Abstract: Accurate extraction of land use types is an important way to manage ecological environment of watershed, but there are many classification algorithms with different effects on classified objects, and the Erhai Basin, as a plateau fracture lake, has its unique characteristics, and it is of great significance to carry out research on applicability of classification methods．Based on the Sentinel-2 image data, 34 optimal classification features were selected by innovatively integrating spectral features, geometric structure and texture features, and feature space optimization algorithm was applied to compare and analyze applicability of three classification methods of Support Vector Machine（SVM）, random forest（RF）, and decision tree（DT）in the Erhai Basin as study area．Results showed that, in ranking of feature importance, red edge index（NDREI）showed a more significant advantage over tnormalized vegetation index（NDVI）and ratio vegetation index（RVI）, and ranked first in contribution rate among all features．After optimizing feature space, SVM has the best effect on overall accuracy and Kappa coefficient of classification, with an improvement of 3.75 and 5.06 percentage points, and DT has the worst effect, with an improvement of only 1 and 1.17 percentage points．Compared with the other two classification methods, RFF with optimized feature space has the highest overall classification accuracy and Kappa coefficient, reaching 90.63% and 88.87%, and was more in line with distribution of real features in details, such as distinguishing wetlands, water bodies and construction land in fine-grained area of feature distribution．The use of feature space optimization algorithm improved accuracy of classification method, and the RFFs has the strongest applicability to land use classification in Erhai Basin．