Gabor Feature Based Spatial-Spectral Classification Of Hyperspectral Image

Gabor wavelet transform shows excellent feature extraction performance in frequency domain.For the particularity of hyperspectral data that contains both spatial and spectral information,3-d Gabor filtering can simultaneously extract spatial-spectral features.Therefore,3-d Gabor transform has been widely applied as feature extraction method in spatial-spectral classification of hyperspectral image in recent years.However multiple sets of Gabor parameters result in huge gabor feature dimension while covering high resolution,which poses a great challenge to both classification algorithms and computation resource requirements.For the purpose of solving the memory overflowing problem while computing high dimensional gabor features,an incremental residuals least square algorithm has been proposed based on the collaborative representation theory.The coefficients of collaborative representation are updated with the single new-coming gabor magnitude feature block iteratively,and the local residuals are accumulated successfully.Finally,the classification is based on the minimum local residuals criterion after traversing all Gabor parameter settings.We analyze the classification performance of the proposed 3D-Gabor-IRLS algorithm with different percentages of training samples,and verifies the efficiency by comparing 3D-Gabor-IRLS with two methods that classifying with dimension reduced by feature selection and PCA.Combining gabor wavelet with the parameter self-learning ability of CNN,a 3-d Gabor convolution based network(3D-GCNN)has been proposed.It explores the representative ability of extracting abstract features with multi-layer Gabor convolutions,and the parameter scale of 3D-GCNN is light-weighted comparing to the standard CNN,which makes it possible to fit the speciality of insufficient training samples of hyperspectral data.In addition,the Gabor Joint Activation function for nonlinear mapping the real and imagine features extracted by 3-D Gabor come up as the activation function of 3D-GCNN.The classification accuracies of different number of training samples,different input patch-size and different depth of 3d-gcnn are illustrated on two common hyperspectral datasets.In experiments,3D-GCNN has shown a better performance than other CNN-based methods on small datasets.