A new wavelet based multi-resolution texture segmentation scheme of remotely sensed images for vegetation extraction

Texture segmentation via wavelet transform traditionally adopts textural features based approach. However, applying this method can lead to oversegmentation problems. To overcome this limitation, we propose a new scheme of texture segmentation. The proposed approach will be applied to remotely sensed images for vegetation extraction. The key idea is that we precede wavelet transform by a preliminary step for increasing the homogeneity of each texture region and removing noise, consisting on smoothing the image by means of anisotropic diffusion technique. The smoothed image is then decomposed into sub-images with different spatial frequencies. Separating the high spatial frequency details corresponding to vegetations, from the other spatial low frequencies, dictates a wavelet reconstruction while neglecting the approximation coefficients. Then, by applying a triangle algorithm based thresholding, we isolate regions corresponding to vegetations. Natural remote sensed images of olive trees parcels with different geometries and different characteristics were considered for texture segmentation. Simulation results prove that trees textures are successfully extracted.