Mode de Vie
-
MOdélisation et DEtection de Végétaux
en Interaction avec leur Environnement

Navigation:

Agenda

Participants

Project Ariana, shared by INRIA/CNRS/UNSA. Participants : Xavier Descombes, Mats Eriksson , Guillaume Perrin, Josiane Zerubia.

Project DigiPlante, shared by INRIA/Laboratoire MAS de l'ECP/UMR AMAP du CIRAD. Participants : Philippe de Reffye, Paul Henry Cournède, Véronique Letort, Amélie Mathieu.

The MAS Laboratory (École Centrale Paris). Participant : Christian Saguez

LIAMA (Lab. franco-chinois en Informatique, Automatique et Mathématiques Appliquées). Participants : Remote sensing data fusion and understanding group : Véronique Prinet, Cyril Cassissa, GreenLab : Thierry Fourcaud, Marc Jaeger.

Coordinator of the project: Xavier Descombes , I.N.R.I.A..

Abstract: The subject of this ARC, a collaboration between INRIA, École Centrale Paris, and LIAMA, falls into the frame of the larger theme of study of the development of the biodiversity. In particular, the wish of rationalise the maintenance of the eco system. The project has two ambitions. The first objective is to develop an automatic utility for foresters to participate in image analysis. This utility includes many functionalities. Among them for instance, count the trees, classify the trees into there species, and follow the evolution of the trees. The evaluation of the developed method is based on simulations from the models proposed by the team developing Digiplante. The second objective concerns the modulation of the growth of the population. The models developed by the team behind Digiplante will confront the extracted information from the aerial images.

Very brief summary of the first obtained results (non technical powerpoint presentation)

Publications of the MODEdeVIE Project :

Title / Original	Date	Consult
2D and 3D Vegetation Resource Parameters Assessment using Marked Point Processes, Conference ICPR	August 2006
A comparative study of three methods for identifying individual tree crowns in aerial images covering different types of forests, Conference ISPRS	July 2006
Evaluation des Ressources Forestières à l'aide de Processus Ponctuels Marqués, Conference RFIA	January 2006
Adaptive Simulated Annealing for Energy Minimization Problem in a Marked Point Process Application, EMMCVPR Workshop, © Springer Verlag	November 2005
A Marked Point Process Model for Tree Crown Extraction in Plantations, conference ICIP	September 2005
Image Processing for Forest Monitoring, ERCIM News No. 61	April 2005

Description of the main goals

Introduction

The interest of finding individual tree crowns in digital images has its roots in 1920, which is the first time aerial photo-interpretation was used in forest inventories. Since the breakthrough, after World War II, the technique of using remote sensing in forestry has been frequently used. Today, a common technique is to use a method based on manual interpretation of remotely sensed data (aerial photographs) in combination with field measurements, in order to find the forest parameters used to maintain the forest. Important parameters are stem volume, age, tree species composition, and ecological values such as key biotopes and habitats. These parameters can be estimated by interpretation of aerial photographs using a stereoscope. However, manual interpretation is time consuming and labour. By using automatic techniques a decrease in the work-load might be possible.

Identification of trees in aerial images

To identify the trees in the aerial images point processes are uesd. The point processes takes advantage of both stochatical methods and geometrical approaches. Marked point processes take advantages from both the stochastic and the geometrical approaches. In fact, the stochastic models, especially the Markov chains, have proved their capacity in analysis images by taking account unkown and a priori information of the structure of the wanted objects in the image. Nevertheless, the type of the a priori information in this approach remains often to be pixelwise. If the homogenity criteria are easily modelable, it is most difficult to include also the geometrical information in the solution. Moreover, the attached end of the tasks, calculating the height of the pixel, remains bad for unknown correlation or geomerical type. However, the realisations of the marked point process are composed by a group of points whos distribution is derived from a density relative the Poisson measure. This density permitts to model the mean number of the points such as relations between these points (neighbourhood, alignement, etc.). All of these points are associated marks, realisations of the random variables, definitions of the geometry of the underlying object (radii, length, compactness, etc.) The problem this ARC is dealing with is well adapted to modelisation of a marked point process. In fact, the knowledge of the object is of importance since the goal of the project is to count the trees. The other part, the parameters of the objects allows to characterize as well as recognition of the different trees. We thus defines a geometrical object to model tree crowns. Once the model is defined, its use will require the development of an algorithm of optimization. To solve this problem, we use an algorithm called MCMC with reversible jumps. This algorithm is iterative. With each iteration, a random disturbance of the current configuration is proposed. The new configuration obtained is accepted with a probability which depends on the model, evaluated for the current configuration and the new configuration, as well as law of proposal. The various disturbances suggested consist typically of the addition or the removal (birth or died) of an object in the current configuration or of the modification of the mark or the position of an object of the current configuration. The various disturbances suggested as well as the manner of combining them have a paramount influence on the properties of convergence. Indeed, if convergence is theoretically assured, the speed of convergence remains an open problem. A great effort will thus be devoted to the definition of an algorithm of optimization adapted to the problem and the model considered. We will define an estimator and an algorithm converging towards this estimator to extract the tree crowns starting from a scene. We will then have directly access to the number of trees (a number of objects). The statistics of the configuration obtained will also make it possible to recover invaluable information on forest cover like the density of timbering, the space distribution of the trees or the distribution of their size.

Classification of tree species

Classification will be carried out along two axes. The first axis relates to pixel-wise classification on the level. This classification will make it possible to delimit the zones of vegetable cover in the studied image. With this intention, one will base classification on an analysis of texture, jointly with radiometry. One will distinguish in particular the dense zones of settlements, for which one cannot individualize the tree crowns, of the sparser zones. A possible approach will consist in carrying out a classification on radiometry, then to classify this first result. Indeed, the nondense settlements are characterized by a mixture of classes to the radiometric direction (alternation of tree crowns, shades and ground). The texture of classification on radiometry should thus make it possible to characterize them. This work will be strongly based on the experiment of the Ariana project classification analysis of textures. This module of classification has a double vocation. On the one hand, that will allow a total analysis of the image by distinguishing the zones from interest (zones of settlements) of the other zones (fields, urban zones). In addition, that will enable us to distinguish the zones on which the detection of tree crowns can be done, and to gauge the model of detection (estimate of the orientations and spacings to define the interactions between objects).

Populative change detection

We plan to apply the tools for extraction of tree crowns to a multitemporelle series (possibly restricted at two dates), then to estimate the displacement of the summits in the two iamges. The interest in such a study relates to the bond (direct or indirect) which exists between the displacement of the summit and the movement of the ground itself. The knowledge of displacements on the level of the summit is thus an important indicator for the analyzes movements, landslides. The knowledge of the history of the movement of the summits makes it possible in addition to estimate, via the mechanisms of tropism, the quality of wood when this one is used at industrial ends (Ancelin 2001, Ancelin and Al 2004, Fourcaud and Al 2003). The objective of this work is thus to set up a methodology making it possible to find the local displacement of the summits of the trees, from two or of a sequence of aerial images containing trees in a forest. The problem amounts pairing in a single way each trunk of the first image with its counterpart in the second image: This is a spot of mapping. It will be supposed that the position of the summits and the circumference of the trees in each image are known (stage of detection). It is also supposed that: 1) the displacement is local 2) the variation of the vectors of displacements on the image is relatively continuous almost everywhere. The approach recommended will consist of two stages: 1 - calculation of the field of displacement by optical flood. It should result a first estimate from it from distribution of the field. 2 - regularization of the field by Markovian approach. The introduction of constraints of interdependence and proximity should make it possible to find a smooth variation of the field in the image.

Modulation of the population

The objective is to develop a model the growth of the population. It is a question in particular of extending on a settlement scale the GreenLab model, model of growth structure-function (The fascinating be-with-statement of account at the same time the architecture of the plants and their photosynthetic operation) developed for the plant insulated by Reffye and Al to CIRAD AMAP, with the LIAMA, and from now on in the action of national scale Digiplante (action INRIA, ECP, CIRAD). The plant is described like a dynamic system in interaction with its environment. Recent work allowed a new mathematical formalization in particular, making it possible to carry out studies of behavior and analyses of sensitivity like implementing techniques of optimal control. The passage to the settlement requires to introduce into the system the interactions between plants by modelling the phenomena of competition, in particular for the light and water. Once this modeling carried out, growth of the settlement perhaps simulated and visualized.