Research On Algorithms For Hyperspectral Remote Sensing Image Classification

Hyperspectral images(HSIs)can depict the meticulous spectral characteristics of different land-cover with hundreds of contiguous bands.Classification of HSI,which aims to allocate a possible category for each pixel using the spectral and spatial features,has drawn attention increasingly from the relevant fields.Traditional machine learning methods are limited by their trait of extracting shallow features and poor robustness.Recently,convolutional neural network based models,including residual network,have been the preferable architecture to extract deep spectral-spatial features.However,there are generally some interfering pixels in the neighborhoods of the center pixel,which are unfavorable for feature extraction and will lead to a restraint classification performance.Therefore,attention mechanism is introduced to emphasize the salient bands and crucial spatial positions.But there are still some deficiencies in current attention-based methods.The first one is that the spectral and spatial attentions were employed separately to extract features in spectral and spatial channel,which restricts the supplement of spectral and spatial features.Another one is that current attention modules are weak in emphasizing the benefit of the center pixel to the deduction of spatial attention.This may cause spatial attention to deviate from the relevant spatial areas,which will impair the extraction of the discriminating features greatly and may result in wrong predictions.To solve the first issue,a spectral-spatial fused attention module is proposed.It contains three parts.The first part is designed to extract the correlation among the bands.The second part aims to acquire the common spatial positions.Different from the former two parts,the last part explores the stable spatial features and the contributions of neighborhoods to the center pixel.The identical fused attention modules are stacked sequentially in the proposed network to refine the spectral-spatial feature extraction.For the second issue,two novel models,centralized spatial attention constrained network(CSpaACN)and spectral-similarity-based spatial attention module(S~3AM)are proposed to mitigate it.The latter is the updated version of the former.The CSpaACN model contains a centralized spatial attention module(CSpa AM),a spectral attention module(SpeAM),and a spectral-spatial feature extraction module(SSFEM).Based on the spectral similarity,the CSpa AM is capable of capturing the relevant spatial areas composed of the pixels of the same category as the center pixel from HSI cube with a novel inverted-shifted-scaled sigmoid activation function.The Spe AM aims to select the bands which are beneficial to the spectral features representation.The SSFEM is exploited to extract the discriminating spectral-spatial features.Two well-designed spatial attention masks generated by the CSpa AM are employed to guide the works of the Spe AM and the SSFEM,respectively.Moreover,a spatial consistency loss function is installed to maintain the consistency between the two spatial attention masks so that the network enables the distinction of the relevant features exactly.The novel S~3AM is proposed to emphasize the relevant spatial areas in HSI.It adopts the weighted Euclidean and cosine distances to measure the spectral similarities between the center pixel and its neighborhoods.To alleviate the negative influence of the spectral variability,the full-band convolutional layers are deployed to reweight the bands for the robust spectral similarities.Both kinds of weighted spectral similarities are then fused adaptively to take their relative importance into full account.Finally,a scalable Gaussian activation function,which can suppress the interfering pixels dynamically,is installed to transform the spectral similarities into the appropriate spatial weights.The S~3AM is integrated with the residual network to build the S~3AM-Net model which is able to extract the discriminating spectral-spatial features.Experimental results on multiple HSI data sets demonstrate the effectiveness and the outstanding classification performances of the proposed methods.