Flood segmentation using supervised deep learning models like U-Net plays a pivotal role in disaster response by facilitating rapid and accurate identification of flood-affected ar- eas. However, ensuring the reliability of these models’ predictions is essential, especially in high-stakes applications. Conformal Prediction (CP) provides statistically valid uncertainty estimates and is increasingly recognized as a robust tool for uncertainty quantification. This paper investigates the performance of the two major approaches in CP, namely Inductive CP and k-fold Cross- Validation CP (CV+), in the context of flood segmentation. By evaluating these methods on a baseline bitemporal U-Net model, we demonstrate that CP can offer critical insights into model confidence. Our findings highlight the limitations of Inductive CP in data-scarce scenarios and underscore the advantages of CV+ in achieving a superior balance between calibration and training data usage. This study highlights the importance of CP techniques in enhancing trustworthiness in flood segmentation models.