Challenges in vision based prediction of lightning peak currents using xception-BiLSTM regression

Accurate characterization of return-stroke peak current – the key design parameter for surge arresters and grounding systems – is critical for designing effective lightning protection. A Phantom v310 high-speed camera, operating at 15 000 fps, recorded 270 downward-negative first stroke stepped leader events over urban and semi-urban Johannesburg between 2016 and 2021. Image sequences were used as input to predict return-stroke peak current. Our architecture employs transfer-learned Xception convolutional blocks (truncated by 10 layers), whose feature maps feed into a bidirectional LSTM stack followed by fully connected layers. We systematically assessed: (1) three model variants of increasing complexity (vA→vB→vC), (2) two sequence-processing strategies (terminal-frame windows versus padded downsampling) and (3) input lengths ranging from 10 to 90 frames. The best performance was achieved with variant vB using the 20 frames before attachment and excluding strokes exceeding 55 kA, yielding a mean absolute error of 6.51 kA and a coefficient of determination ($R^2$) of 0.30. This is the first demonstration of a neural network predicting peak current directly from stepped-leader imagery while highlighting salient spatio-temporal leader dependencies. Our results lay the groundwork for vision-based deep-learning investigations into stepped-leader physics.