With the increased damage caused by wildfires under global warming, accurately predicting wildfires remains challenging yet crucial for regional disaster mitigation. The boreal summer (June–September: JJAS) wildfire burned area (BA) of Central Asia reaches the greatest seasonal extent and has the largest interannual variance over the Eurasian continent. Considering both physical relationships and prediction skills for the BA, two independent physical meteorological variables are found to dominate the interannual variance in Central Asian summer BA during 1996–2016. One is the decreased March snow water equivalent (SWE), consistent with the regional warming and increased soil moisture in March, which favors subsequent plant growth. The other is the increased JJAS 500 hPa geopotential height (GHT500) over Central Asia, corresponding to the summer dryness anomaly, which enhances the local fuel flammability. Accordingly, a physically-based statistical-dynamical seasonal model, starting from April 1st each year, is established for predicting Central Asian summer BA using the observed March SWE and dynamically-predicted JJAS GHT500. This model has the highest operational prediction skill (of 0.58) among numerous tests and is shown to be stable by using K-Fold cross-validation. Although the pre-winter ENSO is correlated with the Central Asian summer circulation anomaly, the prediction skills using ENSO-based models (both statistical and statistical-dynamical) are either low or unstable. This study suggests that regional wildfires can be accurately predicted by choosing relevant meteorological variables and optimal dynamically-predicted fields. This approach has great potential to be useful for seasonal wildfire prediction and future projections under climate change.