Abstract: [Objective] By comparing the GLAS data denoising algorithm based on different window functions and the accuracy of different inversion models of forest crowns, the waveform denoising algorithm is optimized and the inversion model with high estimation accuracy is obtained, which can lay the foundation for the study of forest biomass estimation.[Method] Firstly, the GLAS data are denoised by the Blackman window function and the Gaussian window function, and the effects of two kinds of waveform denoising method is quantitatively compared with the root mean square error(RMSE)and the signal to noise ratio(SNR). Then, the waveform data denoised by the window function that produced a better denoising effect were used to extract waveform data parameters. Four forest types (coniferous forests, broadleaf forests, mixed coniferous and broadleaf forests, and all forests)were included in this study. The linear regression method was used to develop a waveform-parameter model with the waveform length and a topographic-index model with the waveform length and the topographic-index. A full-parameter model was based on the topographic-index model with the height of median energy, the waveform leading edge, and the waveform trailing edge. Finally, the result of the crown height estimations of these three models were compared.[Result] The result show that the Gaussian window function gave a lower RMSE and a higher SNR than the Blackman window function. This indicates that use of the Gaussian window function for denoising provided more accurate data. For all forest types, the predictions of crown height given by the full-parameter model were more accurate than those given by other two models. The predictions of crown height for the coniferous forest given by the topographic-index model(R²=0.853, RMSE=2.519 7 m) were more accurate than the predictions of that model for other forest types. The predictions given for the mixed coniferous and broadleaf forest by the full-parameter model(R²=0.972,RMSE=1.001 4 m)were more accurate than the predictions of that model for other forest types.[Conclusion] In conclusion, the Gaussian window function performed better than the Blackman window function in denoising the GLAS waveform data. For the waveforms reflected from complex terrain, when the new multiple linear regression model was developed that included several waveform data parameters and the topographic-index, the performance of the model in interpreting the maximum crown height was significantly improved. This makes our model overcome the difficulty of interpreting the maximum crown height of forest on the highly sloping terrain, and thus an accurate estimation of forest crown height on complex terrains can be achieved.

Key words: large footprint laser radar, Blackman window denoising, Gaussian window denoising, forest crown height inversion model