ABSTRACT A new approach is proposed for the initialization of the primitive equations model (hereafter, NWP model), taking account of the effect of non-adiabatic heating due to release of the latent heat into humidity and divergent wind fields. In the present paper, an operational model of the Japan Meteorological Agency (JMA), the Japanese Spectral Model (JSM) is taken up as the NWP model. For the purpose of the present study, rainfall amount is estimated, with a resolution of about 5-km mesh, from radar-AMeDAS composite precipitation data (hereafter, R-A data), where AMeDAS stands for the Automated Meteorological Data Acquisition System which is a network of automated surface observation stations including raingauge measurements. The observed rainfall is considered to represent the mesoscale precipitation field. The total heating rate calculated from the R-A data is assumed to be partitioned parabolically in the vertical between the lifting condensation level (LCL) and the cloud top (TBB CT) estimated by the GMS (Japanese geostationary meteorological satellite), where TBB means the black body temperature. The diabatic initialization method proposed here consists of two steps. The first step is a physical initialization of the water vapour field, where the relative humidity field is changed to be more moist over the area of rainfall (>1 mm hr−1). The moistening is performed in such a way that, if the air is not moist enough over the area of rainfall, the lapse rate at each model layer from the LCL to the TBB CT becomes a critical value presumably required by the moist convective adjustment scheme used in the NWP model concerned. On the other hand, over the area where no rainfall is observed, the relative humidity is left untouched. The second step is a nonlinear normal mode initialization (NNMI) which includes the non-adiabatic heating rate obtained from the R-A data among other factors. The first four vertical modes with the period less than 6 hours are adopted. It should be remarked that the divergent component of wind field which is necessary to continuously maintain the condensation due to mesoscale rainfall is also initialized at the stage of NNMI with the large-scale condensation and the moist convective adjustment. Improvement in the operational performance of the NWP model with the use of the proposed diabatic initialization method is demonstrated for a case study and statistical verification. The forecasts with and without the present diabatic initialization are compared. Particularly, amelioration of the spin-up of mesoscale precipitation calculation for the first 6 hours after the start is clearly seen. The initialization method shows stable execution and seems to be useful for the prediction of the mesoscale phenomena, particularly in the tropics where meteoroligical data are relatively sparse.