Team Ariana

Scientific Foundations
Application Domains
New Results
Contracts and Grants with Industry
Other Grants and Activities

Section: New Results

ARC Mode-de-Vie

Tree crown extraction using marked point processes

Keywords : forest, tree crown, automatic feature extraction, marked point process, RJMCMC, simulated annealing.

Participants : Guillaume Perrin, Xavier Descombes, Josiane Zerubia.

This PhD was funded by a MAS Laboratory grant, with a complement provided by INRIA. It was performed as part of the ARC Mode De Vie [ ].

This work addresses the problem of tree crown extraction from Colour InfraRed (CIR) aerial images of forests. Our models are based on object processes, i.e. marked point processes [25] , [26] , [35] . These mathematical objects are random variables whose realizations are configurations of geometrical shapes. This approach yields an energy minimization problem, where the energy is composed of a regularization term (prior density), which introduces some constraints on the objects and their interactions, and a data term, which links the objects to the features to be extracted. Once the reference object has been chosen, we sample the process and extract the best configuration of objects with respect to the energy, using a Reversible Jump Markov Chain Monte Carlo (RJMCMC) algorithm embedded in a simulated annealing scheme.

New results have been obtained this year by proposing new models to deal with different densities of stand. In dense areas, we use an ellipse process, while in sparse vegetation an ellipsoid process is used. As a result, we obtain the number of stems, their position, the diameters of the crowns (see figure 10 ) and the heights of the trees for sparse areas and coppice-with-standards structures (see figure 11 ). The resulting algorithms have been tested on high resolution CIR aerial images provided by the French National Forest Inventory (IFN) [17] , [27] .

Figure 10. Top left : poplar plantation (©IFN). Top right : tree crown extraction. Bottom left : timber forest (©IFN). Bottom right : position of the extracted tree crowns.
Figure 11. Top left and middle : tree crown extraction on the border of a poplar plantation. Top right : timber extraction on a coppice-with-standards stand. Bottom left : sparse vegetation (©IFN). Bottom middle and right : extraction of the tree crowns in sparse vegetation.

Shape recognition for classification of tree species

Keywords : tree crown, planar shape, shape space, angle function, geodesic path.

Participants : Maria S. Kulikova, Xavier Descombes, Josiane Zerubia.

This internship was funded by ARC Mode de Vie [ ] and it was done partly in collaboration with the Professor A. Srivastava from Florida State University, USA.

The objective of this work is the classification of trees according to their species based on the contour of the tree crowns extracted from very high resolution aerial images (see figure 12 ).

Figure 12. Tree species classification by shape

The theoretical part of the work is based on the theory of planar shape analysis using the notion of geodesic path on shape space developed by Klassen et al. (see figure 13 ).

Figure 13. Examples of geodesics

The application of this theory, particularly the use of the geodesic as a metric for classification, did not give the desired results, so a new approach was undertaken: a set of features was developed and used in learning algorithms. Figure 14 shows the results obtained.

Figure 14. Classification performance. Left: Nearest Neighbour algorithm, right: SVM.

Calibration of high resolution aerial images of trees using plant growth models

Keywords : image calibration, parameter extraction.

Participants : Meena Mani, Xavier Descombes.

This internship was funded by the ARC Mode de Vie project [ ] and conducted in collaboration with P.-H. Cournède (MAS Laboratory, ECP and Digiplante project, INRIA).

In this work, we seek to establish a link between parameters extracted from aerial images of forest stands and predictions from a generic plant growth model. The plant model used here was developed by Greenlab, through a collaboration between Paul-Henry Cournède, Phillipe de Reffye, and LIAMA. Two parameters that can be extracted from an image, tree crown size and the spatial density of trees, were used for this calibration. The spatial density served as an input to the plant growth simulation. The model then computed the tree crown surface area for a given stage (chronological age) in the growth process. The calibration was done by verifying that the surface area predicted by the model matched the average size of tree crowns in the images used. In this case, the match was nearly exact: for a 40-year old plantation, the model predicted a surface area of 12.5 m 2 while the tree crown distribution in the image yielded an average of 12.4±4.38 m 2 (see figure 15 ).

Figure 15. Left: image tree crown distribution, mean = 12.4 m 2 , std. dev = 4.38 m 2 ; right: model prediction, surface area = 12.5 m 2 , age = 40 years.

Texture analysis of tree crowns

Keywords : forest, co-occurrence matrix, parameter extraction.

Participants : Meena Mani, Xavier Descombes.

This internship was funded by the ARC Mode de Vie project [ ].

In this work, we analyze colour infrared (CIR) images for features that could be incorporated into a classifier. These images, obtained from a Swedish site, had been used by M. Eriksson in his thesis [39] . As Eriksson pointed out, the four classes of trees present—aspen, birch, spruce and pine—could easily be identified as deciduous and coniferous from preliminary summary statistics (mean, standard deviation). This is because deciduous trees reflect a substantially greater percentage of infrared light (figure 16 ).

To further distinguish within the deciduous and coniferous classes, we performed texture analysis using co-occurrence matrices. A co-occurrence matrix is a two-dimensional quantitative representation of spatial relationship. We generated nine such matrices for each tree, each matrix representing a different direction. We were able to separate the trees into four classes by plotting co-occurrence energies versus contrast. With these two texture features, the two CIR histogram statistics, and shape parameters, we were able to identify parameters to build a successful classifier (see figure 16 ).

Figure 16. Top left: infrared profile for a deciduous tree (aspen), mean = 203.8 m 2 , std. dev = 53.8 m 2 ; top right: infrared profile for a coniferous tree (pine), surface area = 152.8 m 2 , age = 39.3 years; bottom left: co-occurrence energy vs contrast; bottom right: standard deviation separates the coniferous and deciduous trees.

Segmentation and classification of forest stands

Keywords : forest stand, image segmentation, classification.

Participants : Mats Eriksson, Xavier Descombes, Josiane Zerubia.

This post-doc is part of the ARC Mode de Vie project [ ] and has been partly funded by an MESR grant.

Before applying an individual tree crown delineation algorithm to an image covering a large region it is necessary to divide the image into forest and non-forest. Furthermore, an algorithm produces different results depending on the denseness of the forest. Thus, it is also important to divide forested areas into different parts according to the crown closure. In this work, the segmentation of the forest is performed by a region growing algorithm based on radiometrical information together with a merging step [17] . The classification of the regions is done using a k-means classifier. Segmentation and classification results on an aerial image can be seen in figure 17 .

Figure 17. Left: Original image (©IFN); middle: segmentation result; right: classification result.

Tree detection from aerial images

Keywords : tree detection, forest stand.

Participants : Sotiris Raptis, Xavier Descombes, Josiane Zerubia.

Sotiris Raptis is funded by a contract with the IFN.

The goal of this work is to improve and transfer to the IFN to a software dedicated to tree crown extraction from aerial images. An interface adapted to the final user will be developed. A first step consists in computing a mask which eliminates areas of the image which do not contain any trees. The goal of this mask is twofold. First, it will improve the computational time required for tree detection. Areas such as lakes and urban areas will automatically be removed from the study area, which will reduce the volume of study. The second goal is to eliminate false alarms in the detection process. Preliminary results have been obtained using a k-means algorithm based on a local variance (see figure 18 ).

Figure 18. Left: Original image (©IFN); Right : mask of planted areas.


Logo Inria