Statistical methods to estimate wind resources at unsampled locations in a region can serve as an initial step to identify locations that warrant further investigation. There has been an ongoing effort to develop approaches for mapping the parameters of the wind speed distribution with statistical methods. This approach enables a comprehensive understanding of the wind resource variability across the entire region by considering the full wind speed distribution rather than focusing solely on mean values. The present study proposes a non-parametric approach to map the wind speed distribution. The method's main advantage is that it avoids constraining the region to a single distribution family and is thus more flexible than existing methods. In the proposed approach, a number of wind speed quantiles are first mapped in the region using machine learning techniques. Afterwards, the wind speed distribution is estimated by fitting an asymmetric kernel estimator to the estimated wind speed quantiles at unsampled locations. The new approach was compared to the standard statistical method based on mapping the regional wind speed distribution parameters. The results indicate that the non-parametric approach leads in the best scenario to a 9% and 6% drop in the Kolmogorov-Smirnov statistic on average during cross-validation and validation, respectively. The Birnbaum-Saunders and the Log-Normal kernels gave a better fit to the estimated wind speed quantiles than the Weibull kernel. The proposed approach is recommended in regions with high wind regime variability.