Accurate and timely flood extent mapping is essential for emergency response. Optical satellite imagery is widely used for rapid flood mapping due to its global coverage and free availability. A common approach for delineating surface water from optical imagery involves the Normalized Difference Water Index (NDWI), which detects water features using green and near-infrared spectral bands. However, NDWI-based flood mapping requires the selection of a threshold value, and small variations in this threshold can lead to substantial differences in the estimated flood extent. At the same time, recent foundation segmentation models such as the Segment Anything Model (SAM) can identify object boundaries without task-specific training but requires manual prompting. This study proposes a risk-guided flood segmentation framework that integrates spectral thresholding with SAM refinement. First, NDWI thresholds are optimized using a risk score derived from flood hazard maps, allowing the threshold selection process to prioritize water detections in areas where flooding is more likely to occur. Then, the resulting NDWI-based flood mask is refined using SAM to improve boundary delineation and recover missed flood pixels. The method is evaluated using imagery from the 2018 spring flood along the Saint John (Wolastoq) River in New Brunswick, Canada, across five study regions using both Sentinel-2 and Landsat-8 scenes. Results show that the proposed risk-guided NDWI threshold selection with SAM refinement improves recall while maintaining stable precision. The framework requires no model training and provides a reproducible workflow for automated flood mapping from optical satellite imagery.