Estimates of Forest Growing Stock Volume of the Northern Hemisphere from Envisat ASAR
Santoro, Maurizio1; Schmullius, Christiane2; Pathe, Carsten2; Schwilk, Julian2; Beer, Christian3; Thurner, Martin3; Fransson, Johan E.S.4; Shvidenko, Anatoly5; Schepaschenko, Dmitry5; McCallum, Ian5; Beaudoin, Andre6; Hall, Ron6
1Gamma Remote Sensing, SWITZERLAND; 2Friedrich-Schiller University, GERMANY; 3Max Planck Institute for Biogeochemistry, GERMANY; 4Swedish University of Agricultural Sciences, SWEDEN; 5International Institute of Applied Systems Analysis, AUSTRIA; 6Natural Resources Canada, Canadian Forest Service, CANADA

The feasibility of mapping forest parameters, such as aboveground biomass or height, at regional to global scale with spaceborne remote sensing data has increased in recent years because of the availability of datasets with full and repeated coverage within short periods of time. Synthetic Aperture Radar (SAR) intensity is particularly suitable for retrieval of forest parameters because forest structural properties affect the response of the radar signal. The availability of large stacks of backscatter measurements acquired by Envisat ASAR in ScanSAR mode has favored the development of an approach, referred to as BIOMASAR, to retrieve forest growing stock volume (GSV) from hyper-temporal datasets of C-band backscatter. The BIOMASAR algorithm exploits a Water Cloud-like model linking the forest backscatter to GSV and a weighted combination of GSV estimates from individual backscatter measurements. The stacks of backscatter measurements are obtained from Envisat ASAR ScanSAR data in Global Monitoring and Wide Swath mode after multi-looking, terrain geocoding, radiometric normalization for slope-induced effects and multi-channel speckle filter. The entire processing chain runs automatically without the need of training data for the forest backscatter model. Unknown model parameters are estimated from central tendency statistics of the backscatter for unvegetated areas and areas with high canopy cover. This information is provided by the MODIS Vegetation Continuous Fields (VCF) product. The estimation of the model parameters and the inversion of the model to retrieve the GSV are implemented at pixel level. The retrieval is constrained within a range of realistic GSV values, which are obtained on a grid of 2 x 2 degrees from different sources (inventory data, inventory reports, other EO datasets of GSV). The weights of the multi-temporal combination correspond to the backscatter difference between dense forest and unvegetated areas.

The BIOMASAR algorithm has been applied recently to obtain estimates of GSV representative for the year 2010 for latitudes above 30° N latitude. The entire processing chain (i.e., SAR data processing and retrieval of GSV) has been implemented on the European Space Agency (ESA) Grid Processing on Demand (G-POD) environment into two separate services. The Envisat dataset consisted of approximately 70,000 image strips acquired in ScanSAR mode and acquired mostly between October 2009 and February 2011. The spatial resolution is 1,000 m; data have been geocoded to a pixel spacing of 0.01 degree. The GSV estimates were refined on local servers following a first evaluation of the data product and interaction with local experts.

Validation of the GSV estimates follows a dual approach. On one hand the BIOMASAR dataset was compared against in situ measurements, where available, and similar data products based on Earth Observation data. On the other hand, regional subsets have been provided to local experts for internal evaluation and validation. Here, we report on the internal validation. Consistent estimates of GSV were obtained across three continents. Trends in the BIOMASAR GSV estimates and reference datasets agreed up to 250-300 m3/ha. Above this level, the BIOMASAR GSV presented increasing underestimation in consequence of the saturation of the C-band signal in very dense forest. The relative difference with respect to reference data was on the order of 15-30% at aggregated level (> 0.5 degree). Larger differences occurred at the full resolution (0.01 degree) in consequence of the limited sensitivity of the SAR backscatter at C-band to GSV.

The validation also indicate a number of shortcomings and bottlenecks when dealing with coarse resolution data in an environment of limited data availability to be used as reference. Forest plot inventory data are not suitable for the validation because the measurement at the scale of a plot represents a different forest structure and composition compared to the kilometric scale. The relative difference on the order of 60-100% obtained in Sweden and Siberia do not represent true retrieval statistics. Inventory units at 1 km scale suitable for correct validation of GSV estimates derived from low resolution remote sensing data are highly desirable but practically do not exist in such form. Hence, rigorous validation of such GSV estimates is not possible. Aggregated inventory data to low resolution present reduced spatial heterogeneity, thus containing information that is closer to the information within a coarse resolution remote sensing pixel. The drawback of comparing aggregated measurements of GSV from inventory data and remote sensing based estimates of GSV is the impossibility to provide conclusions on the spatial distribution of the estimated GSV. Dealing with a research area that included several countries, requires a thorough screening and interpretation of the data to be used as reference. Typically, large-scale averages at the level of provinces, counties or countries are sufficiently reliable to act as reference.