Timely and accurate mapping of rice cultivation is crucial for ensuring global food security and monitoring water usage. Feature selection methods play critical roles in identifying and mapping paddy rice as they reduce redundant information in feature subsets and improve computational efficiency. However, the optimal feature sets selected by existing feature selection methods still encounter challenges such as redundant information or local optimal, limiting their accuracy in rice identification. To address these issues, we developed a novel hierarchical clustering sequential forward selection (HCSFS) method to accurately determine the optimal feature set for paddy rice identification. HCSFS first employs hierarchical clustering to classify all features into different classes. Each independent feature class is filtered by the existing advanced sequential forward selection (SFS) method. Then, all the filtered features are merged to select the optimal feature set for rice identification. The proposed HCSFS method was tested on 8 common machine learning classifiers. The results show that, compared with existing feature selection methods, the feature subset obtained by HCSFS reduced redundant information and demonstrated superior performance. Specifically, the optimal feature set selected by HCSFS yielded the highest accurate rice map, with overall accuracy exceeding 0.95 and Kappa exceeding 0.83 across all classifiers. In addition, this paper found that in regions of southern China with cloudy and rainy weather and complex crop planting structures, the combination of the rice growth period images with LSWI, SWIR2, and RE2 can improve the accuracy of paddy rice identification or mapping. The case validated the applicability and efficiency of the HCSFS method in rice identification for regions with cloudy and rainy and implied the potential use in other similar or less complex regions.