For large area data collection and environmental monitoring in remote areas with no mobile network coverage, this article first designs a LoRa communication protocol between the UAV mobile gateway and the ground nodes. Based on this, a spreading factor prediction model based on the improved extreme learning machine ( PG-ELM) is proposed to achieve dynamic optimization and adjustment of the spreading factor. To improve the prediction accuracy and efficiency, the model uses signal strength, signal-to-noise ratio, distance, packet loss rate, temperature and relative humidity as inputs. The particle swarm optimization algorithm and the grey wolf optimization algorithm are fused to optimize the ELM model. The LoRa communication data sample sets are obtained through the UAV mobile communication experiment, which are then used to train and optimize the PG-ELM model. The results show that, with a data size of 20 kB, the proposed scheme reduces the data collection time by about 78% and 26% compared with single SF12 and SF7. It also lowers the average communication energy consumption by more than 70% compared with single SF12, achieves a packet delivery rate of 98% , and has significant advantages in energy efficiency and prediction real-time performance.