Research On Multiple Spectrum Remote Sensing Images Feature Extraction And Synergic Interpretation

With the development of remote sensing earth observation technology, multiplespectrum remote sensing data are increasingly used in the field of military target detection,urban vegetation monitoring, agricultural remote sensing, and so on. For instance,hyperspectral data has almost continuous information of spectrum, high spatial resolutionremote sensing images can provide more detailed spatial information such as shape andtexture, infrared data can reflect thermal radiation and Lidar can build digital elevationmodels (DEM). Application of these multi-platform data, however, are facing with theproblems of amount of data being greater and information being more complex. Therefore,processing for both the high spectral resolution and the high spatial resolution remotesensing images need further study in information extraction and interpretation desperately.This paper aims at synergic processing of hyperspectral and high spatial resolution remotesensing images. Synergic processing of multiple spectrum is implement with the unionfeature.First of all, for the hyperspectral data, characteristic of the data type is analysed. Dueto its nonlinearity and sparsity, the theory of restricted Boltzmann Machine (RBM) andthe learning algorithm of deep belief networks (DBN) framework are introduced. Theclassical feature extraction methods such as principal components analysis (PCA) andnegative matrix factorization are also performed as comparison in hyperspectral datainterpretation accuracy and time cost. This lies research foundation for subsequentinterpretation. It is verified that deep belief networks has an excellent performance forcomplex spectral information.Secondly, with the improvement of the resolution of panchromatic images and thetexture information being more and more detailed and complex, there is a significant dropin the classification accuracy instead. In order to solve this problem, we perform multi-resolution segmentation at the optimal segmentation scale of different landscape wecalculate previously. Both sample-based standard nearest neighbor classifier and multi-level architecture based object-oriented information extraction method are adopted.Support Vector Machine (SVM) classifier based on pixel level is taken as comparison.Finally, merge the adjacent objects which belong to the same class in the informationextraction result of high spatial resolution data and extract spatial features of the merged objects as the spatial features of each pixel in the object. Fuse the spatial features with thespectral information and take the spectral-spatial joint feature as the input of the deepbelief networks. Influence of learning rate, hidden unit number, layer number and ratio oftraining samples on the interpretation accuracy is discussed. In this way, high precisionsynergic interpretation for multiple spectrum image in the feature level is achievedmaking full use of advantages of both datasets.