A Supervised Segmentation Network for Hyper Spectral Image Classification Using Convolution Neural Network

Numerous workshop have concentrated on elaborately designing colorful spectral- spatial networks, where convolutional neural network (CNN) is one of the most popular structures. To explore the spatial information for HSI bracket, pixels with its conterminous pixels are generally directly cropped from hyperspectral data to form HIS cells in CNN- grounded styles. Still, the spatial land- cover distributions of cropped HSI cells are generally complicated. The land- cover marker of a cropped HSI cell cannot simply be determined by its center pixel. In addition, the spatial land cover distribution of a cropped HSI cell is fixed and has lower diversity. For CNN- grounded styles, training with cropped HSI cells will affect in poor conception to the changes of spatial land- cover distributions. In this paper, an end- to- end completely convolutional segmentation network (FCSN) is proposed to contemporaneously identify land- cover markers of all pixels in a HIS cell. First, several trials are conducted to demonstrate that recent CNN- grounded styles show the weak conception capabilities. Second, a fine marker style is proposed to marker all pixels of HSI cells to give detailed spatial land- cover distributions of HSI cells. Third, a HSI cell generation system is proposed to induce generous HSI cells with fine markers to ameliorate the diversity of spatial land- cover distributions. Eventually, a FCSN is proposed to explore spectral- spatial features from finely labeled HSI cells for HSI bracket. Experimental results show that FCSN has the superior conception capability to the changes of spatial land- cover distributions.