A federated learning-based approach for land surface temperature prediction considering cost-efficiency in seasonally frozen ground region

The communication environment in traffic corridors such as the Sichuan-Tibet Railway is generally unsatisfactory, making it challenging to upload Land surface temperature (LST) monitoring data to a cloud server for large-scale and accurate prediction in the seasonally frozen ground region. Instead, the LST prediction can be conducted individually on local edge devices; but, the prediction accuracy may decrease due to insufficient monitoring data. Here, we propose a federated learning-based approach for LST prediction considering cost-efficiency in seasonally frozen ground region. The proposed method can reduce communication cost while achieving higher accuracy in prediction. Federated learning is a machine learning paradigm that enables the training of models on decentralised data sources. In the proposed method, (1) the initial parameters are distributed to monitoring stations; (2) the parameters are updated by monitoring stations using local data and aggregated to obtain the latest global parameters; (3) the latest parameters are delivered to the selected clients for the subsequent updates. The proposed method was applied at three monitoring stations in Southeast Tibet by comparing with the individual learning-based approach. Results indicated that the proposed approach achieves an average reduction of 14.0% in Mean Absolute Error and 13.6% in Root Mean Square Error.