We introduce GST-UNet, a neural framework for valid causal inference from spatiotemporal observational data with time-varying confounding and complex spatiotemporal dependencies.