Knowledge of soil properties is essential for risk assessment of vapor intrusion (VI). Data assimilation (DA) provides a valuable means to characterize contaminated sites by fusing the information contained in the measurement data (such as concentrations of volatile organic chemicals). Nevertheless, the application of DA in risk assessment of VI is quite limited. Moreover, soil heterogeneity is often overlooked in VI-related research. To fill these knowledge gaps, we apply a state-of-the-art DA method based on deep learning (DL), that is, ES(DL), to better characterize the contaminated sites in VI risk assessment. The effectiveness of ES(DL) is well demonstrated by three representative scenarios with increasing soil heterogeneity. The results clearly show that ignoring soil heterogeneity will significantly undermine one’s ability to make reasonable decisions in VI risk assessment. As a preliminary attempt of applying an advanced DA method in VI research, this work provides implications for the potential of using DL and DA in complex problems that couple hydrological and environmental processes.