Correction for remote estimation of return-stroke peak current using semi-supervised learning

The return-stroke peak current is an important parameter provided by the lightning location system and plays an important role in atmospheric electricity. In this paper, a new method combining classical estimation methods and semi-supervised learning techniques is proposed for more accurate remote estimation of return-stroke peak currents. The model achieves the prediction of propagation error, and hence the correction of return-stroke peak current estimation, in a large amount of unlabeled data by using the pseudo labels established by the multi-station consistency assumption to train the feature extraction capability, and then fine-tuning the output layer with a small number of 100 (±10) km real electric field strengths as the target. The results show a 22 % improvement in the mean value of the corrected estimated peak current, an improvement in the coefficient of determination from 0.8362 to 0.9178 and a reduction in the root mean square error from 21.51 % to 10.67 % with respect to the peak current calculated using empirical equation derived from rocket-triggered lightning and the electric field at close range. The proposed method is not only applicable to the Chinese region and our network but also can be extended to the global and other networks.