Skip to Main content Skip to Navigation

Fusion of 3D point clouds and hyperspectral data for the extraction of geometric and radiometric features of trees.

Abstract : Mountain forests provide environmental ecosystem services (EES) to communities: supplying of recreational landscapes, protection against natural hazards, supporting biodiversity conservation, among others. The preservation of these EES through space and time requires a good characterization of the resources. Especially in mountains, stands are very heterogeneous and timber harvesting is economically possible thanks to trees of higher value. This is why we want to be able to map each tree and estimate its characteristics, including quality, which is related to its shape and growth conditions. Field inventories are not able to provide a wall to wall cover of detailed tree-level information on a large scale. On the other hand, remote sensing tools seem to be a promising technology because of the time efficient and the affordable costs for studying forest areas. LiDAR data provide detailed information from the vertical distribution and location of the trees, but it is limited for mapping species. Hyperspectral data are associated to absorption features in the canopy reflectance spectrum, but is not effective for characterizing tree geometry. Hyperspectral and LiDAR systems provide independent and complementary data that are relevant for the assessment of biophysical and biochemical attributes of forest areas. This PhD thesis deals with the fusion of LiDAR and hyperspectral data to characterize individual forest trees. The leading idea is to improve methods to derive forest information at tree-level by extracting geometric and radiometric features. The contributions of this research work relies on: i) an updated review of data fusion methods of LiDAR and hyperspectral data for forest monitoring, ii) an improved 3D segmentation algorithm for delineating individual tree crowns based on Adaptive Mean Shift (AMS3D) and an ellipsoid crown shape model, iii) a criterion for feature selection based on random forests score, $5$-fold cross validation and a cumulative error function for forest tree species classification. The two main methods used to derive forest information at tree level are tested with remote sensing data acquired in the French Alps.
Complete list of metadata
Contributor : Abes Star :  Contact
Submitted on : Thursday, April 29, 2021 - 3:58:11 PM
Last modification on : Friday, April 30, 2021 - 3:52:22 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03212453, version 1


Eduardo Alejandro Tusa Jumbo. Fusion of 3D point clouds and hyperspectral data for the extraction of geometric and radiometric features of trees.. Signal and Image processing. Université Grenoble Alpes [2020-..], 2020. English. ⟨NNT : 2020GRALT072⟩. ⟨tel-03212453⟩



Record views


Files downloads