Combining object-based texture measures with a neural network for vegetation mapping in the Everglades from hyperspectral imagery

Abstract

An informative and accurate vegetation map for the Greater Everglades of South Florida is in an urgent need to assist with the Comprehensive Everglades Restoration Plan (CERP), a $10.5-billion mission to restore the south Florida ecosystem in 30+ years. In this study, we examined the capability of fine spatial resolution hyperspectral imagery collected from Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) for vegetation mapping in the Everglades. In order to obtain an efficient and accurate procedure for vegetation discrimination, we developed a neural network classifier first and then combined the object-based texture measures with the classifier to examine the contribution of the spatial information for vegetation mapping. The neural network is capable of modeling the characteristics of multiple spectral and spatial signatures within a class by an internally unsupervised engine and characterizing spectral and spatial differences between classes by an externally supervised system. The designed procedure was tested in a portion of the Everglades. An object-based vegetation map was generated with an overall classification accuracy averaged 94% and a kappa value averaged 0.94 in discriminating 15 classes. The results are significantly better than those obtained from conventional classifiers such as maximum likelihood and spectral angle mapper. The study illustrates that combining object-based texture measures in the neural network classifier can significantly improve the classification. It is concluded that fine spatial resolution hyperspectral data is an effective solution to accurate vegetation mapping in the Everglades which has a rich plant community with a high degree of spatial and spectral heterogeneity.