In order to evacuate residents in time, flood warning systems must have rapid data processing algorithms to translate detailed numerical data into simple warnings and navigation aids. Although it is common to display warning messages by directly drawing red circles on a map embedded in a Web page to indicate the regions about to be inundated, such an approach has some drawbacks. We propose an alternative way by summarizing the warning messages based on landmarks, so that the messages can be short and convey even more information. We have designed two approaches to output such messages. They are the nearest landmark approach and the threshold approach, which differ in the way of determining which landmarks will be influenced by a certain flooded region. These two methods are implemented and we compare their performance through real and synthetic datasets. Experimental results show that the threshold approach usually takes less execution time than the nearest landmark approach. Its severity ranking of landmarks is also better suited to human behaviour.