The radioactive materials are generally concentrated downwind of their origins when the prevailing winds blow continuously in one direction. If this principle determined the pattern of dispersion in all cases, dispersion directions could be estimated by wind patterns. However, this hypothesis has not been sufficiently verified because of the complexity of dispersion processes and weather systems. Here, we show that dispersion directions, which are divided into four ranges, can be estimated by wind patterns using a machine learning approach. The five-year average hit rates of the directions of dispersion estimated using near-surface winds exceed 0.85 in all months. The dispersion directions can be estimated up to 33 hours in advance using forecast winds. In particular, high hit rates exceeding 0.95 are achieved in January and March, when large-scale weather systems dominate. These results indicate that the dispersion directions are determined by the wind patterns that correspond to large-scale weather systems and diurnal circulation patterns in most cases. Our findings also provide more reliable information on dispersion patterns with reduced uncertainties, given that reasonable skill is achieved at a sufficient lead time for evacuation.