Abstract: Characterisation of forest structure is a major stake for forestry, species conservation, carbon stock estimates and many forest ecology and management issues. At large scale natural forest structure tends to vary according to climate and geomorphomology (Paget, 1999; Steege et al., 2006) while soil characteristics (and notably water regime) and syMgenetic stage add some finer scale variation (Oldeman, 1989; Sabatier et al., 1997). Forest structure characterisation traditionally relies on field-based collection of individual tree dimensions such as stem diameter and stem height sampled across tracks of forest (Hall et al., 1998). However, such field intensive methods are costly, and of low accuracy regarding measures of tree heights. Airborne light detection and ranging (LiDAR) technology provides horizontal and vertical Information at high spatial resolutions and vertical accuracies (Lim et al., 2003; Hyyppä et al., 2004). It has the potential for gathering vegetation structure data over large areas rapidly at moderate cost and hence is of particular relevance for poorly sampled, difficult to access and largely unexplored tropical rainforests. In this study we examined the ability of airborne LiDAR to detect spatial changes in the structure of dense tropical rain forest and we probed this remote sensing approach against local statistics derived from stem diameters (i.e. classical field data information) mapped across a large track of forest at a long term experimental site in French Guyana. The large variability in forest structure occurring at the experimental site is du to natural variation of the soil cover (and notably drainage properties) combined with various logging intensities applied 15 years before the LiDAR data were acquired. On this basis ten different forest types were identified at the site (figure 1 and 3). Various stem based statistics were computed for a series of meshes with cells ranging from 30 by 30 m plots to 250 by 250 m plots. These statistics included basal area, stem density, quadratic mean diameter, and diameter distribution percentiles. Similarly local statistics were extracted either from the Canopy Height Model (e.g. median height, mean height, standard height deviation, height coefficient of variation, height percentiles, frequency of hits below 5 m above ground level). Additionally a wetness index (Böhner et al., 2002) was computed at each node of a 5 by 5 m grid from the Digital Terrain Model also extracted from the LiDAR data set. We used both types of cell statistics to discriminate the various forest types. Comparison between the two approaches for a range of spatial resolution is available from in table 1. Results indicate that LiDAR based statistics are essentially as powerful as field based statistics to discriminate forest types at coarse scale. This reflects the very strong correlation between the CHM and the field based stem diameter data. For example (figure 5) the Pearson correlation coefficient between median height and quadratic mean diameter for cells of 125 by 125 m is 0.945 (n=0.72). When a finer resolution is required however as for the detection of seasonally flooded bottomland forest along thalwegs, then LiDAR technology proves more efficient than field based inventories as it combines information from the DTM and the CHM. The wetness index alone correctly retrieves about 2 thirds of the seasonally flooded areas. All in all, discriminant analysis performance of the LiDAR derived information approaches 80% when classifying forests cover at the finest scale of 5 by 5m into 10 different types and reaches 87% when a coarser classification Into 6 forest types is considered (figure 4).