Accurate and timely crop mapping using remote sensing technology is crucial for precision agriculture, yield estimation, and food security. Deep learning models trained with proper loss functions are widely used in crop mapping and have achieved promising results. However, most of the existing loss functions focus on loss optimization in a universal way, i.e., problems regarding sample imbalance, and neglect the uniqueness of crop mapping task, for which its target often shows phenological characteristics. Given this, this letter proposes a crop phenological prior cross entropy loss (PPCE) function, which focuses on guiding the training processing in the direction where crops can be better identified. It is practical and easy to use. The phenological prior is quantified using normalized difference yellow index and normalized difference vegetation index obtained in different growing periods. Under the supervision of PPCE, if a crop pixel is misclassified to other class, the prior knowledge will increase the contribution of its loss to the final loss and thus guide the network to extract more discriminative features for crop mapping. To demonstrate the performance of PPCE, five widely used loss functions combined with three typical deep learning models (LSTM, DNN, and 1D-CNN) are compared. Experimental results show better performance of PPCE than the existing loss functions with different deep learning models.