Improved tropical deep convective cloud detection using MODIS observations with an active sensor trained machine learning algorithm

Tropical deep convective clouds (DCCs) play an essential role in the Earth system. However, separating DCCs from thick anvils with satellite passive measurements is challenging. In this work, a day/night unified tropical DCC detection approach for MODIS observations is developed with a CloudSat active sensor measurement trained machine learning model. This approach combines brightness temperatures (BTs) at 6.7 μm (BT6.7) and 11 μm (BT11) with derived variables describing horizontal structures of DCCs and improves the tropical DCC probability of detection (POD) from 45.0% for the widely-used BT11 < 215 K method to 67.9%, while the false alarm ratio (FAR) decreases from 43.9% to 32.6%. The Heidke skill score (HSS) of the new algorithm is 0.425, which is more than doubled compared to the HSS of the BT11 < 215 K method (0.182). Moreover, the new algorithm has small biases in DCC's total amount and spatial distribution compared with the CloudSat results. The monthly DCC variations from MODIS wide-swath measurements with the new algorithm demonstrate the ability to reliably provide high temporal and spatial resolution DCC distributions to support intra-seasonal to interannual climate variability studies, which the CloudSat narrow-swath measurements lack enough samples to do. Meanwhile, the new algorithm can also be applied to measurements from other satellite passive sensors when BT6.7 and BT11 are available.