The simple dependence of the intensity in annular dark field scanning transmission electron microscopy images on the atomic number provides (to some extent) chemical information about the sample, and even allows an elemental identification in the case of light-element single-layer samples. However, the intensity of individual atoms and atomic columns is affected by residual aberrations and the confidence of an identification is limited by the available signal to noise. Here, we show that matching a simulation to an experimental image by iterative optimization provides a reliable analysis of atomic intensities even in presence of residual non-round aberrations. We compare our new method with other established approaches demonstrating its high reliability for images recorded at limited dose and with different aberrations. This is of particular relevance for analyzing moderately beam-sensitive materials, such as most 2D materials, where the limited sample stability often makes it difficult to obtain spectroscopic information at atomic resolution.