Estimating Forest Aboveground Biomass by Combining ALOS PALSAR and WorldView-2 Data: A Case Study at Purple Mountain National Park, Nanjing, China

• Published in: Remote sensing (Basel, Switzerland) Vol. 6; no. 9; pp. 7878 - 7910
• Main Authors: Deng, Songqiu, Katoh, Masato, Guan, Qingwei, Yin, Na, Li, Mingyang
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.09.2014
• Subjects: Biomass; Forest biomass; forest biomass estimation; Mixed forests; National parks; neural network model; PALSAR; Principal components analysis; Remote sensing; stepwise regression; synergy; WorldView-2 data
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs6097878

Enhanced methods are required for mapping the forest aboveground biomass (AGB) over a large area in Chinese forests. This study attempted to develop an improved approach to retrieving biomass by combining PALSAR (Phased Array type L-band Synthetic Aperture Radar) and WorldView-2 data. A total of 33 variables with potential correlations with forest biomass were extracted from the above data. However, these parameters had poor fits to the observed biomass. Accordingly, the synergies of several variables were explored to identify improved relationships with the AGB. Using principal component analysis and multivariate linear regression (MLR), the accuracies of the biomass estimates obtained using PALSAR and WorldView-2 data were improved to approximately 65% to 71%. In addition, using the additional dataset developed from the fusion of FBD (fine beam dual-polarization) and WorldView-2 data improved the performance to 79% with an RMSE (root mean square error) of 35.13 Mg/ha when using the MLR method. Moreover, a further improvement (R2 = 0.89, relative RMSE = 17.08%) was obtained by combining all the variables mentioned above. For the purpose of comparison with MLR, a neural network approach was also used to estimate the biomass. However, this approach did not produce significant improvements in the AGB estimates. Consequently, the final MLR model was recommended to map the AGB of the study area. Finally, analyses of estimated error in distinguishing forest types and vertical structures suggested that the RMSE decreases gradually from broad-leaved to coniferous to mixed forest. In terms of different vertical structures (VS), VS3 has a high error because the forest lacks undergrowth trees, while VS4 forest, which has approximately the same amounts of stems in each of the three DBH (diameter at breast height) classes (DBH > 20, 10 =< DBH =< 20, and DBH < 10 cm), has the lowest RMSE. This study demonstrates that the combination of PALSAR and WorldView-2 data is a promising approach to improve biomass estimation.