The weighted stochastic response surface method (WSRSM) has been demonstrated to be effective in improving the accuracy of the estimation of statistical moments and probability of failure (PoF) upon the stochastic response surface method (SRSM). However, it has been noticed that the weighting method in WSRSM may have little and sometimes negative impact on PoF estimation especially in the cases of low PoF. To mitigate this issue, a new double weighted SRSM (DWSRSM) is proposed that the weights of sample points are determined based on their importance not only to regression but also to PoF estimation. Specifically, relatively larger weights are assigned to points closer to the failure surface, which significantly accounts for the accuracy of PoF estimation. Comparative studies show that DWSRSM outperforms WSRSM producing more accurate PoF estimation without incurring additional function evaluations. An application of DWSRSM to the rocket design further demonstrates its effectiveness for PoF estimation.