The somatic mutation detection could contribute to revealing the pathogenesis and ultimately exploring the prevention and treatment of diseases, as well as in the development of targeted drugs and personalized treatment. Due to that sequencing data contain a lot of noise, it is very difficult to identify somatic variation. In this paper, Multi-Grained Cascade Forest (gcForest) is utilized to identify somatic mutations from whole-genome and exome sequencing datasets. The attributes of each somatic mutation are input into the multi-grained scanning part for preprocessing. The features spliced are input into cascade forest part for training. Three real somatic mutation datasets from exome capture data and whole genome shotgun data are utilized. The results show that gcForest could obtain the better ROC curves and AUC values than random forest (RF), support vector machine (SVM) and forest graph-embedded deep feedforward network (forgeNet). gcForest can detect somatic mutations more accurately.