Cooperative Institute for Research in the Atmosphere

Tropical cyclone intensity prediction based on satellite cloud feature extraction and machine learning

Convection intensity and organization are significant contributors to the intensification of tropical cyclones but suffer from the lack of in situ observations over the ocean. This study developed i) the geostationary satellite cloud feature extraction (CFE) algorithm to quantify the dynamic process of Tropical cyclone (TC) rapid intensification (RI) and ii) machine-learning-based TC intensity prediction. In the CFE algorithm, we newly developed satellite indices to quantify the degree of convection organization, such as Largest Patch Index and Effective Mesh Size, in addition to the well-known indices of overall convection intensity and symmetry from the satellite images. Regular observations from geostationary satellites in RI TCs in comparison with slow- and neutral-intensifying TCs in the western North Pacific for 2015-2019 are used to extract the time series of primary convective features related to RI. We constructed a model based on deep convolutional neural networks based on the CFE algorithm and re-analysis data to predict 6, 12, and 24-hour TC intensity. Machine-learning-based TC intensity prediction with multi-scale TC modulating factors is expected to reduce the TC intensity forecast errors.