The ability to produce truthful maps of the distribution of one or more gases is beneficial for applications ranging from environmental monitoring to mines and industrial plants surveillance. Realistic environments are often too complicated for applying analytical gas plume models or performing reliable CFD simulations, making data-driven statistical gas distribution models the most attractive alternative. However, statistical models for gas distribution modelling, often rely on a set of meta-parameters that need to be learned from the data through Cross Validation (CV) techniques. CV techniques are computationally expensive and therefore need to be computed offline. As a faster alternative, we propose a parameter selection method based on Virtual Leave-One-Out Cross Validation (VLOOCV) that enables online learning of meta-parameters. In particular, we consider the Kernel DM+V, one of the most well studied algorithms for statistical gas distribution mapping, which relies on a meta-parameter, the kernel bandwidth. We validate the proposed VLOOCV method on a set of indoor and outdoor experiments where a mobile robot with a Photo Ionization Detector (PID) was collecting gas measurements. The approximation provided by the proposed VLOOCV method achieves very similar results to plain Cross Validation at a fraction of the computational cost. This is an important step in the development of on-line statistical gas distribution modelling algorithms.