Research On Collaborative Classification Of Multi-source Remote Sensing Data Based On Hyperspectral Images

In recent years,with the rapid development of earth observation technology,single type of remote sensing data has been unable to satisfy the increased application demand of landcover classification.As result of differences and complementarities between different sensing images,the cooperative classification of multi-source remote sensing images has gradually attracted the attention of scholars.Hyperspectral image plays an important role in the landcover classification depending on its fine spectral resolution.However there are lots of mixed pixels in the low spatial resolution hyperspectral images due to the limitation of the resolution of the satellite sensor,which cause troubles in classification of hyperspectral image.So this thesis takes hyperspectral images as the analysis subject to research the cooperative classification of multi-source remote sensing images.The main work of this thesis includes the following parts.(1)This thesis proposes a fusion method of hyperspectral and multispectral image based on multi-scale guide filter.The method regards the top three principal components of multispectral image after principal component analysis as color guided image,and establishes a multi-scale detail injection model.The missing low-level details information are injected into the original hyperspectral image to achieve the improvement of hyperspectral image spatial resolution.Furthermore,image enhancement are realized on two sets of hyperspectral and multispectral data.(2)Considering the lack of spatial information of hyperspectral images and the lack of label data in reality,this thesis proposes a cooperative classification method of multi-source remote sensing data based on spectral-spatial feature.Through feature dimension reduction and feature extraction for hyperspectral image and multispectral image respectively,the redundant information is removed and the main features are retained.The spectral-spatial feature is taken as the input of the stacked auto-encoder,and the generalization ability of the model on new samples is improved through unsupervised pre-training and supervised fine-tuning.The overfitting problem caused by high-dimensional data is avoided.Experimental results show that the proposed method can effectively improve the classification accuracy of new samples caused by inadequate model training.(3)In order to take advantage of the comprehensive utilization of multi-source remote sensing data,this thesis analyzes the cooperative classification based on decision layer.A cooperative classification method of hyperspectral and multispectral images based on decision correction is proposed.The first principal component of multispectral images is used to correct the initial forecast results of hyperspectral images and some misclassifications are removed.In view of the differences in spatial distribution of different types of ground objects,an adaptive correction strategy is proposed to select the best filter kernel size for each type of land cover.Through contrast experiments with other methods,it is found that with the aid of multispectral images,the adaptive correction method proposed in this thesis can effectively obtain higher accuracies and give a strong support to fine land-cover classification.