Detection of Aircraft Lightning Potential Areas by using a Deep Neural Network with Interpretability

Aircraft lightning strike is one of issues in the aviation weather field. Once it occurs, aircraft have to be fully inspected, and any damages have to be repaired. When taking serious damages, aircraft have to stop servicing. Even when small or no damages are found, following flight can be delayed by their inspection and first-aid repairs [1].

Japan aerospace exploration agency carried out a feasibility study to develop tactical support information for avoiding aircraft lightning strikes. In the feasibility study, weather observation data with respect to actual flight data (including flights with and without aircraft lightning strikes) were analyzed. The analysis revealed that vertical integrated radar reflectivity and radar reflectivity at altitudes where air temperature is around -10 degC were highly relevant to aircraft lightning strikes which actually occurred. Furthermore, the analysis indicated that aircraft lightning potential areas can be detected by thresholds on the vertical integrated reflectivity and reflectivity at altitudes of -10 degC, and 60—80 % of current aircraft lightning strikes can be reduced by avoiding the detected areas. Figure 1 shows an example of the detected aircraft lightning potential areas [2].

However, it is anticipated that the detection of aircraft lightning potential areas has to be modified depending on situations. For example, proper thresholds would be different between summer and winter lightning areas. Climate of an airport may affect the thresholds. A scheme which determines the thresholds adaptively to an airport is useful or necessary to widely apply the detection of aircraft lightning potential areas.

This work studied for applying a machine learning approach to the detection of aircraft lightning potential areas in order to obtain adaptivity. A deep neural network was designed so that outputs of mid-layers are studied in comparisons to the previous thresholding algorithm in order to possess interpretability. The deep neural network was tested by separating the data set of weather observation and actual flights into training and validation. As a result, the training was properly converged, and the validation showed a detection rate more than 80 % and a false-alarm rate less than 10 % (Figure 2).

[1] WEATHER-Eye Consortium, 2017: Weather-eye vision. Tech. rep., Japan Aerospace Exploration Agency, in Japanese.

[2] E. Yoshikawa and T. Ushio, “Tactical Decision-Making Support Information for Aircraft Lightning Avoidance —Feasibility Study in Area of Winter Lightning,” Bulletin of American Meteorological Society, https://doi.org/10.1175/BAMS-D-18-0078.1.