Abstract
Tropical forests, representing 6.4% of the Earth's surface, host the greatest biodiversity of any terrestrial ecosystem and play a fundamental role in the carbon cycle on a worldwide scale. The sustainable use of tropical forests is a fundamental issue from both the point of view of biodiversity conservation and the reduction of emissions from deforestation and forest degradation (REDD+). The Office National des Forêts is responsible for the conservation and management of 6 million hectares of forests in French Guiana. The possibility of mapping species in the canopy by remote sensing is of obvious interest. For both applied and scientific purposes, the use of airborne observation measurements can enable local field information that is difficult to collect over large areas of tropical forest to be extrapolated.The specific spatialized inventories at the landscape scale would contribute to advancing fundamental knowledge of this complex and threatened biome and assist in its sustainable management. Maps of species distribution can in fact be cross-referenced with maps of environmental factors and thus provide keys for interpreting the organization patterns of forest stands. From a management point of view, species distribution maps are an help to the rationalization of forestry operations. The mapping of commercial species could promote forestry practices that minimize the environmental impact of logging. The identification of species would in particular enable priority to be given to areas that are particularly rich in commercial species, while avoiding the opening up of exploitation tracks in areas with low levels of exploitable resources. Remote sensing also offers the possibility of monitoring the spread of pervasive species, such as lianas.Hyperspectral imagers and LiDAR sensor have been used on board an aircraft to identify species in the Guyanese tropical forests. A wide spectral range from hyperspectral sensors (400-2500 nm) is measured allowing to have many descriptors. LiDAR provides a detailed description of canopy structure and facilitates the segmentation of canopies. The fusion of these two types of information improve the characterization of the resource.In order to make the most of the hyperspectral data, different radiometric preprocessing has been evaluated. Spatial smoothing and shadow filtering are the main factors that improve species discrimination. The full spectral range rather than only the visible-near-infrared region (400-1000nm) is also beneficial. These classification results were obtained on a group of 20 abundant species. The identification of these same species in a mixture within a hyperdiverse stand was the second step of this work.We thus assessed the level of spectral information required and the degree of confusion tolerable in the learning data when the task is to find a target species in a hyperdiverse canopy. A special classification method was implemented in order not to be sensitive to contamination between focal/non-focal classes for training. Even in the case where the non-focal class contains up to 5% of pixels of the focal class (species to be identified), the classifiers developed proved to be efficient.The third step deals with the problem of transposability of the classifiers from one acquisition to another. The characterization of the acquisition conditions and the consideration of their effects are necessary to convert the radiance data into surface reflectance. However, this standardization operation remains an extremely delicate step given the many variability sources to be considered: state of the atmosphere, sun-sensor geometry and illumination conditions. By comparing repeated flights on the same site, we evaluate the contribution of the various acquisition characteristics to the spectral divergence between dates. This work aims to propose ways to develop species recognition methods that are more robust to variations in acquisition characteristics.