This study compared detection skill for tropical cyclone (TC) formation using models based on three different machine learning (ML) algorithms-decision trees (DT), random forest (RF), and support vector machines (SVM)-and a model based on Linear Discriminant Analysis (LDA). Eight predictors were derived from WindSat satellite measurements of ocean surface wind and precipitation over the western North Pacific for 2005–2009. All of the ML approaches performed better with significantly higher hit rates ranging from 94 to 96% compared with LDA performance (~77%), although false alarm rate by MLs is slightly higher (21–28%) than that by LDA (~13%). Besides, MLs could detect TC formation at the time as early as 26–30 h before the first time diagnosed as tropical depression by the JTWC best track, which was also 5 to 9 h earlier than that by LDA. The skill differences across MLs were relatively smaller than difference between MLs and LDA. Large yearly variation in forecast lead time was common in all models due to the limitation in sampling from orbiting satellite. This study highlights that ML approaches provide an improved skill for detecting TC formation compared with conventional linear approaches.