Binary classifier performance for 5 × 10⁸ M_⊙ when training on images generated from the middle slice (Z ∈ [−20, 20] kpc) of the simulation box. The different coloured bands depict the performance when training on different features: red – overdensity, yellow – divergence, green – mean speed, and blue – speed dispersion. Model performance where training was done on Gaussian smoothed features is depicted by the solid lines, whereas dashed lines show when training was done on unsmoothed features. We observed that best performance is achieved by using smoothed features out of which overdensity and velocity divergence are most effective.