In order to improve the precipitation forecasting skill of CMA-MESO (China Meteorological Administration Mesoscale Model) at 3 km resolution, three different horizontal correlation characteristic scales of background error covariance are obtained with 2-D discrete cosine transform filter from three months’ (2 June to 31 August 2018) background error samples. The three horizontal co-correlation scales are fitted and implemented in CMA-MESO operational testing system with recursive filter of three different scales, so as to replace the single-scaled recursive filtering. The results show that the profiles of three horizontal co-correlation scales with height are similar, with tens to hundreds km apart. The analysis qualities and verification of precipitation forecasting between the control experiments (single-scaled recursive filterings) and sensitivity experiments (three different scaled recursive filterings) with CMA-MESO system at 3 km resolution are compared. The numerical results indicate that the wind and relative humidity analyses are more close to observation in sensitivity experiments. The increment difference of temperature analysis is very small. In addition, the precipitation forecast skill is improved in sensitivity experiments. The first 6 h precipitation forecast TS value over 1-31 July 2018 with cold start is higher and the bias value is more close to 1 in sensitivity experiments. Meanwhile, the TS value of every 6 h precipitation in 24 h forecasting term with warm start is improved too.