The damage caused by forest fires to humans and the environment cannot be ignored. However, there are few works about the traceability of tree smoke in current time. In this paper, a new system of laser-induced breakdown spectroscopy (LIBS) combined with machine learning was developed to detect and identify the smoke from different tree species. In this new system, spectral lines of heavy elements including Fe, Sr and Ba can be observed, which, together with spectral lines of some other common elements, a total of 57 lines, are regarded as characteristic fingerprints of smoke spectra. These spectral lines are used as variables to reduce the dimensionality of the data using principal component analysis (PCA). Then, a supervised machine learning algorithm of back propagation artificial neural networks (BP-ANN) are applied to identify the source of tree smoke. The average precision rate for smoke achieved over 85%, indicating that LIBS combined with machine learning has a wide application prospect in situ online detection and identification of local forest fires.