ESTIMATION OF REGIONAL FOREST ABOVEGROUND BIOMASS COMBINING ICESAT-GLAS WAVEFORMS AND HJ-1A/HSI HYPERSPECTRAL IMAGERIES
Keywords: LiDAR, ICESat-GLAS, Waveform, HJ-1A/HSI, Hyperspectral imagery, Forest aboveground biomass, Support vector machines
Abstract. Estimation of forest aboveground biomass (AGB) is a critical challenge for understanding the global carbon cycle because it dominates the dynamics of the terrestrial carbon cycle. Light Detection and Ranging (LiDAR) system has a unique capability for estimating accurately forest canopy height, which has a direct relationship and can provide better understanding to the forest AGB. The Geoscience Laser Altimeter System (GLAS) onboard the Ice, Cloud, and land Elevation Satellite (ICESat) is the first polarorbiting LiDAR instrument for global observations of Earth, and it has been widely used for extracting forest AGB with footprints of nominally 70 m in diameter on the earth's surface. However, the GLAS footprints are discrete geographically, and thus it has been restricted to produce the regional full coverage of forest AGB. To overcome the limit of discontinuity, the Hyper Spectral Imager (HSI) of HJ-1A with 115 bands was combined with GLAS waveforms to predict the regional forest AGB in the study. Corresponding with the field investigation in Wangqing of Changbai Mountain, China, the GLAS waveform metrics were derived and employed to establish the AGB model, which was used further for estimating the AGB within GLAS footprints. For HSI imagery, the Minimum Noise Fraction (MNF) method was used to decrease noise and reduce the dimensionality of spectral bands, and consequently the first three of MNF were able to offer almost 98% spectral information and qualified to regress with the GLAS estimated AGB. Afterwards, the support vector regression (SVR) method was employed in the study to establish the relationship between GLAS estimated AGB and three of HSI MNF (i.e. MNF1, MNF2 and MNF3), and accordingly the full covered regional forest AGB map was produced. The results showed that the adj.R2 and RMSE of SVR-AGB models were 0.75 and 4.68 t hm−2 for broadleaf forests, 0.73 and 5.39 t hm−2 for coniferous forests and 0.71 and 6.15 t hm−2 for mixed forests respectively. The full covered regional forest AGB map of the study area had 0.62 of accuracy and 11.11 t hm−2 of RMSE. The study demonstrated that it holds great potential to achieve the full covered regional forest AGB distribution with higher accuracy by combing LiDAR data and hyperspectral imageries.