At z>1, the distinction between merging and 'normal' star-forming galaxies based on single band morphology is often hampered by the presence of large clumps which result in a disturbed, merger-like appearance even in rotationally supported disks. In this paper we discuss how a classification based on canonical, non-parametric structural indices measured on resolved stellar mass maps, rather than on single-band images, reduces the misclassification of clumpy but not merging galaxies. We calibrate the mass-based selection of mergers using the MIRAGE hydrodynamical numerical simulations of isolated and merging galaxies which span a stellar mass range of $10^{9.8}$-$10^{10.6}M_{sun}$ and merger ratios between 1:1-1:6.3. These simulations are processed to reproduce the typical depth and spatial resolution of observed HUDF data. We test our approach on a sample of real z~2 galaxies with kinematic classification into disks or mergers and on ~100 galaxies in the HUDF field with photometric/spectroscopic redshift between 1.5$⩽ z ⩽$3 and $M>10^{9.4}M_{sun}$. We find that a combination of the asymmetry $A_{MASS}$ and $M_{20, MASS}$ indices measured on the stellar mass maps can efficiently identify real (major) mergers with $\lesssim$20% contamination from clumpy disks in the merger sample. This mass-based classification cannot be reproduced in star-forming galaxies by $H-$band measurements alone, which instead result in a contamination from clumpy galaxies that can be as high as 50%. Moreover, we find that the mass-based classification always results in a lower contamination from clumpy galaxies than an $H-$band classification, regardless of the depth of the imaging used (e.g., CANDELS vs. HUDF). ; Comment: Accepted for publication in ApJ. The complete Table 1 can be found at this URL: http://www.phys.susx.ac.uk/~ac625/Table1_full.pdf