Deep learning-based monitoring of overshooting cloud tops from geostationary satellite data

Overshooting tops can cause a variety of severe weather conditions, such as cloud-to-ground lightning, strong winds, and heavy rainfall, which can affect flight and ground operations. Many previous studies have developed overshooting top (OT) detection models. However, rather than identifying individual pixels in satellite images as OTs or non-OTs, we typically find OTs through visual inspection of the contextual information of pixels (i.e., dome-like shape). Such an approach is more intuitive, accurate, and generalizable regardless of the OT characteristics that are used in the existing OT detection algorithms. In this paper, a new approach is proposed for OT detection using deep learning, more specifically a convolutional neural network (CNN), which can mimic the human process by convolution operation. Himawari-8 satellite images were used as input data, which were chopped into patches (i.e., grids) with a 31 × 31 window size and binary detection (OT or non-OT) for each patch was the output of the model. The validation results show that CNN can be successfully applied for the detection of OTs over the tropical regions, showing a mean probability of detection (POD) of 79.68% and a mean false alarm ratio (FAR) of 9.78%.