Typhoon Damage Scale Forecasting with Self-Organizing Maps Trained by Selective Presentation Learning

We previously proposed a new typhoon warning system which forecasts the likely extent of damage associated with a typhoon towards humans and buildings. The relation between typhoon data and damage data was learned by self-organizing maps (SOM) and typhoon damage scale (small, middle or large) was forecast by the SOM using typhoon data. Although average accuracy for actually small scale damage data was comparatively high (96.2%), average accuracy for actually large scale damage data was comparatively low (65.2%). Thus, we apply a selective presentation learning technique for improving the predictability of large scale damage by SOM. Learning data corresponding to middle and large scale damage are presented more often. Average accuracy for actually large scale damage data was increased by about 9%. The accuracy for actually large scale of numbers of fatalities and houses under water was increased by 25% and 20%, respectively.