Study On Joint Multitask And Sparse Representation For Hyperspectral Image Classification

Hyperspectral images(HSIs),with a wide spectral range and high spectral resolution,provide abundant spatial and spectral information to discriminate similar land-cover materials.It has been widely applied to a variety of fields,such as precision agriculture,mineralogy,atmospheric science and national defense.However,the highdimensional characteristics and the correlation among neighboring bands also bring the challenges of the large volume and redundant information.Facing these challenges,how to give full play to the advantages of hyperspectral images and improve the classification accuracy has become a research hotspot in the field of hyperspectral image classification.To handle the problem of the high correlation among neighboring bands,most of current methods focus on using data dimension reduction techniques,such as feature extraction and band selection.But these methods still have some limitations,mainly reflected in the following two points: 1)The band utilization is low,and the selected spectral quality,which depends on a reasonable rate of reduction and a suitable number of bands,is usually uncertain.2)Lacking reasonable analysis and the use of relevant information,the spectral correlation is omitted as interference information,which resulting in the loss of the original spectral information.Based on the sparse characteristic of hyperspectral images and the correlation between adjacent bands,this paper proposes a joint multi-task learning and sparse representation classification method by combining multi-task learning and sparse representation.The main contributions of this paper are as follows: 1)All band information is used in training,which preserves the original spectral features and improves the band utilization rate.2)The spectral dimension and the spectral redundancy of each sub-dataset are effectively reduced.3)The inter-band correlation of images is fully utilized and the spectral information is further explored.The main contents of this paper are as follows.(1)The band cross-grouping strategy is used to construct several related sub-tasks based on the correlation among the bands of hyperspectral images.This strategy divides the highly correlated adjacent bands into different sub-datasets,thus a number of related sub-datasets are constructed and each sub-dataset corresponds to a sub-task.In the grouping process,the strategy just regroup the original spectral,which retaining the original spectral information and reducing the dimensionality and redundancy of the sub-data sets at the same time.(2)This paper constructs a multi-task shared sparse structure and proposes a multitask joint sparse representation classification algorithm.Based on the sparsity of hyperspectral images,the sparse representation model for each sub-task is established firstly.Secondly,with the correlation between sub-tasks,a shared sparse structure with high robustness is constructed by using ?",$ mixed-norm regularization constraints to integrate all tasks into a whole union.These tasks can share a common sparse pattern,which can extract cross-feature information of different tasks and share common feature and information.Moreover,the model integrates all the original band information,which improves band utilization and provides more discriminant information.Then,the accelerated proximal gradient algorithm is used to optimize the non-convex optimization problem of the model,which obtain the weight vectors of multiple tasks.Finally,the classifier is constructed by using the cumulative residual of multiple tasks.(3)This paper designs several experiments based on two sets of public datasets.And this paper analyze the classification effect of the model and verify the validity of the multi-task joint sparse representation classification.In the experiment,the influence of the correlation among multiple tasks on the model is explained.Six groups of comparative experiments are designed and the results indicates the superiority of multi-task joint sparse representation model.Finally,two key parameters of the algorithm are analyzed,meanwhile the variation trend of the model with the parameter setting is illustrated.