Multiscale Texture And Shape Feature Extraction And Object-Oriented Classification For Very High Resolution Remotely Sensed Imagery

Very high resolution (VHR) remotely sensed imagery, such as QuickBird, IKONOS, SPOT-5, can provide a large amount of information, thus opening up avenues for new remote sensing applications. However, their availability poses challenges to image classification. Due to the complex spatial arrangement and spectral heterogeneity even within the same class, conventional spectral classification methods are grossly inadequate for classification of VHR imagery. Detailed features and small objects can be detected in VHR images and, consequently, the spectral signatures inside an information class become more heterogeneous and different objects become more spectrally similar. The resulting high intraclass and low interclass variabilities lead to a reduction in the statistical separability of the different land-cover classes in the spectral domain. In order to overcome the inadequacy, the textural, strauctural, scale and object-based features should be exploited effectively, in order to complement the spectral feature space.This paper aims to investigate the textural, shape and object-based features from VHR imagery. Furthermore, these spatial features are extended to multiscale approaches. Afterwards, three case studies on urban, agricultural and forest regions were conducted for validation and application for the proposed algorithms in this paper.At first, we analyzed the effects and performance of different classifiers for VHR image interpretation. Accordingly, we proposed to use the neural network classifiers and machine learning approaches. The notable advantages of these classification approaches consist in the adaptive learning rule and the non-parametric characteristic, which is especially efficient for the complex data distribution of VHR imagery. In experiments, the MLP (Multi-Level Perceptron), PNN (Probability Neural Network), SVM (Support Vector Machines), and RVM (Relevance Vector Machine) were employed. The experiments on the HYDICE airborne dataset revealed that the machine learning approaches substantially outperformed the conventional classifiers, such as MLC (Maximum Likelihood Classification), MDM (Minimum Distance to Mean), SAM (Spectral Angular Mapping) etc.However, the experimental results revealed that although the advanced classification techniques could give higher accuracies, their results also showed some confident mis-allocations. These mis-classifications should be improved using further discriminatory variables (e.g. additional wavelengths, textural and structural information). In other words, some additional information is essential for the recognition of spectrally similar classes. Therefore, in this paper, we focused on the texture, shape and object-based feature extraction from the VHR imagery.For the texture information extraction, in this paper, a novel multiscale and multidirectional texture measure based on the NSCT transform was proposed. NSCT transform can be divided into two shift-invariant parts:1) a non-subsampled pyramid structure that satisfies the multiscale property and 2) a directional filter bank that ensures directionality. The advantages of NSCT over the wavelet consist in the representations of image edges and lines, therefore, it is more potential for textural and structural feature extraction from the VHR imagery. The experiments showed the introduction of additional texture information could obviously improve the spectral classification in terms of both visual inspection and statistical accuracies. In the other hand, the proposed NSCT feature gave higher accuracy than the GLCM textures, spatial autocorrelation measure and the feature based on stationary wavelet transform.The shape and structural features are of interest because they are more sensitive to the human vision system and they can provide more discriminative information. We proposed a novel shape feature index, namely pixel shape index (PSI), to describe the shape and contour in a local area surrounding a central pixel. PSI is a pixel based feature, which measures the gray similarity distance in multiple directions. The results of PSI were compared with some spatial features extracted using wavelet transform (WT), gray level co-occurrence matrix (GLCM) in order to test its effectiveness. The experiments demonstrated that PSI was capable of describing the shape features effectively and resulted in more accurate classifications than other methods. Based on the extracted direction-lines histogram, we proposed an extension of PSI algorithm, namely SFS (structural feature set). Some new statistical measures were designed to extract structural features from the direction-lines, such as weighted mean, length-width ratio, and standard deviation, in order to overcome the inadequacy of the PSI. Afterwards, some dimension reduction approaches were employed in order to reduce information redundancy. BPNN, EM-PNN and SVM were used to process the hybrid spectral-structural features after the steps of spatial feature extraction and dimension reduction. The proposed SFS approach was evaluated using two QuickBird datasets and the HYDICE Washington dataset. The results revealed that the new set of reduced spatial features had better performance than PSI.With respect to the object-based analysis, an adaptive mean shift analysis framework was proposed for object extraction and classification of VHR imagery. The basic idea is to apply a mean shift to obtain an object-oriented representation of VHR data and then use support vector machine to interpret the feature set. In order to employ mean shift effectively, two bandwidth selection algorithms were proposed for the mean shift procedure. One is based on the local structure and the other exploits separability analysis. Experiments were conducted on two VHR datasets, the DC Mall HYDICE image and the Purdue campus HYMAP image. We evaluated and compared the proposed approach with the well-known commercial software eCognition (object-based analysis approach) and an effective spectral/spatial classifier for hyperspectral data, namely the derivative of the morphological profile (DMP). Experimental results verified that the proposed mean shift-based analysis system was robust and obviously outperformed other methods.The proposed textural, shape and object-based features are extended to multiscale approaches. We aim to 1) extract multiscale features from VHR urban imagery and 2) integrate the multiscale information using three approaches: vector-stacking (VS) SVM, Fuzzy SVM and multi-classifier voting. The VS SVM concatenates the multilevel features in a single SVM. The fuzzy approach deals with the SVM function values for each scale and then chooses the optimal scale according to the membership value. In experiments, three VHR datasets were used for validation of the presented multiscale fusion schemes:ROSIS Pavia datasets and the HYDICE DC Mall image. Experiments showed that both vector-stacking and fuzzy approaches were able to give comparable or higher results than that one obtained by the optimal scale in terms of accuracies. In most cases, the VS SVM provided higher accuracies than the Fuzzy SVM and multi-classifier voting.At last, we used three case studies to validate and test the multiscale features and the classification methods proposed in this paper. The applications involved 1) the feature analysis and information fusion using aerial photograph and LiDAR data, 2) the mangrove mapping from IKONOS multispectral imagery using multiscale texture information in a study area on the Caribbean coast of Panama, and 3) detailed agriculture mapping using multiscale spectral-textural method based on the PHI airborne data in Xiaqiao, Jiangsu Province.