Recent monitoring applications extensively rely on low-power wide-area networks, such as those provided by LoRa and LoRaWAN, in order to enable end-devices to communicate over long distances. However, in large-scale deployments, the small throughput of LoRa is further reduced due to collisions. In this paper, we propose a new scheme SF-DS that aims to decode colliding frames. SF-DS relies on frequency detection at each symbol frontier. It reduces detection errors when symbols of similar values are superposed. It is compatible with legacy LoRaWAN communications, and only requires modifications at the gateway. Our simulation results show that SF-DS is able to decode more collisions than the other protocols from the literature (about twice more for 16 nodes and SF12), therefore increasing the throughput and thus the scalability of the network.