Identifying the location of nodes in wireless networks is a major challenge because its accuracy impacts the efficiency of location-aware protocols and applications. This paper presents a novel localization scheme called Multi-hop Localization through Model Selection (MLMS) which can significantly improve localization accuracy in different irregular networks. The proposed scheme contains three steps: data collection, establishing the skeleton model, and location estimation. In the data collection step, the proximity information of the irregular network is collected. In the establishment of the skeleton model step, the skeleton model among the practical distances and the proximity among nodes is constructed using regression which is supervised by Bayesian Information Criterion (BIC). Specifically, the skeleton model when the BIC value is the smallest is deemed to be optimal. In the location estimation step, any unknown node calculates its location in a distributed manner using the optimal skeleton model. The simulation results demonstrated that the proposed scheme can greatly reduce the estimation error and quickly achieve estimation location in networks with different irregular topologies. Simulation results show that, compared with similar algorithms recently reported, MLMS improves localization accuracy by more than 32 %.