An input subset including average radar reflectivity () and its standard deviation (SD) is proposed to improve radar estimates of rainfall based on a radial basis function (RBF) neural network. The RBF derives a relationship from a historical input subset, called a training dataset, consisting of radar measurements such as reflectivity () aloft and associated rainfall observation () on the ground. The unknown rainfall rate can then be predicted over the derived relationship with known radar measurements. The selection of the input subset has a significant impact on the prediction performance. This study simplified the selection of input subsets and studied its improvement in rainfall estimation. The proposed subset includes: (1) the of the observed within a given distance from the ground observation to represent the intensity of a storm system and (2) the SD of the observed to describe the spatial variability. Using three historical rainfall events in 1999 near Darwin, Australia, the performance evaluation is conducted using three approaches: an empirical relation, RBF with , and RBF with and SD. The results showed that the RBF with both and SD achieved better rainfall estimations than the RBF using only . Two performance measures were used: (1) the Pearson correlation coefficient improved from 0.15 to 0.58 and (2) the average root-mean-square error decreased from 14.14 mm to 11.43 mm. The proposed model and findings can be used for further applications involving the use of neural networks for radar estimates of rainfall.