Deep convection can cause a variety of severe
weather conditions such as thunderstorms, strong winds, and
heavy rainfall. Satellite observations provide all-weather and
multi-directional observations, facilitating the timely
detection of such weather systems, which is crucial to saving
lives and property. However, previous methods based on
channel feature extraction and threshold filtering did not
make full use of information in satellite images, which led to
limitations on such complex problems as strong convection
detection. In this study, we propose a novel framework of a
deep learning-based model Convection-UNet to detect
convection. We use channel 4 to 7 of FY-4B GHI that we select
according to the microphysical properties of convection as
input and radar reflectivity as label. We combine the detailed
training time and test time data augmentation strategies and
build a deep neural network to automatically extract spatial
context features and achieve end-to-end learning. Results
show that the performance of our method far exceeds the
previous channel extraction combined with threshold filtering
methods such as BT and BTD at least 0.24 on F1-measure.
We also show that our channel selection and data
augmentation strategies are of great significance to detect
convection.