Abstract: Tobacco lands in the southern mountainous areas of China are mostly characterized with small patches and fragmentedly mixed with other croplands. The local cloudy and rainy weather also hinders capturing land cover by optical remote sensing and thereafter affects remote sensing investigation. To solve the above issues of tobacco land recognization, we proposed a tobacco growing area extraction approach by utilizing the nature of multiple scale object-based classification method and taking full advantage of the Gaofen multi-source and multi-temporal remote sensing imagery. The tobacco land recognization experiment was conducted in Dahe town, Qixingguan District, Bijie, and the object-based classification was compared with the pixel-based maximum likelihood (ML), neural network (NN), and support vector machine (SVM). The results showed that the accuracy of the object-based method was the best, followed by ML, SVM, and NN, with the Kappa coefficients of 0.948, 0.936, 0.930, and 0.905 respectively. In addition, the shape of tobacco patches extracted by the object-based method were relatively complete, which effectively avoided the "salt and pepper phenomenon", and the visual effect was significantly better than that of the pixel-based method. The object-based classification method combined with Gaofen remote sensing constellation can accurately extract the distribution of tobacco planting areas in southern China, which is helpful to the macro-management, regulation, and decision-making.