基于局部保留降维卷积神经网络的高光谱图像分类算法

为提高高光谱遥感图像的分类精度,通过局部保留判别式分析与深度卷积神经网络(DCNN)算法,提出了基于局部保留降维卷积神经网络的高光谱图像分类算法。首先,用局部保留判别式分析对高光谱数据降维,再用二维Gabor滤波器对降维后的高光谱数据进行滤波,生成空间隧道信息;其次,用卷积神经网络对原始高光谱数据进行特征提取,生成光谱隧道信息;再次,融合空间隧道信息与光谱隧道信息,形成空间-光谱特征信息,并将其输入到深度卷积神经网络,提取更加有效的特征;最后,采用双重优选分类器对最终提取的特征进行分类。将本文方法与CNN、PCA-SVM、CD-CNN和CNN-PPF等算法在Indian Pines、University of Pavia高光谱遥感数据库上进行性能比较。在Indian Pines、University of Pavia数据库上,本文算法识别的整体精度比传统CNN方法的整体精度分别高3. 81个百分点与6. 62个百分点。实验结果表明,本文算法无论在分类精度还是Kappa系数都优于另外4种算法。

Hyperspectral Image Classification Algorithm Based on Locally Retained Reduced Dimensional Convolution Neural Network

In order to improve the classification accuracy of hyperspectral remote sensing images, a novel hyperspectral image classification algorithm based on local preserving reduced dimensional convolutional neural network (DCNN) was proposed by using local preserving discriminant analysis and deep convolutional neural network (DCNN) algorithm. Firstly, the dimensionality reduction of hyperspectral data was analyzed by local reserved discriminant, and then the spatial tunnel information was filtered by two dimensional Gabor filter. Secondly, the original hyperspectral data were extracted by convolution neural network to generate spectral tunnel information. Thirdly, the spatial tunnel information and spectral tunnel information were integrated to form the air spectrum characteristic information, which was input into deep convolutional neural network to extract more effective features. Finally, the feature of the final extraction was classified by using the dual optimization classifier. The proposed method was compared with CNN, PCA-SVM, CD-CNN and CNN-PPF in the performance of Indian Pines and University of Pavia hyperspectral remote sensing databases. In the database of Indian Pines and University of Pavia, the overall recognition accuracy of the proposed method was 3.81 percentage points and 6.62 percentage points higher than that of the traditional CNN method. Experimental results on two databases showed that the proposed method was superior to the other four methods in both classification accuracy and Kappa coefficient, and it was a better classification method for hyperspectral remote sensing data classification.