Estimation of above-ground biomass of grassland based on multi-source remote sensing data.

Grassland is one of the most widely distributed terrestrial ecosystems on Earth. It is critical to accurately estimate grassland biomass of the desert steppe, and to understand its dynamics changes in order to study the regional carbon cycle and the sustainable use of grassland resource. The integration of multisensory data provides the opportunity to explore the benefits of grassland biomass effective estimation via multiple data sources. Based on the field survey data, quadrature-polarization (qual-pol) RADARSAT-2 Synthetic Aperture Radar (SAR) C-band data was utilized to develop a biomass regression model and estimate the aboveground biomass (AGB) of the Caragana microphylla shrubbery in the desert steppe region in the northwest of China. The research area was located at Yangzhaizi Village in Ningxia Autonomous region. Grassland inventory was carried out in 45 randomly distributed plots (30 m × 30 m), and the data was used for either model development or validation. An allometric regression model was established to estimate its biomass for every Caragana microphylla shrub with CH (crown width multiple plant height) variable. The local allometric regression equation was applied to calculate AGB per plot. Furthermore, the correlation between the aboveground biomass of Caragana microphylla shrubbery and the radar backscatter coefficient was analyzed. The AGB regression model was developed by integrating field measurements of 25 sample plots with RADARSAT-2 backscatter remotely sensed data. The multiple stepwise regressions algorithm was applied to develop the AGB model and estimate the grassland above-ground biomass from RADARSAT-2 backscatter data. The developed model was validated by using 20 independent sample plots. Simultaneously, RADARSAT-2 images were fused with the optical HJ1B data by using a discrete wavelet transform for the land cover classification. The image classification based on the objects was performed by using the empirical-statistical machine learning techniques, such as a classification and regression trees (CART) algorithm. The overall accuracy and Kappa value of the proposed method was 90.2% and 0.88, respectively. It indicated that the proposed method performed well for the land use and land cover (LULC) classification. An AGB biomass distribution map was produced by RADARSAT-2 backscatter data in combination with the land cover classification image and AGB regression model. As a comparison, the AGB from RADARSAT-2 estimates were compared with the results from the HJ1B normalized difference vegetation index (NDVI) model. The result showed that there was a good quantitative relationship between the AGB from the microphylla shrubbery and the RADARSAT-2 radar backscatter coefficient. A good fit was found between AGB estimated by RADARSAT-2 and ground-measured biomass with a R² (coefficient of determination) and Root Mean-Square Error (RMSE) of 0.71 and 14.2 kg/hm² respectively. Its estimated accuracy was higher than that of the HJ1B NDVI model (R²=0.27, RMSE=20.58 kg/hm²). Consequently, the fusion of optical and radar data for the land cover classification could effectively improve the accuracy of the object recognition for the land cover classification and the estimation accuracy of AGB estimation. Radar remote sensing data could be used for quantitative studies on grassland structural parameters. Moreover, it demonstrated a high potential for monitoring indicators of grassland ecosystem by combining the optical with polarimetric SAR remote sensing imanges. [ABSTRACT FROM AUTHOR]