Floods are one of the most dangerous natural hazards and the greatest challenge for hydrologists due to their mass force and short response time. Taiwan is located in the northwestern Pacific Ocean where the activities of the subtropical jet stream are frequent. In the last century, there were about 360 typhoons, an average of 3.6 annually that hit the Taiwan Island. Typhoons are usually coupled with huge amounts of rain from June to October, and disastrous flooding results from the intense bursts of rainfall. The rivers in this island are short and steep, and their flows are relatively quick with floods lasting only for a few hours and usually less than one day. The large flood peaks with fast-rising limbs would unavoidably cause serious disasters. Last year Typhoon Morakot struck south Taiwan with stunning rainfall on August 8th with the highest precipitation reaching 1166 mm/day. It caused 665 deaths, 34 missing, many civilian injuries, and even a small village was buried under the following debris flow. Estimation of flood depths and extents may provide the disaster information for dealing with contingency and alleviating risk and loss of life and property. We proposed serial-propagated back-propagation neural networks (BPNNs) to forecast one to six-hour-ahead flood inundation depths. The practicability and effectiveness of the proposed approach is tested on several inundation-prone spots of three counties in Taiwan. The results show that the proposed serial-propagated BPNNs can adequately provide one to six-hour-ahead flood inundation depths that well match the simulation flood inundation results.