Abstract: Rice is one of the main food crops in China. It is of great importance for the timely, rapid, and accurate acquisition of rice field area and its spatial distribution in agricultural production, particularly for national food security. However, the traditional classification variables of remote sensing and pixel-based machine learning have posed great challenges in identifying the paddy fields with high fragmentation. The phenological parameters have been used to represent the growth dynamics of vegetation. Object-oriented random forests can also effectively avoid the 'salt and pepper' noise for the high accuracy of paddy fields. In this study, the object-oriented algorithm was selected to identify the paddy field, according to the rice phenological parameters. The research area was taken as the paddy field in Zhangzhou City, Fujian Province, a typical mountainous and hilly area in southern China. Firstly, the recursive feature elimination (RFE) and variance inflation factor (VIF) were applied to optimize the remote sensing variables. As a result, the normalized difference vegetation index (NDVI), modified normalized difference water index (MNDWI), soil-adjusted vegetation index (SAVI), vertical transmit/vertical receive (VV), vertical transmit/horizontal receive (VH), and six phenological parameters were taken to construct the classification model. Secondly, the object-oriented and pixel-based random forest was used to identify the paddy fields. Three combinations of the variables were used as the input parameters of the classification model. The effectiveness of phenological parameters and the object-oriented random forest were verified to improve the accuracy of paddy field identification in the complex terrain areas in southern China. The results are as follows: 1) The highest identification accuracy of paddy fields was 94.47%, where the Kappa coefficient was 0.92. A better performance was achieved to identify the paddy fields using the remote sensing variables, phenological parameters, and object-oriented random forest; 2) The phenological parameters shared the significant advantages to characterizing the vegetation growth among various types. Thus, the phenological parameters and remote sensing variables were combined to improve the accuracy of paddy field identification by 8.78-9.36 percentage points, compared with the experimental group with remote sensing variables only; 3) The object-oriented random forest was outstanding defined the shape and boundary of paddy fields with a high degree of fragmentation in the complex terrain areas. The artifact of 'salt and pepper' was effectively avoided. The object-oriented classification improved the accuracy of paddy fields by 0.58-1.53 percentage points, compared with the traditional pixel-based one. Therefore, the improved model was more suitable for the extraction of fragmented farmland in the complex topographic areas using remote sensing images. The finding can provide a strong reference to further improve the accuracy of paddy field mapping products in complex terrain areas.