Spectral-spatial based remote sensing image classification

Abstract

Spatial information is important for remote sensing image classification. How to extract spatial information and incorporate it into classification procedure is a challenging issue in order to improve the traditionally spectral based classification. New frameworks and models are developed in this thesis. A new method based on the Conditional Random Fields (CRF) model is constructed to incorporate both spatial and spectral neighbouring information into the classification simultaneously. We develop a simplified version to cope with the complex training procedure for CRF. The new model integrates the boundary constraint into the classification. Comparing to traditional Markov Random Fields (MRF) model, this method incorporates discriminative model instead of generative model and takes into account of the observed data dependency into the classification to improve the result. A new framework is also developed based on the super pixels which are spatially connected homogenous regions. We develop an irregular graphical model based on the super pixels to incorporate long distance dependency into classification. Computation complexity is reduced significantly because of the reduction of the nodes and the edges in the construction of this graphical model. New methods are built to calculate the node and edge potentials with pixel level samples. A new algorithm for boundary information is developed to avoid over-smoothing. The third part of this thesis is the investigation of incorporating the spatial features into hyperspectral remote sensing image classification. These generated extra features make the high dimensional problem even worse for the hyperspectral data. A Sparse Multiclass Logistic Regression model is applied in this thesis to combine the spectral features, spectral interacted features and textural features into one procedure. This model realizes the feature selection and classification simultaneously. It can identify an effective subset of features and cope with a small size of training sample problem, even for the situation where the number of training samples is smaller than the dimensionality of the features. Experiments were conducted using real remote sensing images and encouraging results are acquired comparing to traditional methods.