Multi-object spectroscopic galaxy surveys typically make use of photometric and colour criteria to select targets. Conversely, the Euclid NISP slitless spectrograph will record spectra for every source over its field of view. Slitless spectroscopy has the advantage of avoiding defining a priori a galaxy sample, but at the price of making the selection function harder to quantify. The Euclid Wide Survey aims at building robust statistical samples of emission-line galaxies with fluxes in the Halpha-NII complex brighter than 2e-16 erg/s/cm^2 and within 0.9<z<1.8. At faint fluxes, we expect significant contamination by wrongly measured redshifts, either due to emission-line misidentification or noise fluctuations, with the consequence of reducing the purity of the final samples. This can be significantly improved by exploiting Euclid photometric information to identify emission-line galaxies over the redshifts of interest. To this goal, we compare and quantify the performance of six machine-learning classification algorithms. We consider the case when only Euclid photometric and morphological measurements are used and when these are supplemented by ground-based photometric data. We train and test the classifiers on two mock galaxy samples, the EL-COSMOS and Euclid Flagship2 catalogues. Dense neural networks and support vector classifiers obtain the best performance, with comparable results in terms of the adopted metrics. When training on Euclid photometry alone, these can remove 87% of the sources that are fainter than the nominal flux limit or lie outside the range 0.9<z<1.8, a figure that increases to 97% when ground-based photometry is included. These results show how by using the photometric information available to Euclid it will be possible to efficiently identify and discard spurious interlopers, allowing us to build robust spectroscopic samples for cosmological investigations.