arXiv:2610.23682v1 Announce Type: new Abstract: Long range wide area network (LoRaWAN) has a long communication range and low power consumption, making it a potential technology for direct-to-satellite (DtS) Internet of Things (IoT) connectivity. However, packet collisions and Doppler impairments brought on by satellite motion can greatly increase the packet loss ratio (PLR) in large-scale DtS-IoT networks. In this paper, we study resource allocation in DtSIoT LoRaWAN by first developing an analytical PLR model that simultaneously takes Doppler impairments and packet collisions in the temporal and frequency domains into account. In order to maximize resource allocation, the generated PLR is then used as the reward function for two reinforcement learning (RL) schemes: deep Q-network (DQN) and proximal policy optimization (PPO). Simulation results validate the analytical PLR model against Monte Carlo simulations and demonstrate significant improvements over random resource allocation. At a target PLR of 3*10^-2, while random allocation supports only 100 end devices (EDs), PPO and DQN support approximately 650 and 690 EDs, respectively, representing nearly a seven-fold increase. Furthermore, for 800 EDs, PPO and DQN increase the total number of transmitted packets by 106.7% and 88.0%, respectively, compared with random allocation.