Autonomous underwater vehicles must plan through observations affected by color cast, low illumination, scattering, and blur. This study develops a reproducible framework that converts the expert pixel annotations of the Segmentation of Underwater IMagery (SUIM) dataset into binary free-water/obstacle traversability labels, trains SegFormer-B0 with U-Net and DeepLabv3 baselines, estimates pixel-wise softmax entropy, and fuses obstacle probability with uncertainty into a grid risk field for A* replanning. Navigation under perceptual uncertainty is posed as a shortest-path problem on a 32 × 32 image-plane grid: the per-cell obstacle probability and entropy obtained from the network outputs enter the A* cost, so ambiguous cells are penalized without using labels at run time. All navigation experiments are image-plane grid simulations in which the vehicle cell is known exactly; real inertial navigation, localization error, and vehicle dynamics are not included. Four deterministic degradation levels are applied to held-out images, and 300 paired navigation scenarios with fixed start and goal cells are built from 60 test scenes, both for static planning and for a receding-horizon loop with limited sensing and route hysteresis. SegFormer-B0 reaches 0.913 obstacle-class IoU and 0.954 F1 on Clear and retains 0.843 obstacle-class IoU and 0.915 F1 at L4, with an expected calibration error below 0.04; retraining with five seeds confirms the backbone ranking from Clear to L3 but shows a large spread at L4 (0.733 ± 0.085 obstacle-class IoU). Mean normalized entropy increases from 0.373 to 0.414, while the top-10% hit rate in boundary or error regions increases from 29.4% to 34.9%. Risk-aware A* improves safety success over probability-cost A* on all conditions and reduces unsafe crossing by 2.86–10.23% with a mean path-length cost below 1.27%; the gain is small but consistent (8 of 8 discordant scenarios, McNemar p = 0.008) and shrinks when the baseline cost scale is tuned on validation data. The segmentation backbone has a larger effect (72.0% vs. 56.0% safe routes against DeepLabv3), and route hysteresis removes 96% of route switches in receding-horizon replanning.