An applicable method to retrieve cloud top height (CTH) is proposed. Using the ATrain data in 2007, CTH in low latitude area (15°S~15°N) is retrieved. Firstly, using the brightness temperatures at Channels 31 and 32 of MODIS, clouds are classified into three categories including opaque clouds, semitransparent clouds and transparent clouds based on SVM method with 90.6% accuracy. Then, the CTH of the three categories are retrieved by kernel regression.The root mean square errors (RMSE) are 0.95 km, 1.17 km and 1.27 km, respectively, compared with CloudSat 2BGEOPROFLIDAR product.The errors of the three categories of clouds all decrease after being classified. Lastly, three typical cases are analyzed. The algorithm could be carried out in other satellites containing infrared split window channels.