Abstract: Crop identification is limited by sample quality and the interference from similar land types. To improve crop identification accuracy, this study explored an optimized crop identification method based on the phenological features of crops and a multi-sample dataset. Specifically, by analyzing the changes in phenological features of crops in time-series remote sensing data, this study determined the optimal phenophases for monitoring various crops, thereby improving the accuracy and stability of crop identification. Furthermore, a multi-sample dataset containing various sample sizes and types was constructed. Finally, a deep convolutional neural network (DCNN) model was developed to conduct comparative experiments under different combinations of phenophases and samples. The experimental results indicate that sample quality exerted a significant influence on crop identification accuracy. Under the same sample quality, selecting the optimal monitoring phenological period can significantly improve recognition performance. Under the combination of optimal sample quality and phenophases, the overall identification accuracy reached 84.88%, with a Kappa coefficient of 0.770 1. Overall, by constructing a multi-sample dataset based on the optimal phenophases for monitoring and optimizing the deep learning model, this study offers a novel approach for addressing the crop identification challenge posed by insufficient sample quality and similar growth cycles.