Your U-Net model is overfitting likely due to a limited dataset or insufficient regularization. In addition to data augmentation, try these steps:

1. Use dropout layers in the encoder and decoder.

2. Apply L2 regularization on convolution layers.

3. Reduce the model complexity (e.g., fewer filters per layer).

4. Implement early stopping based on validation loss.

5. Consider using a pre-trained encoder (e.g., ResNet as backbone).
   Also, ensure your validation set is truly representative and not too small.