TOPMODEL is a storage-type semi-distributed-parameter model with advantages existing in each of the dostributed-parameter and lumped-parameter models. It can consider the physical characteristics nin the rainfall-runoff process in mountainous basins to some existent and has lately attracted considerable attention in the research field of rainfall-runoff analysis. In the TOPMODEL, water balances in groundwater storage and root zone are calculated by a sinmple lumped-parameter model. A basin area is horizontally divided into small square grids and phenomena in surfacesoil layers, such as surfacee runoff and groundwater recharge, are expressed by distributed-parameter models on each square grid. The time-varying runoff-contributing area during flood events can also be calculated in the distributed-parameter models by judging the generation of surface runoff from the soil moisture deficit in each square grid. In this paper, we made some modifications in the TOPMODEL and tried to improve the prediction accuracy of river discharge. First, in the modified-model 1, we introduced fast runoff component in the runoff-contributing area. Using this model, we tried to improve the prediction accuracy during flood events that can not be expressed only by groundwater storage. Next, in the modified-model2, water bakance in root zone was calculated by a distributed-parameter model to consider the special variation of thickness of surface soil layers on each square grod. Unknown parameters in each model were optimally searched by a simple generation algorithm with the fitness function defined by a 2 error between observed and caluclated river discharges. For the results, in the modified-model 1, fast runoff occured at the same time as peak discharge and it improved the prediction accuracy during flood events. In the modified-model 2, the predintion accuracy during low-discharge term was fairly improved.