Very-short-term quantitative precipitation forecasting (QPF; 0–3 h) is crucial for flood-risk management, but reliability is degraded by data (aleatoric) and model (epistemic) uncertainty. To isolate and quantify the effect of epistemic uncertainty on forecast performance, we generated nine synthetic rainfall fields (three motion types \(\times \) three intensity evolutions) by rotating an elliptical 2-D Gaussian and trained two convolutional neural network (CNN) QPF models (U-Net and Convolutional Long Short-Term Memory (ConvLSTM)). Monte Carlo Dropout (MCD) was applied to sample predictive distributions. Performance was evaluated using modified Kling–Gupta efficiency (KGE′), Root Mean Squared Error (RMSE), Critical Success Index (CSI), Continuous Ranked Probability Score (CRPS), 95% prediction-interval Coverage and Width, and Error Contribution by Intensity (ECI). Across all cases, deterministic skill ranged from KGE′ 0.83–0.99, CSI 0.90–0.99, and RMSE 0.20–3.18 mm·h−1. U-Net captured structure at short leads but increasingly showed boundary blurring and central overestimation/dispersion, whereas ConvLSTM preserved spatial coherence yet accumulated smoothing with intensity-dependent over/underestimation; ECI indicated a lead-time shift of relative error contribution toward moderate intensities. At t + 1, U-Net achieved CRPS 0.132–0.253 mm·h−1 with Coverage 0.953–0.984 (Width 1.70–3.17 mm·h−1). At t + 6, U-Net became strongly overconfident (Coverage 0.325–0.430; CRPS 0.917–2.069), whereas ConvLSTM maintained higher Coverage (0.623–0.847) and lower CRPS (0.498–0.981 mm·h−1). These findings quantify a sharpness–calibration trade-off in CNN-based QPF and motivate physics-informed constraints, joint aleatoric–epistemic uncertainty modeling, and post-hoc calibration for reliable nowcasting.
