Adaptive Mesh Refinement (AMR) is an appealing technique to mitigate simulation cost by refining and coarsening the mesh locally and dynamically during a simulation. In feature-based adaptation, metric criteria are used to determine which regions of the flow are of interest and require a fine mesh resolution. For reactive flows, several criteria exist but are rarely compared to assess which one induces minimal error. This paper provides a comparison of gradient-based and Hessian-based criteria and a newly introduced criterion based on the direct closure of subgrid reaction rates. This criterion uses the scale similarity hypothesis with a test filtering of the reaction rates to estimate the unresolved subgrid-scale reaction rates. The subgrid-scale contribution is then used as an error estimator to construct a mesh adaptation metric. First, a 2D hydrogen/air triple flame is used for comparison and assessment of the criteria. The new Scale Similarity Resolved Reaction Rates (SSRRRM) criterion provides the same level of accuracy than more traditional criteria with half the mesh size. Then, large-eddy simulation of the CORIA Rouen Spray Burner (CRSB) is performed with AMR based on the SSRRRM criterion and an aerodynamic vorticity-based criterion. Comparison to prior simulations and experimental results is conducted. The inner flame front presents realistic wrinkling and local extinction of the flame due to turbulent shear stress and droplets traversing the flame front. The AMR cost is largely compensated by the reduction of mesh size. Still, the major advantage of the AMR strategy is to capture all the important turbulent spray flame features without any prior knowledge on the flame topology making it particularly suitable to multi-regime combustion.