Mangrove Species Discrimination from Very High Resolution Imagery Using Gaussian Markov Random Field Model

Abstract

© 2018, Society of Wetland Scientists. Mangrove forests are the most productive ecosystem for tropical and subtropical flora, but they are experiencing serious loss. Urgent measures should be taken for conservation and restoration, including investigation of mangrove species constitution using satellite imagery in which texture provides effective information. However, differentiation by texture for various species of mangroves is underexplored. The Gray Level Co-occurrence Matrix (GLCM) has proved to be a common approach to describing texture in most previous studies, while other textural measurements such as the Gaussian Markov Random Field (GMRF) have seldom been tested and applied. This study aimed to provide a comprehensive assessment and comparison of textures using GLCM and GMRF to offer a better understanding of their roles in mangrove species discrimination using very high-resolution satellite data. The experiments were designed to highlight two aspects of textural features - local pattern and textured region size - to assess their effect on mangrove species discrimination based on texture definition. The results indicated that adjusting the textured region size can easily improve texture extraction, and GMRF outperformed GLCM in texture representation of mangrove forest using the same parameters, with an increase in overall accuracy (1.54%, 6.47% and 10.66% on average at three different local pattern sizes).