The natural tropical rain forest is an irregular mosaic of ever-changing developmental phases of vegetation. Its structure is a key information for forest monitoring and sustainable management. This experimental study aims to improve the understanding of the relationships between the forest structure and the anisotropic behaviour of the remotely sensed forest reflectance. It relies on intensive field-based measurements and synchronous remote sensing data collection. Three different tropical rain forests have been observed using six different SPOT–HRVIR images. Texture information is related to forest structure by means of geostatistical analysis based on directional variograms for the NIR channel. The results demonstrate a strong interaction between the elements contributing to the forest structure, i.e. the tree crowns and the gaps, and the viewing and illumination conditions of each observation. The concept of the Bi-directional Standard deviation Distribution Function (BSDF) is introduced to discriminate the forest structures. A simple geometric-optical gap model explains more than 80% of the variability of the NIR reflectance standard deviation. Finally, a cross-thresholding approach using the NIR reflectance and the Local Directional Contrast (LDC) is proposed and successfully applied to discriminate the canopy roughness of the different forest sites.