The number of moles of components in reaction systems can be decomposed into reaction variants (states that change with the progress of reactions) and reaction invariants (states that do not change with progress of reactions). The concept of reaction variants/invariants is used in modeling, control, and design applications. For computation of reaction variants, it has been shown that a subset of components needs to be measured and labeled as reference components. This idea ensures that measurement of a subset of components is sufficient in order to estimate the unmeasured components using reference components and known information on flow rates, inlet compositions, volume, and initial conditions. There are several feasible subsets of components which can be measured for computation of reaction variants and estimation of unmeasured components. In practice, measurements are corrupted with noise. Further, there is a cost associated with measurement of each component due to the nature of the instrument or protocol. In this work, we address the problem of selection of reference components or measurements. First, structural conditions for selecting the minimum number of reference components are revisited for homogeneous and heterogeneous reaction systems. It is shown that the component mole numbers lie in a lower dimensional affine subspace under the assumption of known inlet flow rates and compositions, initial conditions, and volumes. Then, we formulate optimization problems for optimal selection of reference components that maximize the quality of the estimates with constraints on cardinality (number of components to be measured) and cost associated with measurement. These formulations are demonstrated using examples of homogeneous reactions and heterogeneous reactions for different reactor configurations and cost associated with sensors.