Hydrological and flood forecasting programs require accurate estimation of precipitation data. In particular, accurate, high-resolution and real-time quantitative precipitation estimation (QPE) plays an important role during severe convective rainfall events in localized regions (Cole & Moore, 2008;Jian et al., 2013). Rain gauges and radar are the two most widely used rainfall sensors, providing a direct and indirect measurements of rainfall intensity, respectively. In order to obtain the spatial distribution of precipitation from available gauges that are usually unevenly distributed, spatial interpolation techniques are required. The Kriging-based method is the most widely used rain gauge interpolation method (Jin & Heap, 2011), which is used to estimate the continuous attribute values of the true value at an unknown position. The methods based on the Kriging have the basic assumption of Gaussianness, and the fitting of the variogram model is complicated. To relax the Gauss hypothesis and make the interpolation method more flexible, some researchers have attempted it from different perspectives. Based on an artificial neural network (ANN), Sivapragasam et al. (2010) used data from 18 rain gauges to interpolate the average monthly precipitation in the Tasmilabarani Basin, India. The results suggested that ANN is a powerful alternative to the traditional Kriging spatial interpolation method. Verdin et al. ( 2016) compared two interpolation methods, the Ordinary Kriging (OK) and the K-Nearest Neighbor (K-NN) local polynomials, and fused them with precipitation data from the mountains of Central America and Colombia. Experimental results illustrate that these fusion methods improve precipitation estimation based on satellite data and increase the ability to capture extreme precipitation. To obtain the variogram of the OK method with better adaptability in daily precipitation interpolation, Q. Wang et al. ( 2016) compared and analyzed the four commonly used variograms