New Faster Multiscale Feature Learning Networks For Hyperspectral Image Classification

With the significant development of remote sensing in theory and technology,hyperspectral imaging technology has become more and more mature.Hyperspectral image（HSI）,which consists of hundreds of continuous and narrow spectral bands and contains an enormous amount of spatial contexture information,offers rich and valuable information for subsequent image processing.HSI classification,known as image segmentation in computer vision,which attempts to assign a specific class to each pixel,is one of the most active topics in HSI analyses and has wide applications including global environmental monitoring,precision agriculture,materials analysis and national defense security.However,although convolutional neural networks based on deep learning in recent years have shown excellent image multi-scale feature extraction capabilities in hyperspectral image classification tasks,due to the complexity of hyperspectral images,hyperspectral image classification tasks still face Many challenges:（1）Most of the existing multi-scale methods based on residual networks for HSI classification represent multi-scale characteristics by increasing parallel convolution.This process not only increases the amount of calculation,but also causes feature redundancy to a certain extent.（2）Existing attention mechanism methods do not make full use of the spatial,spectral and channel 3-D features data of HSI.This paper studies the above-mentioned existing problems,and the specific contributions are as follows:（1）In order to solve the problem of high parameter redundancy in the traditional convolution used in the capsule layer of the capsule network.This paper proposes a faster multiscale capsule network with octave convolution（MSOctCaps）for hyperspectral image classification.In the proposed MSOctCaps,this paper design multiple kernels of different sizes with parallel convolution to extract deep multiscale features.To feasibly reduce the redundancy of parameters and achieve high accuracy,the octave convolution is explored in the capsule layer,instead of the traditional convolution,which improves the accuracy of the capsule layer above predicted by the capsule layer below.The comparison experiments with six state-of-the-arts on two challenging contest data sets demonstrate the proposed MSOctCaps is able to produce competitive advantages in terms of both classification accuracy and computational time.（2）In order to solve the problem that there is a large amount of feature redundancy in the extraction of multi-scale features with multiple parallel convolutional layers,and the feature is weak for extracting pixel-level hyperspectral images.This paper proposes a res2 net with spectral-spatial and channel attention（SSCAR2N）for hyperspectral image classification.In the proposed SSCAR2 N,this paper adopts the res2 net block which constructs residual connections in a separate residual block,thereby effectively extracting multi-scale features at the granularity level while ensuring a low amount of calculation and reduce the redundancy of parameters.To further explore the discriminative features,the spectral-spatial and channel attention mechanism block,including a channel attention module and a spectral-spatial attention module,is adopted to optimize the feature maps and can be effectively combined with the res2 net block.The comparison experiments with five state-of-the-arts on three hyperspectral image data sets demonstrate the proposed SSCAR2N is able to produce competitive advantages.