Development of an Efficient Contextual Algorithm for Discrimination of Tall Vegetation and Urban for PALSAR Data

Fully polarimetric synthetic aperture radar based land cover classification has been intensively investigated for past several decades, but it is still a challenging task to segregate tall vegetation and urban because scattering mechanism involved for both the classes is not sufficient to get the proper threshold in order to differentiate them. Therefore, there is a need to develop such a technique that has the capability to classify these classes with significantly better accuracy. Textural information of an image is known to be an alternate source of extracting useful information of targets. While dealing with natural targets, such as tall vegetation, characteristic of textural feature, i.e., roughness, may be an important parameter which could identify these targets, since both the classes possess different types of roughness. Henceforth, commonly used texture features, i.e., fractal dimension, lacunarity, Moran's I, entropy, and correlation were critically analyzed and realized that these features are still lacking in the concerned segregation, because generally they are pixel-based. Consequently, neighboring pixels are taken into account and an approach has been developed by considering the randomness response (or manner of distribution of scatterers) based on relative similarity of total backscattering power of neighboring pixels by proposing a similarity entropy feature. An optimized threshold method is also developed by means of the contextual thresholding in order to provide a proper decision boundary between the two classes. The proposed approach is successfully tested and validated on different Phased Array type L-band Synthetic Aperture Radar data with sensitivity of tall vegetation and urban as 0.93 and 0.932, respectively.