|
|
Publications sur Croissance de Region
Résultat de la recherche dans la liste des publications :
Thèse de Doctorat et Habilitation |
1 - Détection de zones brûlées après un feu de forêt à partir d'une seule image satellitaire SPOT 5 par techniques SVM.
O. Zammit. Thèse de Doctorat, Universite de Nice Sophia Antipolis, septembre 2008.
Mots-clés : Classification, Imagerie satellitaire, Zones brûlées, Feux de foret, Support Vector Machines, Croissance de Region.
Copyright :
@PHDTHESIS{zammit_these_08,
|
| author |
= |
{Zammit, O.}, |
| title |
= |
{Détection de zones brûlées après un feu de forêt à partir d'une seule image satellitaire SPOT 5 par techniques SVM}, |
| year |
= |
{2008}, |
| month |
= |
{septembre}, |
| school |
= |
{Universite de Nice Sophia Antipolis}, |
| url |
= |
{http://tel.archives-ouvertes.fr/tel-00345683/fr/}, |
| keyword |
= |
{Classification, Imagerie satellitaire, Zones brûlées, Feux de foret, Support Vector Machines, Croissance de Region} |
| } |
Résumé :
Cette thèse aborde le problème de cartographie de zones brûlées à partir d'images satellitaires haute résolution. Nos modèles reposent sur le traitement d'une seule image SPOT 5, acquise après le feu afin de détecter automatiquement les zones brûlées.
Le modèle est fondé sur les Séparateurs à Vaste Marge (SVM), une technique de classification supervisée qui a démontré une meilleure précision et une meilleure capacité de généralisation que les algorithmes de classification plus traditionnels. Concernant notre problème de détection, les différentes zones brûlées possèdent des caractéristiques spectrales assez similaires, au contraire des zones non brûlées (végétation, routes, eau, zones urbaines, nuage, ombre...) dont les caractéristiques spectrales varient énormément. Nous proposons donc d'utiliser les One-Class SVM, une technique qui dérive des SVM mais qui n'utilise que des exemples de pixels brûlés pour les phases d'apprentissage et de classification.
Afin de prendre en compte l'information spatiale de l'image, l'algorithme OC-SVM est utilisé comme une technique de croissance de régions, ce qui permet de diminuer les fausses alarmes et d'améliorer les contours des zones brûlées.
De plus, la base d'exemple de pixels brûlés nécessaire à l'apprentissage des techniques SVM est déterminée automatiquement à partir de l'histogramme de l'image.
Finalement, la méthode de classification proposée est testée sur plusieurs images satellitaires afin de valider son efficacité selon le type de végétation et la surface des zones brûlées. Les zones brûlées obtenues sont comparées aux vérités de terrain fournies par le CNES, Infoterra France, le SERTIT, les Services Départementaux d'Incendies et de Secours ou l'Office National des Forêts. |
|
 haut de la page
Article de conférence |
1 - A comparative study of three methods for identifying individual tree crowns in aerial images covering different types of forests.
M. Eriksson et G. Perrin et X. Descombes et J. Zerubia. Dans Proc. International Society for Photogrammetry and Remote Sensing (ISPRS), Marne La Vallee, France, juillet 2006.
Mots-clés : Croissance de Region, Processus ponctuels marques, Champs de Markov, Extraction d'objets, Extraction de Houppiers.
@INPROCEEDINGS{eriksson06a,
|
| author |
= |
{Eriksson, M. and Perrin, G. and Descombes, X. and Zerubia, J.}, |
| title |
= |
{A comparative study of three methods for identifying individual tree crowns in aerial images covering different types of forests}, |
| year |
= |
{2006}, |
| month |
= |
{juillet}, |
| booktitle |
= |
{Proc. International Society for Photogrammetry and Remote Sensing (ISPRS)}, |
| address |
= |
{Marne La Vallee, France}, |
| pdf |
= |
{ftp://ftp-sop.inria.fr/ariana/Articles/2006_eriksson06a.pdf}, |
| keyword |
= |
{Croissance de Region, Processus ponctuels marques, Champs de Markov, Extraction d'objets, Extraction de Houppiers} |
| } |
Abstract :
Most of today's silviculture methods has the goal to optimise the outcome of the forest in stem volume when it is cut. It might also be relevant to save parts of the forest, for instance, to protect a habitat. In order to get a good survey of the forest, remote sensed images are often used. These images are most often manually interpreted in combination with field measurements in order to estimate the forest parameters that are of importance in the decision how to optimally maintain the forest. Among these parameters the most common are stem number, stem volume, and tree species. Interpretation of images are often labour and time consuming. Thus, automatically developed methods for interpretation can lower the work load and speed up the interpretation time.
The interpretation is often done using images captured from a far distance from the ground in order to capture as large area as possible. However, this lower the accuracy of the estimates since it must be done stand wise. Knowledge of where each individual trees in the forest is located together with its size will increase accuracy. It makes it also possible to plan the cutting in detail. With this knowledge in mind, research about finding automatically methods for finding individual tree crowns in aerial images has been a subject for researchers the last decades.
Today's methods are not capable to alone handle all kind of forests. Therefore, comparative studies of different segmentation methods with different types of forests are of importance in order to clarify how much a method is reliable at a certain type of forest. This knowledge can, for instance, be used to build up an expert system which are supposed to be able to find individual tree crowns in any kind of forests. The comparison is done using images covering different types of forests. The types of forests that are included in the study ranges from isolated tree crown where the ground is clearly visible between the crowns to dense forest which is naturally regenerated via planted forest.
In this study we compare three existing segmentation methods for extracting individual tree crowns from aerial images. The first two methods are probabilistic methods which minimises some energy function while the third is a region growing algorithm. The first probabilistic method is based on a Markov Random Field modelling. We define a prior Markov model to segment the image into three classes (background, vegetation and tree centres). The prior model embed a circular shape model of the tree crown with a random radius. The data term allows to well position the tree centres onto the image and to describe the tree shape as fluctuations around the circular template. Besides, some long range interactions models the relations between the trees locations, such as some periodicity in case of plantations.
The second probabilistic method consists in modeling the trees in the forestry images as random configurations of ellipses or ellipsoids, whose points are the positions of the stems and marks their geometric features. The density of this process embeds a regularization term (prior density), which introduces some interactions between the objects, and a data term, which links the objects to the features to be extracted. We estimate the best configuration of an unknown number of objects, from which 2D and 3D vegetation resource parameters can be extracted. To sample this marked point process, we use Monte Carlo dynamics, while the optimization is performed via a Simulated Annealing algorithm, which results in a fully automatic approach. This approach works well on plantations, where there are high spatial relations between the trees, and on isolated trees where 3D parameters can be extracted, but some difficulties remain in dense areas.
The third method, the region growing algorithm, relies as all region growing methods on good seed points, i.e. in this case approximate locations of the tree crowns. From the seed points the segments are grown according to a grey level value of the neighbouring pixels. The larger the value is the sooner it is connected to the neighbouring segment. The segments stops to grow when all pixels belongs to a segment. This method, contrary the others, will have as a result, segments that have captured the actual shape of the tree crown if the forest is not too sparse. If the forest is too sparse such that the ground is visible, there are problems of finding the seed points. In the cases when the forest is sparse, there are difficulties to separate the tree crowns from the ground. Even if the seed points would be located only at the tree crowns the result will contain a lot of errors since all pixels most belong to a segment, i.e. even the ground pixels must be connected to a segment in this case. |
|
haut de la page
Ces pages sont générées par 
|
Rapport d'activité 2011
bibtex
