Rice mapping can provide invaluable technical assistance for food security assurance and field farming management. However, Rice identification still faces several challenges, including low automation in sample generation, complex feature selection, and limited model transferability. This paper proposes a Cross-Domain Rice Recognition Integrating Multi-Feature Optimization and Phenological Feature Matching (CDRR). The method initially utilises long-term remote sensing images as data sources, calculates the parameters of rice growth at six critical stages using a ratio model, and automatically generates samples by combining multiple surface cover rice thematic products. Then, a total of 80 feature indicators, distributed across five categories (including climatic statistics, spectral indices, spatial texture, and topographic parameters), are calculated for the time-series images in order to form the basic feature vector. Additionally, 12 parallel Random Forest classifiers are constructed on a monthly basis to perform feature optimization. Next, the separable index is calculated on a monthly basis, based on the predicted probabilities of the preferred features and the 12 classifiers. Its normalised value is then used as a sequence of phenological weights. Subsequently, the traditional Dynamic Time Warping approach is enhanced by incorporating the above phenological weights as a time weight. This enabled the alignment of phenomena features across diverse regions and facilitated the identificationof rice through a combined approach with the K-Nearest Neighbor algorithm. Finally, some experiments are conducted in four study areas with disparate cropping systems in Myanmar. The results demonstrate: 1) In the same-domain experiment, the mean overall accuracy of CDRR is 95.1%. In comparison with other conventional rice identification methods, this method yields improvements between 8.8% and 19.1%. 2) Cross-domain experiments reveal an improvement in CDRR's overall accuracy by 12.0%~32.4% compared to other methods. Notably, the accuracy of CDRR remaines consistent when compared with the same region, with a mere average 1% decline. The above results illustrate that the CDRR is effective in enhancing the accuracy of rice recognition in complex regions, meanwhile improves the migration ability of the recognition model.