Research On Hyperspectral Image Classification And Target Detection Based On Deep Learning

Hyperspectral images are obtained using advanced sensor technology and imaging systems.Since hyperspectral image data has wide application prospects in many fields of social life,the analysis and processing of hyperspectral images has become a research hotspot in the field of remote sensing image research.Hyperspectral image target detection and classification are the two most important branches of hyperspectral image technology.However,in the target detection task of hyperspectral images,the existing methods cannot effectively use the background contour structure information of the input hyperspectral data itself.At the same time,the existing hyperspectral classification algorithms do not take into account the impact of non-neighbor pixel correlation on the classification results,resulting in the final classification accuracy not being guaranteed.TTherefore,this paper puts forward the corresponding solutions to the problems in the process of hyperspectral image target detection and classification.The main work of this paper is as follows:(1)A target detection method for hyperspectral images based on background structure assistance is proposed,which fully uses the background contour structure information of hyperspectral data as an aid to complete the detection task of target objects.First select an existing hyperspectral image target detection method with relatively high efficiency and relatively low algorithm complexity to obtain an initial detection result.The image decomposition method of fusion component analysis and comprehensive sparse representation is used to obtain background contour structure information of hyperspectral image data after dimensionality reduction by principal component analysis.The obtained background contour structure information is combined with the initial detection result obtained before,and the false alarm of the non-target area is filtered to obtain the final target detection result after correction.Through comparison experiments on real data,it can be seen that the false detection rate of the final target detection result obtained by the proposed hyperspectral image target detection method based on background structure assistance is reduced and the accuracy rate is significantly improved.(2)A hyperspectral image classification method based on recurrent neural network is proposed,which integrates the texture and morphological features of hyperspectral image data and introduces the idea ofnon-neighbor pixel features.Extract the texture and morphological features of the dimensionality-reduced hyperspectral image,and then use the multi-feature fusion Stracking integration method to integrate and construct the neighborhood pixel features of the hyperspectral image and construct it by the K nearest neighbor algorithm Out of non-neighborhood pixel features.On the basis of introducing a deep learning recurrent neural network framework,the constructed non-neighborhood features are used as its input to optimize the system parameters.Finally,the final image features generated by the recurrent neural network are input to the softmax layer to complete the final hyperspectral image classification.Non-neighbor pixel features not only retain the excellent properties of neighbor pixel features,but also increase the amount of information contained in the features,thereby effectively improving the classification accuracy of hyperspectral images.Through experiments on public data sets,the classification method is compared with other hyperspectral image methods combined with deep learning framework.The experimental results show that the proposed hyperspectral image classification method based on recurrent neural network The three indicators of total accuracy,average accuracy and Kappa coefficient are the most prominent.