This paper compares and contrasts three satellite infrared methods of estimating rainfall. The World Climate Research Programme's Global Precipitation Climatology Project held its first algorithm intercomparison during JuneAugust 1989 over the Japanese archipelago and surroundings. The GOES precipitation index (GPI), the NegriAdlerWetzel Technique, and the convective-stratiform technique (CST) were applied to hourly infrared imagery. A network of radar data calibrated by a dense raingage network was used as ground truth but withheld from the investigators until after submission of the satellite estimates. Scattering signatures in concurrent 86-GHz brightness temperatures from the Special Sensor Microwave/Imager were used to develop a method to discriminate nonraining cirrus from active convection in two of the infrared techniques.
All three of the IR techniques did poorly in estimating the rain maxima over southeastern Japan associated with shallow orographic (warm) rain systems. The statistics for the combined 2-month dataset for both land and ocean (1.25° grid boxes) indicated that the GPI had the lowest bias (30 mm or 25% of the radar mean) but also a low correlation (.48) and high root-mean-square error (rmse) (103 mm or 87% of the mean). This was due to the GPI's overestimate in June (bias was 92 mm) and underestimate in JulyAugust (bias was −32 mm). Despite its increased sophistication, the CST had an rmse of 104 mm, with a large negative bias (−70 mm) but a higher correlation coefficient (0.66). When the dataset was limited to the ocean-only points (to remove the effect of the shallow orographic precipitation), new statistics emerged. Under these restrictions, and for this limited dataset, the CST performed best, with the lowest bias (−39 mm or 42% of the mean), the lowest rmse (65 mm or 71% of the mean), and the highest correlation (0.79). It is believed that the lower scatter (higher correlation) of the CST and NAWT with respect to the GPI is due to the discrimination of thin cirrus used in both the NAWT and CST.
Daily rainfall estimates had rms errors of almost 200% of the mean and negative biases of about 50% of the mean. Hourly estimates for 1.25° grid boxes had rms errors of 200%300% of the mean and negative biases of order 100% of the mean. Spatial averaging to 2.5° showed a slight improvement in these statistics. Despite the poor performance on hourly scales, the satellite techniques were able to identify diurnal signals when averaged over 1 month.