Spectral-Spatial Feature Extraction Techniques For Hyperspectral And Multispectral Image Classification

The main objective of this thesis is the classification of VHR multispectral and hyperspectral remote sensing images.The last generation of remote sensing(RS)sensors are able to acquire images with very high spatial and spectral resolution from satellite and airborne platforms.Particularly,this new generation of remote sensing systems acquires VHR multispectral images characterized by a sub-metric spatial resolution and hyperspectral images,characterized by very fine spectral resolution with hundreds of narrow spectral bands.The remote sensing images characterize precisely the different materials on the ground with spatial coordinates(e.g.,streets,buildings,vegetation,water,etc.)in a given study area.This type of data provides very useful information for several applications related to environment,security,urban studies,monitoring...etc.However,for developing real-world applications with multispectral and hyperspectral data,it is necessary to define efficient and effective techniques for image analysis.In this dissertation,we focus our attention on the development of feature extraction and fusion techniques for remote sensing image classification,which become the most important step in classification frameworks of VHR multispectral and hyperspectral data.Image classification is an important task in remote sensing such as it is a principal application of remotely sensed data with a wide range of spatial,temporal,spectral,and radiometric resolutions.Image classification can be considered as the fulcrum of most image interpretation and representation in remote sensing field.The pixel wise-classification is amongst the prominent topic in remote sensing field and it is a challenging problem for two main reasons.First,a large spectral dimensionality of the data.Second,big spatial variability of spectral signature.For this reason,the general objective of this dissertation is the development of novel techniques for the classification of VHR multispectral and hyperspectral images.The proposed approaches in this thesis are devoted to extract and describe useful features that represent the large spectral and spatial information in order to improve classification performances and exploit good classification results in real applications.Three different strategies are considered in the thesis as described below:1)The first strategy is the development of feature extraction techniques for the classification of hyperspectral and VHR-multispectral images,in order to identify useful spectral and spatial features that exhibit high capability to discriminate among the defined classes and high invariance in the spatial domain of a given scene in investigation.This feature extraction strategy is based on two extractors Extended Multi-Attribute Profiles and Sparse Autoencoder such as it is divided in two approaches exposed in two papers.In the first one,we use Extended Multi-Attribute Profiles(EMAP)for extracting spatial features.Then,we use sparse autoencoder(SAE)for feature extraction and dimensionality reduction of EMAP features vector.The obtained EMAP-SAE features are finally employed for performing supervised classification with Support Vector Machine(SVM).Despite the high performance and speed of this approach “EMAP-SAE”,we found that this process loses much of the spectral information.For this reason,we developed the second approach “Spectral-EMAP-SAE” to recover the loss of spectral information and increase the classification performances.In this approach,Extended Multi-Attribute Profiles is employed to spatial feature extraction,then the spatial features are joined to the whole original spectral information of a given remote sensing image to describe its spectral-spatial property.The obtained features are fed into a Sparse Autoencoder for the efficient extraction of the mixed features.Finally,the learned spectral-spatial features are used for classification by using SVM classifier.2)The second strategy “D-SS Frame”(D-SS Frame: Deep-Spectral-Spatial Frame)is based on Feature Extraction and Fusion techniques using Deep Learning methods for Panchromatic and Multispectral Image Classification.In particular,the proposed approach extracts and fuses spectral and spatial information from panchromatic(PAN)and multispectral(MS)images of a same scene using sparse autoencoder and Stacked Sparse Autoencoders.The developed framework is presented in three stages;the first stage is to extract the spatial information from panchromatic image(high-resolution image)by using Sparse Autoencoder.In The second stage,Stacked Sparse Autoencoders(SSAE)is employed for describing the spectral information of multispectral image(low-resolution image).Finally,in the third stage,the spatial and spectral features obtained from panchromatic and multispectral images are concatenated directly as a simple fusion features then used them as the input of SVM classifier.The experiments carried out with MS and PAN images acquired with the same sensor "WorldView-2 satellite".3)Finally,the third strategy “RS-MSSF Frame”(RS-MSSF: Remote Sensing Multiple Spectral-Spatial Features)is a new contribution for spectral-spatial classification of remote sensing images based on extracting spectral features and multiple spatial features(shape and texture).In addition to the deep spectral features extracted using Stacked Sparse Autoencoders,two different types of spatial features are extracted "shape-geometric features" and "texture features".In our approach,Extended Multi-Attribute Profiles is employed for extracting geometric-shape features and Fast-Gray Level Co-occurrence Matrix(FGLCM)is used for extracting texture features.Thus,the result of spectral and multi-spatial feature fusion is fed into SVM classifier.The fusion of these high-level features achieves high classification accuracy comparing to the others strategies presented in this thesis.