The scattering coefficient of Lightning Whistler (LW) is a key physical quantity for exploring space environment information. Currently, the scattering coefficient of each LW is mainly extracted manually. However, with the daily production of 20G data from the Zhangheng satellite, manual methods are no longer able to meet the demand for massive data. This paper explores an automatic method for extracting the scattering coefficient of LW: First, detailed data of the Zhangheng satellite’s induction magnetometer payload (SCM) is extracted using a sliding window of 0.8 seconds, and it is converted into a time-frequency graph and an audio file. Second, a YOLOV5 neural network is created to automatically locate a single LW and output its time-frequency position information based on its time-frequency graph. The corresponding audio containing the LW is then intercepted according to this time-frequency position information and zero-padded into an audio segment of 0.8 seconds. Finally, the Mel-frequency cepstral coefficients (MFCCs) of the audio data are extracted, and a GRU neural network improved based on multi-head attention mechanism is created. The MFCCs feature is input into the network, which automatically outputs the scattering coefficient. Experimental results on the dataset in February 2020 showed that the average absolute error and average absolute percentage error were 0.453 and 0.176, respectively. Compared with the GRU model and the Unet neural network-based LW scattering coefficient automatic extraction algorithm proposed by Feng Xiaokang, the average absolute error was reduced by 1.908 and 1.079, respectively, and the average absolute percentage error was reduced by 0.7 and 0.148, respectively. The average processing time per image was 0.074 seconds. The method proposed in this paper can extract the scattering coefficient of LW more accurately and quickly.