Modeling and producing flood susceptibility maps represent a scientific approach to identifying prone areas and managing their consequences. Accordingly, the aim of this study is to compare the performance of three widely used Graph Neural Network (GNN) architectures, namely GCN, GAT, and GraphSAGE, in preparing the flood susceptibility map. Furthermore, the combination of these structures with the meta-heuristic Particle Swarm Optimization (PSO) algorithm was evaluated to identify the optimal configuration. In this study, 1,478 samples, including 739 flood occurrence points and 739 non-occurrence points, along with 13 effective variables in flood modeling—comprising topographic, hydrological, climatic, and environmental factors—were utilized. The performance of the models was evaluated and compared using various indices, including the Receiver Operating Characteristic (ROC) curve. The results indicated that all models performed very well in predicting floods, and their performance discrepancies were limited. The GraphSAGE-PSO model, achieving the highest AUC value of 96.99%, demonstrated the best performance; however, its difference from the base GraphSAGE model was marginal. Moreover, the base GraphSAGE model outperformed the hybrid GCN-PSO and GAT-PSO models. Finally, the results showed that employing the PSO algorithm can enhance the performance of certain GNN architectures, although the extent of this improvement depends on the architecture in question. The proposed methods can serve as a potential framework for similar studies; nevertheless, their application to new regions requires independent validation.