Abstract: Large footprint lidar ICESat/GLAS data contain lots of vertical structure information of ground including forest vertical profile and terrain, and these parameters have high correlation with above ground biomass. Based on waveform data process, we extracted waveform parameters, and used linear stepwise regression model and Erf-BP neural network model to set up the relation between waveform parameters and forest above ground biomass, respectively. We used the Erf-BP neural network model and waveform data of GLAS to inverse the biomass of GLAS footprints. We used a machine learning approach named random forests and the MISR data to build biomass model to extend point to planar and validated the accuracy of this model with the field measured data from plots. The prediction accuracy of Erf-BP neural network model (P=0.965, R_MSE=3.81 t·ha^-1) was better than that of the linear stepwise regression model (P=0.86, R_MSE=4.54 t·ha^-1). The predicting precision of the biomass spatial extensive model was good (P=0.81, R_MSE=2.39 t·ha^-1), the range of inverse biomass in research area was 0-144.4 t·ha^-1, and the mean of the biomass was 59.28  t·ha^-1. The retrieval precision verified by sample data was high (R²=0.72, R_MSE=8.98 t·ha^-1). The estimation was close to real value. Therefore, using few field measurement data to inverse forest ABG in large scale and high precision could provide an effective approach for forestry inventory, ecology research and carbon investigation.

Key words: biomass, ICESat/GLAS, MISR