A Local Dwarf Galaxy Search Using Machine Learning

We present a machine learning search for local, low-mass galaxies ($z < 0.02$ and $10^6 M_\odot < M_* < 10^9 M_\odot$) using the combined photometric data from the DESI Imaging Legacy Surveys and the WISE survey. We introduce the spectrally confirmed training sample, discuss evaluation metrics, investigate the features, compare different machine learning algorithms, and find that a 7-class neural network classification model is highly effective in separating the signal (local, low-mass galaxies) from various contaminants, reaching a precision of $95\%$ and a recall of $76\%$. The principal contaminants are nearby sub-$L^*$ galaxies at $0.02 < z < 0.05$ and nearby massive galaxies at $0.05 < z < 0.2$. We find that the features encoding surface brightness information are essential to achieving a correct classification. Our final catalog, which we make available, consists of 112,859 local, low-mass galaxy candidates, where 36,408 have high probability ($p_{\rm signal} > 0.95$), covering the entire Legacy Surveys DR9 footprint. Using DESI-EDR public spectra and data from the SAGA and ELVES surveys, we find that our model has a precision of $\sim 100\%$, $96\%$, and $97\%$, respectively, and a recall of $\sim 51\%$, $68\%$ and $53\%$, respectively. The results of those independent spectral verification demonstrate the effectiveness and efficiency of our machine learning classification model.