Floods are renowned as the most destructive natural phenomena, and their frequency and intensity increase due to climate change. Accurate and timely flood mapping is critical for effective risk mitigation. However, traditional approaches relying on optical remote sensing imagery and synthetic aperture radar (SAR) classification face significant limitations due to cloud cover and misclassification-induced low accuracy. To address these challenges, this study developed a novel hybrid framework of metaheuristic optimization (MO) and deep learning (DL)-based semantic segmentation for more precise flood mapping. Three MO algorithms including artificial bee colony (ABC), genetic algorithm (GA), and swarm-based simulated annealing (SwarmSA) were used to identify the most informative combination set of polarimetric SAR (PolSAR), including VV and VH, PolSAR decomposed features, and textural descriptors from the Gray-Level Co-occurrence Matrix (GLCM). Three convolutional neural network (CNN)-based DL models (e.g., VGG16-U-Net, DRN, and CPNet) were trained to extract flood inundated areas from the Sentinel-1 SAR imagery. The proposed methodology was applied to the April 2019 flood event in Khuzestan province, Iran. The results showed that the CPNet model coupled with the SwarmSA achieved the highest F1-score (flooded areas: 0.901; non-flooded area: 0.976) and IoU (flooded areas: 0.820; non-flooded area: 0.954) in mapping inundated areas. Furthermore, the selected feature set, which includes dissimilarity from VV, GLCM correlation from VV, homogeneity from VH, GLCM mean from VV, and GLCM correlation from VH effectively captured the spectral and textural characteristics of flooded areas. The results highlighted the effectiveness of integrating MO-based feature selection techniques with DL architectures to achieve high-resolution and expeditious flood extent mapping.