Heterogeneous Remote Sensing Image Fusion And Change Detection Based On Hierarchical Autoencoder

With the development of remote sensing technology,there are different types of remote sensing data.The processing of heterogeneous remote sensing images has become an urgent problem to be solved.In this paper,heterogeneous remote sensing images are remote sensing images from different sensors,or in different resolutions,or having different spectral bands.Considering the importance of earth resources surveys,environmental monitoring and earth research,it's an urgent need to obtain accurate target information from remote sensing images.Thus,image fusion and change detection has attracted a lot of research interest.Based on deep learning theory,this paper studies the fusion and change detection methods of heterogeneous remote sensing images.The specific work of this paper includes:1.A fusion method of panchromatic and multispectral image based on sparse autoencoder is proposed.Considering the defects of the existing neural network-based models,which use super-resolution methods and cannot use the multi-source information comprehensively,this paper designs a multi-source mapping network based on hierarchical sparse autoencoder.First,the image fusion problem is transformed into a nonlinear mapping problem from multi-source images to target images.Given a sufficient number of hidden units,adding a sparse regularization,we can learn the relationship between the input and the output.The experiment results on the Geoeye-Hobart and Quick Bird satellite data show that the proposed method has a great improvement in the spectral and spatial details,with an improvement of 0.02-0.1 on the index Q4,compared with the traditional methods.2.A change detection method between optical image and synthetic aperture radar(SAR)image is proposed,which is based on multi-feature and sparse autoencoder.Considering the problem that we cannot get the difference map between the optical image and the SAR image directly,the multi-feature of the image are extracted and pre-classified by fuzzy C-means(FCM)clustering.Then,a sparse autoencoder model is established to represent the unchanged regions obtained by the pre-classification.Since the two images are different representations of the same scene,the unchanged pixels of the two images have similarities in the nonlinear feature space formed by the deep network,which changed pixels don't have.So using the reconstruction error can reflect the difference between the changed and unchanged pixels.The experiment results show that this method has improved in the index overall accuracy and the Kappa coefficient compared to the existing methods.3.A heterogeneous image change detection based on convolutional autoencoder is proposed.In order to obtain a more automated and more intelligent change detection method,a convolutional autoencoder is designed to automatically extract the deep features of images.Then,the difference map is calculated using the feature map,and the difference map is classified by FCM.On the one hand,the use of convolutional autoencoder can automatically extract hierarchical features to avoid manual feature selection and reduce the process of multi-feature fusion.On the other hand,convolutional autoencoder directly process two-dimensional image blocks and avoid transforming the blocks into vectors,which damages the spatial structure of the image.From the analysis of numerical evaluation indexes of experiment results,it can be clearly seen that this method has improved the accuracy and Kappa coefficient over the previous method.