Nonprobability samples have been used frequently in practice due to the lack of sampling frame information, time or budget. Inference by only using nonprobability samples without further adjustments may lead to biased results. Parametric mass imputation approaches have been developed in previous research, but the performances depend on the underlying parametric model assumptions. To overcome this issue, we propose nonparametric mass imputation for data integration. For low dimensional covariate, kernel smoothing approach is proposed. For relatively high dimensional covariate, generalized additive model is used for imputation. Asymptotic theories have been developed. Simulation studies as well as real application show the benefits of our proposed methods compared with parametric methods.