Research On Hyperspectral Image Classification With Convolutional Feature Learning Network

Hyperspectral remote sensing,with wide detection range and strong discrimination ability between different land-cover types in remotely imaged scenes,remains to be one of the key technologies for macro and scientific observation,which plays an important role in scientific research and practical application.Although offering both rich spectral information in hundreds of continuous spectral bands and spatial information,hyperspectral images(HSIs)remain to pose challenges for land-cover classification because of their huge dimensionality and complex non-linearity signatures.The emerging deep learning provides a new way to address the above challenges due to its strong feature learning ability,which has been by far one of the frontier research directions and hot areas.Because of the affection of various noise types and distributions in HSIs,the feature instability of land-covers under the condition of different spatial resolutions,and the spectral-spatial feature heterogeneity in different land-cover types of HSI,current deep learning-based classifiers perform weaker generalization ability to different scenes of HSIs,which greatly limit their practical application range.By combining with the theory of convolution neural network(CNN)and the characteristics of HSIs,we study the algorithms of joint image denoising and classification,robust feature learning and classification,as well as spectral and spatial feature fusion,and provide several novel methods and ideas to improve the ability of noise immunity,classification accuracy and stability to different scenes of HSIs.The main works of this dissertation are summarized as follows:(1)Since the actual noise types and distributions in HSIs are often unknown,and current works perform denoising and classification separately,which easily incur some useful information loss and weaker generalization ability in real noisy images.We introduce here an integrated deep learning framework for joint denoising and classification of hyperspectral imaging(IDCN).In this framework,we design an all-band denoising network that makes use of spectral information,which can be applied to arbitrary HSIs with disparate spectral dimensionality and enables to connect with the classifier in an end-to-end fashion.This denoising network is connected to a spectral classification network,which captures discriminative spectral features and reduces sharply the high spectral dimensionality by processing jointly groups of spectral bands.In view of the fact that existing deep denoisers and classifiers always perform independently and ignore their interplay,formulate a joint learning scheme,which enables denoising and classification to benefit each other.In particular,we define a novel compound loss function to train the proposed framework from scratch.Iteratively,the denoiser outputs directly the denoised result to the classifier for facilitating classification,and the classifier provides class information to the denoiser which helps in preserving semantic details.This way,the denoising and classification processes improve their performance in parallel.Experimental results show that the proposed framework can adapt to both the simulated data with various noise types and two real noisy data due to incorporating the noise reduction process into the classification step and establishing their interplay.Compared with the reference denoisers and spectral classifiers with asynchronous training,our framework indeed improves the classification accuracy and stability,and yields a favourable denoising result,especially in terms of the semantic content.(2)For the feature instability of land-covers under the condition of different spatial resolutions of HSIs,we propose a robust classification framework based on fully group CNNs(FGCNN).To our knowledge,this is the first reported fully group CNN model in general,and we design it in particular for robust classification of HSIs.In the primary feature extraction stage,we introduce multi-kernel depthwise convolution to weight the spatial information at different scales in each band.Based on this concept and group convolutions,we develop a multi-scale spectral feature extraction method to make full use of spectral features,which effectively overcomes the feature instability of land-covers caused by different spatial resolutions.In the subsequent stage,to avoid overusing point convolutions which are known to incur some loss in accuracy,we introduce a discriminative spectral-spatial feature extraction method via combining separable spatial convolution and group convolution,to capture informative spectral-spatial features with relatively few parameters.The final feature fusion stage,is defined as a novel lightweight group feature fusion method that sharply reduces fusion weights so as to mitigate the problem of overfitting,providing reliable features for land-cover classification.Furthermore,we devise an effective approach that is to relate the network hyperparameters to the number of groups,which reduces the number of hyperparameters and the time required for their tuning.Experimental results on three data sets with different spatial resolutions demonstrate that the proposed framework,with the same hyper-parameter settings,yields remarkably stable classification performance between the testing data sets compared to the state-of-the-art,exhibiting robust feature learning ability.(3)Aiming at the problem of spectral-spatial feature heterogeneity in HSIs,we propose a two-region spectral-spatial feature fusion framework(FFTN)for HSI classification.In view of the fact that existing spectral-spatial classifiers often use a single fixed input,which do not make full use of the spectral-spatial information.Instead,we construct two inputs with different sizes and with complementary spectral-spatial information,and we develop a novel two-region feature extraction method based on a two-stream CNN,which extracts spectral,local spatial and global spatial information in parallel,such to adapt well to the spectral-spatial feature heterogeneity in different land-cover types of HSI.Not all the feature maps are equally informative,and it is thus of interest to identify the most important ones and to suppress the others by some proper weighting.We accomplish this by an operation that generalizes the so-called attention mechanism.Moreover,we devise a layer-specific regularization and a smooth normalization fusion scheme to adaptively learn the fusion weights for the spectral-spatial features extracted from the two parallel streams.An important asset of our model is simultaneous training of the feature extraction,fusion and classification processes with the same loss function.Experimental results on five hyperspectral data sets demonstrate that the proposed method is not only very competitive in terms of the accuracy,but also computationally efficient relative to the current state-of-the-art.