Abstract: Accurate tobacco planting distribution and area information is helpful for efficient allocation of tobacco resources,scientific optimization of planting management and improvement of economic benefits.While existing crop classification research has largely concentrated on major crops,the identification of tobacco cultivation areas through the fusion of multi-source remote sensing data remains under-explored.In this study,Xuanwei city,Yunnan province was taken as the research area.Based on field survey samples and existing crop datasets,tobacco/non-tobacco sample points were generated,and Sentinel-1 polarization features,Sentinel-2 spectral/index features and growth phenology features of tobacco were constructed.Then,the random forest algorithm was used for feature selection and classification to explore the optimal feature combination scheme for tobacco planting area recognition and achieve accurate identification of tobacco planting areas.The results indicate that the optimal scheme of 79 features,derived from the fusion and selection of Sentinel-1,Sentinel-2,and phenological data,achieves the highest tobacco identification accuracy.The overall accuracy (OA) reaches 94.51%,the Kappa coefficient is 0.89,and the coefficient of determination (R²) is 0.94 between the identified area and statistical data at the township scale.This study provides efficient and reliable technical support for large-scale and long-term tobacco planting monitoring,precise management and policy formulation.

Key words: tobacco detection, Sentinel-1, Sentinel-2, phenology, random forest