Cloud base height (CBH) is a vital parameter in weather forecasting and aviation safety ensuring. However, it is difficult to obtain the CBH data with high spatio-temporal resolution. Currently, the only available CBH data with global coverage and hourly resolution is provided by ERA5 reanalysis, which has large biases. This study evaluates the CBHs from ERA5 reanalysis for the period from 2006 to 2019 based on satellite observations and revises them by machine learning at low latitudes. Overall, the CBHs from ERA5 reanalysis are lower than those from satellite observations except in part desert areas. The ERA5 overestimates the proportion of low CBH and underestimates the proportion of middle and high CBH. Especially at low latitudes, the proportions of CBH between 0 and 1000 m from observations and ERA5 are 41% and 77%, respectively. The bias of CBH from ERA5 at low latitudes depend heavily on specific cirrus clouds, whose formation is closely associated with deep convection, reflecting the poor performance of ERA5 for high clouds. In addition, revised results show that random forest performs the best revision on CBHs from ERA5 at low latitudes, reducing the root mean square error of CBH from 4210 m to 2557 m and increasing the proportion of middle and high CBH. This study points out the uncertainty in CBH from ERA5 and provides a new methodology for revising the CBH at low latitudes.