The multi-temporal Sentinel-2 images were used to classify the autumn crops in Gaocheng, Shijiazhuang to provide an important basis for the local agricultural planting structure adjustment. The influence of comprehensive multi-features and feature optimization on the extraction accuracy of autumn crop planting area were analyzed. In order to reduce the influence of high-dimensional features on the performance of the classifier, a filter hybrid feature selection model (ReEMISE) based on improved separability and thresholds combined with the Relief F algorithm was proposed. Firstly, the Relief F dimensionality reduction algorithm was used to select the features. Secondly, the improved separability and thresholds (EMISE) combined with image entropy was used to further optimize the preliminary feature set, and then the EMISE feature importance value was given a Relief F feature weight. Finally, the random forest model was used to extract the crop planting area from the optimized feature subset, and compared with the random forest classification accuracy of the Relief F dimensionality reduction algorithm and the EMISE dimensionality reduction algorithm. The purpose was to ensure the accuracy of classification, minimize the feature dimensions and improve the classification efficiency. Six different types of feature variables were generated based on the Sentinel-2 data with multi-phase and rich spectral information, including multi-temporal spectral features, traditional index features, red-edge index features, texture features, difference features of different time-phase bands, and ratio features of different time-phase bands. On the basis of multi-temporal spectral features, adding different features, totally six groups of different feature combination experiments were constructed to extract autumn crop planting area and verify the classification accuracy of different feature combination. At the same time, the influence of different features on the extraction accuracy of crop planting area was analyzed from two aspects: the importance of features and the classification accuracy of different feature combinations. The results showed that the best correlation coefficient threshold of the ReEMISE’s feature-optimized random forest model（RF_ReEMISE) was 0.96. The accuracy was the highest when the number of feature variables was 9 and the overall accuracy and Kappa coefficient were 95.3918% and 0.9397, respectively. The advantages of the RF_ReEMISE mainly included two aspects, i.e., the least number of feature variables at the highest accuracy, and the classification accuracy index of ReEMISE algorithm was the highest when the number of different feature variables of the three algorithms attained the best accuracy. Comprehensive multi-feature was the key to improve the accuracy of crop classification. Based on the multi-temporal spectral features, the traditional index features and the red-edge index features were added respectively, and the overall accuracy was improved by 1.5021 percentage points and 1.5715 percentage points, and the Kappa coefficient was increased by 0.0198 and 0.0207. The accuracy was slightly reduced when adding the texture features, the difference features of different time-phases bands, and the ratio features of different time-phase bands. The RF_ReEMISE can reduce the feature dimensions, improve the classification accuracy, and achieve the balance between efficiency and accuracy. The spectral differences of crops at different phenological stages can be better reflected by introducing the ratio features of different time-phase bands than the difference features of different time-phase bands. It was found that correlation between the red-edge vegetation index and chlorophyll was stronger, and the importance of features was higher. The RF_ReEMISE with multi-features can effectively improve the accuracy and efficiency of autumn crop classification.