Hyperspectral imaging system (HSI) uniquely captures a full spectrum of the reflected radiance of any object in the spatial domain (real world), where each substance exhibits different spectral signatures that combine quantitative and qualitative information. HSI is becoming an overpowering technology for accurate image classification and recognition, for that end, it is pervading many, and increasing, fields of application. However, the high dimension of the data and the shortage of labeled training samples are two majors hindrance to more amelioration of the performance. In this paper, a novel Cross Spatial-Spectral Convolution Network (CSSCN) framework based on the convolutional neural network (CNN) with GoogleNet and principal component analysis (PCA) is proposed. By transforming each pixel into a new spectral channel contains all the spectral signature, the maximum spectral features are exploited, and a concatenated convolutional neural network with a dynamic learning rate based on GoogleNet architecture is employed to extract deep spatial features. We thoroughly evaluate the effectiveness of our method on several commonly used HSI benchmark data sets. Promising results have been achieved when comparing the proposed CSSCN with the state of the art of HSI classification.
Paper
Full text
Cross Spectral-Spatial Convolutional Network for Hyperspectral Image Classification
Semantic Scholar · Environmental Science · 2019
Abstract
Hyperspectral imaging system (HSI) uniquely captures a full spectrum of the reflected radiance of any object in the spatial domain (real world), where each substance exhibits different spectral signatures that combine quantitative and qualitative information. HSI is becoming an overpowering technology for accurate image classification and recognition, for that end, it is pervading many, and increasing, fields of application. However, the high dimension of the data and the shortage of labeled training samples are two majors hindrance to more amelioration of the performance. In this paper, a novel Cross Spatial-Spectral Convolution Network (CSSCN) framework based on the convolutional neural network (CNN) with GoogleNet and principal component analysis (PCA) is proposed. By transforming each pixel into a new spectral channel contains all the spectral signature, the maximum spectral features are exploited, and a concatenated convolutional neural network with a dynamic learning rate based on GoogleNet architecture is employed to extract deep spatial features. We thoroughly evaluate the effectiveness of our method on several commonly used HSI benchmark data sets. Promising results have been achieved when comparing the proposed CSSCN with the state of the art of HSI classification.