NESSI 2021 - Prahalath Anbu Bharathi
Explainable AI for Short-term Lightning Prediction
Severe weather kills over 500 people and causes billions of dollars in property damage every year in the United States. Short-term lightning prediction can help reduce some of these losses. Providing lead time before lightning strikes allows communities to prepare for hazardous weather. Deep learning has shown promising results in other areas of atmospheric science and here we apply it to lightning detection. We train a ResNet neural network (NN) on infrared satellite imagery from the Advanced Baseline Imager and the Geostationary Lightning Mapper of the GOES-16 satellite. Specifically, the NN model has been trained with a lead time of twenty minutes from the following bands of the GOES-16 satellite: tropospheric water vapor (bands 8, 9, 10) and infrared longwave window (band 14). The current model is 95% accurate at predicting lightning; however, it performs poorly when differentiating between high and low lightning count images. We turned to explainable AI methods, Integrated Gradients and Deeplift, to interpret the model’s predictions. Integrated Gradients and Deeplift provide pixel-wise attributions for a NN’s prediction through gradient and backpropagation-based approaches, respectively. These methods have provided insight into the important features learned by the model, like overshooting cloud tops. Preliminary results have indicated that band 14 contributes the most towards the prediction. Retraining the ResNet with only band 14 data achieves comparable accuracy to the original model. Future work will explore the use of the other 12 bands for lightning prediction.
Mentor: John Schreck
Slides and Poster