Deep learning (DL) models for wireless propagation face three critical challenges: robustness to domain shifts, uncertainty quantification, and interpretability. This paper addresses these three critical issues, advancing towards more generalizable, reliable, and explainable DL models for wireless communication. We build upon a hybrid architecture combining scalar metadata with a diffraction parameter profile, using residual and attention blocks to capture long-range terrain-dependent effects in the very high frequency (VHF) band. We demonstrate that domain adatation using deep correlation alignment (Deep CORAL) substantially improves generalization under domain shifts. To ensure trustworthy predictions, we employ conformalized quantile regression (CQR), which provides finite-sample, distribution-free uncertainty guarantees and corrects for overconfident predictions, yielding well-calibrated 90% prediction intervals that enhance reliability for emergency service scenarios. Furthermore, using explainability techniques including saliency maps, attention visualizations, and Grad-CAM, we show that the model captures physically meaningful features. This integrated approach of domain-adaptive, uncertainty-aware, and interpretable modeling paves the way for trustworthy DL deployment in real-world wireless communication systems.