Toward a general tropical forest biomass prediction model from very high resolution optical satellite images

International audienceVery high spatial resolution (VHSR) optical satellite imagery has shown good potential to provide non-saturating proxies of tropical forest aboveground biomass (AGB) from the analysis of canopy texture, for instance through the Fourier Transform Textural Ordination method. Empirical case studies however showed that the relationship between Fourier texture features and forest AGB varies across forest types and regions of the world, limiting model transferability. A better understanding of the biophysical mechanisms on which canopy texture – forest AGB relation relies is a prerequisite to move toward broad scale applications. Here we simulated VHSR optical canopy scenes in identical sun-sensor geometry for 279 1-ha tropical forest inventory plots distributed across the tropics. Our aim was to assess the respective merits and complementarity of two types of texture analysis techniques (i.e. Fourier and lacunarity) on a set of forests with contrasted structure and geographical origin, and develop a general texture-based approach for tropical forest AGB mapping. Across forests, Fourier texture captured a gradient of stands mean crown size reflecting well the progressive changes in stand structure throughout forest aggradation phase (e.g. Pearson's r=−0.42 with basal area) while lacunarity texture captured a gradient of canopy openness (, i.e. Pearson's r=−0.57 with stand gap fraction). Both types of texture indices were highly complementary for predicting forest AGB at the global level (so-called FL-model). The residual error of the FL-model was structured across sites and could be partially captured with a bioclimatic proxy, further improving the performance of the global model (so-called FLE-model) and reducing site-level biases. The FLE model was tested on a set of real Pleiades images covering a mosaic of high-biomass forests in the Congo basin (mean AGB over 49 field plots: 359±98Mgha−1), leading to a significant relationship (R2=0.47 on validation data) with reasonable error levels (