Texture-based forest segmentation in satellite images

Abstract

NDVI images are often used to segment vegetation areas in satellite images. Unfortunately, NDVI lack information to effectively separate forest regions from grass regions, as it is computed basing solely on spectral characteristics of red and infrared bands, and both forest and grass display similar spectral characteristics. On the other hand, forest areas are visually distinguishable, as they contain a lot of small shadows or dark spots in satellite images, while grass areas look flat. In the paper, we use this observation by extracting simple texture features from the panchromatic band. In our case, it was enough to compute standard deviation for a sliding window to separate forest from grass.