| Hyperspectral remote sensing data contains a lot of spectral dimension and spatial dimension information of surface features,which can provide a better basis for classification of surface features.However,the full band hyperspectral remote sensing data collected by spaceborne/airborne take up a large amount of storage space,which makes it difficult to process quickly.At the same time,there will be Hughes phenomenon,which leads to the decline of classification accuracy.On the other hand,due to the limitation of load transmission and computer processing speed,it is difficult to process more spectral bands in real time.Therefore,it is necessary to study band selection based on less characteristic spectra after band selection.It is a great significance to classify and recognize hyperspectral targets based on segment data.At the same time,most of the existing classification methods only consider the spectral information of the image,ignoring the important impact of the strong spatial correlation on the classification and prediction results.In view of the above problems,this paper constructs a spatial information extraction model by optimizing the dimension reduction of hyperspectral remote sensing data,and proposes band selection method based on improved particle swarm optimization and spatial-spectral joint classification method.The specific research contents are as follows:1)A band selection method based on improved particle swarm optimization is proposed aiming at the problem of dimension reduction of spaceborne/Airborne Hyperspectral Data.Three groups of real hyperspectral data sets are selected,and the hyperspectral band subset obtained by using the traditional particle swarm,simulated annealing-particle swarm,genetic-particle swarm and the improved particle swarm algorithm proposed in this paper is input into the KNN classifier for simulation verification and analysis.Among them,the band selection method proposed in this paper has the highest classification accuracy in three data sets.In Indian pines data set,the variance representing stability is 0,and the classification accuracy is 86.94%;in Salinas data set,the variance is 5.05×10-6,and the classification accuracy is 90.18%;in long Kou data set,the variance is 0,and the classification accuracy is 96.85%.2)Aiming at the problem of insufficient utilization of spatial information in traditional classification methods,a space spectrum joint classification method is proposed.The spatial information extraction model is constructed by using Euclidean distance and spatial reconstruction value to obtain the reconstruction value of the central pixel in the neighborhood spatial region.Then,combined with the band subset information after dimension reduction,KNN classifier is used for classification simulation verification.Finally,the classification results are filtered by mathematical morphology.The results show that the classification accuracy can reach 91.37%in the Indian Pines dataset,92.10%in the Salinas dataset and 98.56%in the Long Kou dataset.To sum up,in this paper,a band selection method based on improved particle swarm optimization algorithm and a space-spectrum joint classification method are designed.Firstly,the improved particle swarm algorithm is used to reduce the dimension of hyperspectral data.Then,the spatial feature is obtained by using the spatial information extraction model.Finally,the band subset and spatial feature after dimension reduction are used as the input sources of the classifier for classification,and the final results are corrected by morphological filtering.The task of high-precision classification of ground objects with a small number of bands representing the whole band information is completed.The comparison results with other methods verify the superiority of the proposed algorithm,and the corresponding conclusions can provide certain reference value for the research field of hyperspectral image processing. |
PDF Full Text Request