Automated seismic facies segmentation can improve the consistency and efficiency of subsurface interpretation, but complex stratigraphic boundaries and imbalanced facies remain challenging. This work evaluates a reproducible pipeline for multiclass seismic facies segmentation based on simple probability-level ensembles of three U-Net-derived architectures that share an EfficientNet-B4 encoder: UNet++, CMUNeXt, and MultiResUNet. The models were trained and tested on the real field-acquired Netherlands F3 seismic benchmark, which contains six interpreted facies. During inference, each network produces class-wise softmax probability maps, and the ensemble averages these maps before pixel-wise classification. The individual models achieved mean intersection over union (mIoU) values of 0.710, 0.744, and 0.757, respectively. The CMUNeXt–MultiResUNet ensemble reached an mIoU of 0.762, while the three-model ensemble achieved the best overall balance among the proposed configurations, with mIoU = 0.766, frequency-weighted intersection over union (FWIoU) = 0.883, pixel accuracy (PA) = 0.934, and mean class accuracy (MCA) = 0.860. In a single-run ablation, disabling all training-time augmentation while retaining the same fixed normalization reduced the three-model ensemble mIoU from 0.766 to 0.719; augmentation improved PA, MCA, FWIoU, and mIoU for every evaluated model and ensemble. Relative to published F3 results, the proposed ensemble improves several global metrics over multiple baselines but does not surpass every prior method. These findings show that low-complexity ensemble averaging can exploit complementary class-wise behavior without introducing a separate meta-learner, providing a competitive and computationally practical baseline for automated seismic interpretation.