Kernel-based granulometry of textural pattern measures on satellite imageries for fine-grain sparse woodlands mapping

Abstract Textural measures of target objects have been studied to increase the accuracy of mapping sparse woodlands on satellite images. Kernel-based granulometry analyses were used to calculate BCM, NDC, RR, FI, H, and DI on QuickBird, RapidEye, and ASTER images. The Jeffrey-Matusita distance was used for analyzing the separability of the training ROIs. Finally, supervised segmentation was accomplished using the probabilistic maximum-likelihood rule on the measures. The results indicated that BCM and NDC using 5 × 5 and 7 × 7 kernels are more successful for QuickBird and RapidEye images, considering the local co-occurrence of the openings. On ASTER images, the results were considerable because of the convolution effects, defining the granules. Contextual information of the images increased 20%–30% to the accuracy heuristics of overall accuracy and kappa coefficient, 98% for the ASTER data set. The method is easy to retrieve and is helpful for the segmentation of the scene of fine-grain textural patterns.