The adaptive random forest classifier sets a warning detector and a drift detector on each basic classifier. When the instance is being trained, multiple warning detectors are often triggered at the same time, causing multiple background trees to be trained simultaneously, which requires large memory and long running time. Aiming at this problem, this paper proposes an improved adaptive random forest ensemble classification algorithm. It sets the concept drift detector in the ensemble learning device, removes the drift detectors at each base tree, and determines the number of background trees according to the ensemble prediction accuracy. The improved algorithm classifies balanced data streams. Under the premise of ensuring the classification performance, the running time and the memory consumption is reduced, and the concept drift appearing in the data stream can be more quickly adapted.