Scalable variational Gaussian process classification based on KL divergence(KLSP) is generaly plagued by the fact that both the mean and covariance matrix of inducing variables need to be optimized, therefore an improved algorithm is proposed by using the Laplace's method. Firstly, a lower bound of the posterior distribution of inducing variables is constructed, which can be computed easily. Then a Gaussian approximation of this lower bound is computed, which is by using the Laplace's method taken as the approximation the posterior distribution of inducing variables. As a result, the objective function of the proposed algorithm is a concave function with respect to the mean of inducing variables only, which can be solved easily. The experimental results show that, compared with the original algorithm, the speed and accuracy of the proposed algorithm are improved obviously.