The Self-Organizing Map (SOM) can supportively organize complex datasets such as highly dimensional flood inundation maps. Nevertheless, SOM may produce distinct patterns after being trained with identical samples or may not converge in clustering highly dimensional datasets, which causes usability concerns and prevents its applications from a broader spectrum. Motivated by such concerns, two training strategies (S1 and S2) were proposed to configure SOM based on a large number of highly dimensional flood inundation maps associated with two basins located in southern Taiwan. S1 focused mainly on the weights’ adjustments in the ordering stage, while S2 would methodically balance the ordering and convergence activities on the weights’ adjustments. The effectiveness and suitability of S1 and S2 were inspected in detail by using coverage ratio, flip detector, and five clustering indices based on their configured topological maps in the two basins. The clustering results showed that the flip detector and the coverage ratio could visibly and objectively examine the suitability of the configured topological map. It was noticed that the influences of the ordering and convergence stages upon both training strategies for building SOM could significantly affect the coverage ratio as well as flip condition. Comparing the SOM topological maps implemented separately with each strategy, S2 strategy has a lower probability of causing a flipping situation and takes far fewer iterations to train a model of the same network size, which indicates S2 is more efficient and effective than S1 in configuring the SOM topological map for representing regional flood inundation maps.