Abstract:
Woody vegetation cover affects several ecosystem processes including carbon and water cycling, energy fluxes,
and fire regimes. In order to understand the dynamics of savanna ecosystems, information on the spatial distribution
of woody vegetation over large areas is needed. In this study we sought to assess multi-temporal ALOS
PALSAR L-band backscatter to map woody cover in southern African savannas. The SAR data were acquired
from the JAXA archive, covering various modes and seasons between 2007 and 2010. We used high resolution
airborne LiDAR data as reference data to interpret SAR parameters (including backscatter intensities and polarimetric
decomposition components), to develop SAR-basedmodels aswell as to validate SAR-based woody cover
maps. The LiDAR survey was carried out in April 2008 with the Carnegie Airborne Observatory (CAO, http://cao.
ciw.edu). The highest correlations to the reference data were obtained from SAR backscatters of the dry season,
followed by the wet season, and the end of the wet season. The volume components from polarimetric decompositions
(Freeman-Durden, Van Zyl)were calculated for the end ofwet season, and showed similar correlations
to the LiDAR data, when compared to cross-polarized backscatters (HV).We observed increased correlation between
the SAR and LiDAR datasetswith an increase in the spatial scale atwhich datasetswere integrated,with an
optimum value at 50 m. We modeled woody cover using three scenarios: (1) a single date scenario (i.e., woody
cover map based on a single SAR image), (2) a multi-seasonal scenario (i.e., woody cover map based on SAR images
fromthe same year and different seasons, based on key phonological difference), and (3) amulti-annual scenario
(i.e., woody cover map based on SAR data from different years). Predicted SAR-based woody cover map
based on Fine Beam Dual Polarization dry season SAR backscatters of all years yielded the best performance
with an R2 of 0.71 and RMSE of 7.88%. However, single dry season SAR backscatter achieved only a slightly
lower accuracy (R2 = 0.66, RMSE = 8.45%) as multi-annual SAR data, suggesting that a single SAR scene from
the dry season can also be used for woody cover mapping. Moreover, we investigated the impact of the number
of samples on the model prediction performance and showed the benefits of a larger spatially explicit LiDAR
dataset compared to much smaller number of samples as they can be collected in the field. Collectively, our results
demonstrate that L-band backscatter shows promising sensitivity for the purposes of mapping woody
cover in southern African savannas, particularly during the dry season leaf-off conditions.
