Segmentation d'images à très haute résolution spatiale basée sur l'analyse multifractale
The recent availability, in remote sensing, of very high spatial resolution images brings the texture classification of images to a higher level of complexity.The singularity content of very high spatial resolution images, such as those from IKONOS, is very important due to the high local variability of their gray level. Such images have so many details that classical classification algorithms fail to achieve good results. In the case of IKONOS images of forest areas, a texture can be so different within a same class, that it becomes very difficult, even for a human, to classify or interpret them.The study of the high frequency content of the data seems to be a good way to study those images.The multifractal analysis provides us with global and local information on the singularities which represent the high frequency content of the image. We propose a new approach which uses the singularities of the image to achieve the classification of very high spatial resolution optical forestry images. It is based on the computation of the Hölder regularity exponent at each point in the image. From this parameter we can compute the local Lacunarity spectrum or the large deviation multifractal spectrum which give information about the geometric distribution of the singularities in the image. So we use global and local descriptors of the regularity of the signal as input parameters to a k-means algorithm. The two algorithms are described and applied to an IKONOS image of forestry as well as to two artificial images, one made of Brodatz textures and the other of fractional brownian motions.The classification results are compared to those obtained with the gabor filters, the laws filters, the fractal dimension, the gaussian Markov random fields and the Haralick co-occurrence parameters.The proposed methods give good results and are even able to segment the image in tree density classes. We also devised tests to see the resistance of the discrimination power of the proposed parameters with regards to intensity fields transformations like contrast stretching, luminosity changes and noise addition. Both methods seem not too, sensitive to such alterations. A study of the sensitivity of the proposed algorithms te, the parameters which govern them was carried out to determine the influence of each parameter on the classification result. A rough estimation of the interval of values which they can take was also achieved. The obtained results show that the multifractal theory is a suitable tool for texture classification and analysis.