Automatic detection of overshooting tops and their properties from visible satellite channels

TL;DR

This paper introduces a CNN-based automated method for detecting overshooting tops and estimating their heights from visible satellite images, achieving high detection accuracy and low height estimation error.

Contribution

It presents a novel approach using visible satellite data and shadow length analysis for OT detection and height estimation, improving over traditional infrared methods.

Findings

Detection probability of 97% in visible imagery

Average height estimation error of 0.25 km

Performance expected to improve with higher resolution satellites

Abstract

Overshooting tops (OTs) are informative indicators of convective storm intensity and are widely utilized in meteorological analyses. This study presents an automated algorithm for OT detection and OT height estimation using convolutional neural networks applied to visible satellite imagery. The models are trained and validated on an extensive OT dataset comprising approximately 10,000 manually detected cases over Europe. The OTs were identified from high-resolution visible (HRV) channel of the SEVIRI instrument on board the MSG geostationary satellite, with the heights determined from the length of their shadows in the imagery. While conventional OT detection methods primarily rely on the identification of cold features in thermal infrared channels, our approach extracts information from visible channels, leveraging the ground truth data on OT shadow length provided by the training dataset. In the morning and afternoon hours, when the shadows are visible, the proposed models detect OTs with a probability of detection reaching 97% and estimate their height with an average error of 0.25 km. The performance is expected to further improve once the model is applied to polar and new generation geostationary satellites with increased spatial resolution.