Research On Hyperspectral Image Classification Based On Deep Feature Fusion

Hyperspectral image is a kind of three-dimensional image which contains both spectral and spatial information.With the continuous in-depth research on remote sensing technology,hyperspectral images have been favored by researchers for their advantages of being able to characterize physical,chemical,material and other characteristics of objects.Therefore,hyperspectral images have been widely used in geological exploration,fine agriculture,environmental monitoring,meteorological prediction,marine remote sensing,urban planning and many other fine fields.Because hyperspectral images have high spectral resolution with abundant spectral bands,and each spectral band corresponds to an image with a specific wavelength,which enables hyperspectral images to accurately detect ground objects and distinguish minor differences between different categories.Therefore,hyperspectral image classification has become an important research field.Hyperspectral image classification algorithm is mainly based on machine learning and deep learning algorithm models.This paper mainly focuses on deep learning.At present,the algorithm using deep learning to extract spectral and spatial features has been relatively mature,but there are still some deficiencies.Specifically,it includes : 1)weak recognition of extracted spectral-spatial features,2)insufficient correlation extraction contained in feature images,3)lack of guidance for feature extracted from images,etc.Therefore,this paper proposes three deep feature fusion algorithms to solve the problem of hyperspectral image classification.1.Considering that the recognition ability of spectral spatial features obtained by many current methods is not strong,this paper proposes a deep CNN-based hyperspectral image classification using discriminative multiple spatial-spectral feature fusion.Firstly,Principal Component Analysis(PCA)is used to reduce the dimension of the original remote sensing image to obtain the representative spectral information and remove the redundant information of the remote sensing image.The correlation between features is obtained through multi-layer feature fusion and finally embedded in the classifier.Three data sets were used to carry out four groups of experiments,and comparative analysis was conducted from different perspectives to verify the performance of the model.2.The inherent limitation of many existing algorithm models is that they cannot extract enough spectral and spatial related information.Therefore,this paper proposes a deep collaborative attention network for hyperspectral image classification by combining 2-D CNN and 3-D CNN.First,2D-CNN and 3D-CNN are used to extract rich spectral spatial features.Then,Non Local Block,lightweight Dense block and deep feature fusion are used to extract the correlation in the feature map,which makes the whole model have better image classification performance and stronger generalization ability.Finally,in the experimental part,the effectiveness of Non Local Block and lightweight Dense＿block was first verified,and then the classification performance of the algorithm model was verified.3.In order to achieve to guide extracted features,this paper proposes a hyperspectral image classification based on deep CNN and guidance block network.This algorithm adopts the pixels in the original image and pixel patch taken from after PCA handling as network input.The body of the network structure are composed of three kinds of dimensions corresponding to the block mapping、multiple guiding block and adaptive block,these modules are used to guide the network to extract the corresponding spectral-spatial features.Therefore,these modules also play an important role in assisting the final classification.By comparing the experiments with the basic deep learning algorithms on four kinds of hyperspectral image datasets,it is verified that this algorithm has certain advantages for hyperspectral image classification.