Monitoring and Evaluation of Flooded Areas Based on Fused Texture Descriptors

摘要

The paper presents a new method based on combining textural and color information for patch classification to segment flooded areas from aerial images. To this end, the paper presents a method which combines information provided by various important descriptors of texture like Local Binary Patterns, Histogram of Oriented Gradients, and Fractal Dimension in color version. The remote images were taken by the aid of an Unmanned Aircraft System (UAV) designed and implemented by an authors' team. The algorithm of remote image segmentation uses non-overlapped patches of dimension 128 × 128 pixels. It has two phases: (a) the learning phase to create the representatives of the flood class and (b) the segmentation phase, based on patch classification, to estimate the flood size. The classification is made by a voting criterion which takes into consideration the weights calculating from the three descriptors (fractal dimension, LBP and HOG). The accuracy of segmentation, evaluated from 100 real images, was better than in the separate approaches.