Estimation of tropical forest biomass with image texture of radar images

Interest in the world forests has grown to unprecedented heights, especially with growing awareness of their role in the global carbon cycle. Quantifying carbon in forests is therefore of crucial importance for estimating carbon fluxes at the regional and global scale. Carbon quantities are estimated by inferring wood biomass from forest biomass, and then converting it into carbon by using a value of approximately 0.5 ton of carbon for 1 ton of wood. In order to determine the bio-mass of a forest, significant relationships have therefore been established between radar mean intensity and biophysical variables. However, for mature stands (about 80 t/ha and more) increasing biomass reduces the sensitivity of the backscattering coefficient sigma/biomass relationships. Recent studies have shown that texture could be used instead of the usual intensity-age relationships, even for mature stands up to 140 t/ha, the highest biomass value observed for studied forests (monospecific, even-aged forest, subject to identical silvicultural practices and sampling covering all forest stages from sowing to harvest). The present paper aims at extending these observations to tropical forests which is a large component of the terrestrial carbon pool and the carbon sources generated by deforestation in the tropics. Radar images at P-Band were acquired during the TropiSAR experiment in 2009 over the Paracou experimental site with the SETHI ONERA airborne instrument. Paracou is located in a lowland tropical rain forest near Sinnamary, French Guiana where 15 permanent plots of 6.25 ha each were mapped and regularly measured. Three sets of treatments applied to the 15 forest stands provide biomass values from 260 to 470 T/ha. Plots were selected inside the 15th experimental stands with paying attention to the local topogra-phy. Plots with similar slopes were thus compared. Statistical features were then derived a) from gray level statistics (mean sigma, variance, skewness...) and b) the statistics of pixel pairs (energy, contrast, correlation...) for each plot on the basis of the gray level co-occurrence matrix. It is shown for radar images at P-band and polarisation HV that despite the very homogeneous shape of this regenerating forest, linear relationships between some statistical features and forest biomass can be established which does not saturate even for biomass of more than 350 t/ha. These preliminary results are encouraging and further analysis should be carried out to explore the influence of the different treatments on the retrieval performance.