0.357), readily-available phosphorus(0.120), readily-available potassium (0.189), available magnesium (0.055) and water-soluble chlorine (0.050). 3) The minimum approximate reductions determined by the attribute reduction method in rough set were contents of organic matter, readily-available phosphorus, readily-available potassium and water-soluble chlorine, and their weight coefficients were 0.214, 0.214, 0.321 and 0.251 respectively. 4) The evaluation results of comprehensive state of soil fertility obtained through the AHP and the attribute reduction method in rough set were different. The comprehensive soil fertility results obtained by different methods were significantly linearly correlated with tobacco leaf yields, the correlation coefficients (r) were 0.65 and 0.92, determination coefficients (R²) were 0.42 and 0.86, and root mean square error(RMSE) were 5.92 and 5.89, respectively. The attribute reduction method in rough set had higher correlation coefficients and accuracy.