Abstract
Current climate change affects tropical forests functioning and might jeopardize their role as a global carbon sink. Accurately documenting forest carbon fluxes at a meaningful scale is therefore a pressing challenge. Airborne LiDAR (ALS), which can provide a fine-grained description of canopy structure and dynamics has great potential. This thesis explores the capabilities and limitations of airborne multitemporal LiDAR to map patterns of C fluxes over space and time to reduce uncertainty in models of carbon stocks and fluxes in tropical forests. We relied on a combination of repeated ALS overflights extending over ten years and a large network of plots totaling more than 1.2 km2 of field inventories conducted at the Permanent Research Station of Paracou (French Guiana).The first chapter (Q1. Efflux Modeling Mortality) addresses the possibility of developing reliable estimates of biomass, basal area and stem number loss (efflux) from observed changes in canopy height with repeated ALS overflights and further assesses whether gap dynamics show persistent over time. Absolute basal area loss rate was linearly correlated to gap dynamics at the plot level (R2=0.60) and more strongly so when the analysis was restricted to undisturbed forests (R2 =0.72). The rate of basal area loss was better predicted from gap dynamics than the rate of stem loss. At the landscape scale, LiDAR data revealed that spatial patterns of gap creation were related to local topography and canopy height, where high canopy forests and bottomlands had higher mortality rates. It is concluded that gap dynamics allow tracking the change in forest carbon fluxes, complementing the monitoring of net carbon change derived from static carbon estimates.The second chapter (Q2. Allometry and carbon stock) quantifies the reduction of error in plot-level AGB estimates achieved using locally adjusted height-diameter allometries. Tree height data were obtained either from individually segmented crown in the LiDAR point cloud or from an individual-based forest model (Canopy Constructor) globally adjusted to the Canopy height model. Bayesian multilevel modeling approach incorporated local canopy height and species identity as co-variates. The quadratic error in predicting mean height was reduced by a factor of four by replacing the best universal alternative allometry with the locally derived ALS allometric H-DBH relationship, where nearly half of the reduction in quadratic error was due to reduced bias. A universal model (not adjusted for site-specific H-DBH allometry) underestimated AGB by 12-13 % at the site level. The inclusion of species identity and canopy height was a considerable improvement, dramatically reducing the uncertainty in H prediction.The third chapter (Q3. Influx modeling Productivity) examines whether the canopy height gain derived from ALS can be used to map Aboveground Woody Net Primary Productivity (AGWNPP). The model predicted AGWNPP across the entire range of plots at the spatial resolution of 125m with an R2 of 0.75 and a relative RMSE of 11%. Applying the method to the landscape scale, the maps revealed a spurious spatial pattern of productivity with higher values in seasonally flooded areas. Critically reviewing the model development procedure allowed to identify possible shortcomings. While a final unbiased AGWNPP model was not tested, this analysis also provided an opportunity to critically examine the overall relevance of the exercise. We conclude that measuring change in net carbon change (rather than developing a model of raw carbon influx) might be a more promising avenue, likely to provide more robust and more spatially accurate results than a net primary productivity model.In the last chapter, we synthesize the conclusions of the articles, following with a critical reflection on the work done and discussing the potential of LiDAR in ecology and conservation.