Data driven methods for impact location and severity estimation have great potential for real life application due to their ability to construct meta models that are not affected by changes in structural complexity (e.g. stiffeners and cut-outs). However, to do so, an initial reference database containing known input and output pairs is required to form an accurate meta model. The requirement to collect this data (mainly in the form of experimental tests) is considered by many to be not feasible and hinders the application of data driven methods in large scale, real life structures. Here a new multifidelity approach is presented to reduce the “cost” of constructing the reference database necessary for data driven impact location and severity estimation methods. The proposed cokriging approach is used to combined “cheap” low fidelity FE (Finite Element) simulation data with a limited amount of accurate but “expensive” experimental data to construct a reference database that requires less experimental data sampling (thus lower “cost”) but with similar levels of accuracy compared to reference databases that were constructed from pure experimental data. Using previously developed impact location (kriging based localisation) and maximum force estimation (maximum impact force gradient method) methods, the effect of different reference databases with varying amounts of experimental data samples are tested. The results obtained on a CFRP (Carbon Fibre Reinforced Plastic) coupon and stiffened panel showed that using pure FE data for the reference database yielded poor results. By adding a limited number of experimental points, it was shown that the accuracy increases significantly approaching levels achieved using pure experimental data with significantly less samples. The wider context of data fusion is also explored, highlighting the possibility to combine various sources of data (including inspection history as well as data from sister assets) which can possibly be used to create a data driven framework that requires less initial data whilst also being robust and self-improving.