The　forest　height　estimation　systems　with　GeoScience　Laser　Altimeter　System　(GLAS)　data　have　been　demonstrated　to　yield　an　accurate　estimation　result　in　the　topographically　flat　areas　with　simple　forest　conditions.　However,　because　of　the　complex　forest　structures,　the　GLAS　waveform　data　is　difficult　to　consistently　characterize　the　forest　feature　of　different　forest　land　cover　classes.　To　obtain　an　effective　and　reliable　estimation　model　without　additional　calibration　steps,　in　this　paper,　we　introduce　a　forest　height　estimation　algorithm　based　on　random　forests　leaf　mean　(FLM)　encoding,　which　combines　the　decision　tree　model　with　linear　regression,　and　obtain　a　better　estimation　performance　than　the　single　regression　model.　The　experiment　results　show　that　the　FLM　encoding　method　proposed　in　this　paper　can　implicitly　characterize　the　complex　forest　information,　and　have　a　positive　impact　on　the　overall　estimation　system　performance.　The　complex　forest　features　can　be　extracted　from　GLAS　waveform　by　our　method　and　be　used　as　a　supplemental　feature　to　improve　the　estimation　of　forest　height.　©　2020　Technical　Committee　on　Control　Theory,　Chinese　Association　of　Automation.