Search

International audience; Forest stand delineation is rapidly evolving from traditional photointerpretation to semi-automated GEOBIA techniques. To obtain a good correspondence between image objects and geographic objects, GEOBIA techniques require user decisions considering input data, segmentation algorithms, and classification strategies. GEOBIA applications in forestry have relied mostly on optical remotely sensed data, focusing on the spectral properties of vegetation to identify forest stand boundaries. A limitation of this approach is that optical data have limited sensitivity to forest structural parameters which are the main driver of stand boundaries in even-aged forests. Active sensors such as Light Detection and Ranging (LiDAR) are an alternative, providing a direct estimation of forest structure (e.g. height, density) and potentially leading to more accurate stand maps. In this paper, we propose a semi-automated methodology for even-aged stand delineation using LiDAR data and a two-stage GEOBIA evaluation strategy, combining both unsupervised and supervised evaluation methods to select a suitable segmentation output. The study area is located in the Clear Creek, Selway River & Elk Creek watersheds (~ 54,000 ha) in Northern Idaho (USA), where available LiDAR data was collected in 2009 (Clear Creek watershed) and 2012 (Selway River & Elk Creek). Additionally, a reference dataset of stand-replacing disturbances consisting of yearly clearcut maps compiled from timber harvest records were also available from 1950 as part of the US Forest Service FACTS (Forest ACtivity Tracking System). The proposed methodology involves: (1) image segmentation of several airborne LiDAR metrics using the multiresolution segmentation algorithm implemented on the eCognition software varying consistently the scale, compactness and shape parameters; (2) selection of the best set of parameters for segmentation for each tested LiDAR metric, applying an unsupervised evaluation method based on measures of spatial autocorrelation. This stage ensures that the selected segmentation has the highest possible intra-object uniformity and inter-object heterogeneity; (3) selection of the most suitable LiDAR metric for the segmentation, applying a supervised evaluation method based on measures of area-based dissimilarity, selecting the segmentation with the maximum degree of similarity in size and shape to FACTS reference dataset; and (4) validation using as reference data forest stand perimeters independently derived from visual interpretation. The results show good delineation of even-aged forest, including stands harvested more than 60 years ago that are generally challenging to detect with optical data, because the spectral response of forest canopy saturates at high levels of canopy closure. On a methodological level, the proposed two-stage procedure allows not only accurate image objects delineations but also allows to select the most suitable input data that assure that the image objects are spatially matching with the ground objects. This workflow could be implemented in other studies where different segmentation strategies (e.g., different segmentation algorithms, parameters or resolutions), input data (e.g., Landsat data) or target features (e.g., land cove types) need to be assessed.

We estimated the dominant height of beech regeneration stand relating LiDAR aerial data and field data. The nurnber of sample plots used was 101. LiDAR pulses were mode led using different metrics (percentiles, maximum and minimum, local maximum heights, etc) as exp lana tory variables. The uncertainty in the GPS location of the sample plots and the effect of the variation in the LiDAR retums density were evaluated in order to choose the most robust and stable model. The model using the 95th percentile as the independent variable was selected (RMSEcv= 0.79) and applied to the whole study area getting a mesh of plots with dominant height as the associated predictor variable. Two different mapping methods (since a circular plot mesh and since square plot mesh) are applied into a geographie information system to map the dominant height in the study area. The two methods evaluated are optimal for our work scale.; On a estimé la hauteur dominante de jeunes peuplements de hêtre en croisant l'information LiDAR aérienne et les données de terrain. A partir de 101 placettes de terrain, une série de métrique ont été extraites du nuage de points Lidar (percentiles, valeurs maximales et minimales, hauteurs de maxima locales, etc). Elles ont été utilisées comme les variables explicatives des modèles linéaires testés. L'incertitude du positionnement du GPS dans les placettes de terrain et l'effet de la variation de la densité de points Lidar ont été évalués afin de choisir le modèle le plus robuste et stable possible. Le modèle qui utilise le percentile 95 comme variable indépendante a été choisi (RMSEcv = 0.79) et appliqué à l'ensemble de la zone d'étude obtenant ainsi un maillage de placettes qui ont la hauteur dominante de prédiction comme donnée associée. Deux méthodologies de calcul de carte (maillage de placettes circulaires et carrées) ont été appliquées sous un système d'information géographique afm de cartographier l'ensemble de la zone d'étude. Les deux méthodologies de cartographie testées semblent optimales pour l'échelle de notre travail.