A New Method for Object-Based Hyperspectral Image Classification

Abstract

Hyperspectral remote sensing technology has many applications in the fields of land cover classification and examination of their changes. It seems necessary to use both spectral and spatial information in the hyperspectral image classification due to recent developments and the availability of images at higher spatial resolution. In this study, a new approach for object-based classification of hyperspectral images is introduced. In the proposed approach, first nine spatial features, including mean, standard deviation, contrast, homogeneity, correlation, dissimilarity, energy, wavelet transform and Gabor filter, are extracted from the neighboring pixels of the hyperspectral image. Then, the dimensions of the obtained features are reduced using weighted genetic (WG) algorithm. Next, the hierarchical segmentation (HSEG) algorithm is applied to the reduced features. Then, for the objects obtained from segmentation, nine spatial features, area, perimeter, shape index, strength, maximum intensity, minimum intensity, entropy, relation and adjacency, are extracted. Finally, the classification is performed using the multilayer perceptron neural network (MLP) algorithm. The proposed approach was implemented on three hyperspectral images of Indiana Pine, Berlin and Telops. According to the experimental results, the proposed approach is superior to the MLP classification method. This increase in the overall accuracy is about 12% for the Indiana Pine image, about 11% for the Berlin image, and about 8% for the Telops image.