Understanding GEO-Information from High Resolution Optical SatellitesAnalysis of Imagery Using Object Oriented...

Understanding GEO-Information from High Resolution Optical SatellitesAnalysis of Imagery Using Object Oriented Techinques

SHARE


B. Krishna Mohan
Centre of Studies in Resources Engineering
Indian Institute of Technology, Bombay
Powai, Mumbai – 400076, India
[email protected]

Vijendra Nayak
Centre of Studies in Resources Engineering
Indian Institute of Technology, Bombay
Powai, Mumbai – 400076, India
[email protected]

James Hogg
School of Geography, University of Leeds
Leeds LS6 9JT, U.K.
[email protected]

ABSTRACT
Segmentation and classification of high-resolution imagery is a challenging problem in view of the fact that there is significant spatial and structural information in the image besides the spectral information. The fine spatial resolution implies that each object is now an aggregation of a number of pixels in close spatial proximity, and image segmentation and classification algorithms are being developed to exploit this important facet of high-resolution images. An approach in this direction is to segment the image into a collection of regions based on various criteria, and then classify the regions using spectral, textural and shape attributes. We describe an approach here wherein the high-resolution images are segmented into homogeneous regions. Properties are computed for regions and the regions are classified using artificial neural network where each region is represented by a feature vector comprising its spatial and spectral attributes. We find that the region based approaches are in general superior to per-pixel classifiers with high resolution images.

1. Introduction
High resolution remotely sensed imagery offer an exciting possibility for feature extraction and spatial modeling. In low-resolution satellite images, the data processing is only based on per-pixel spectral information because the spatial structure of objects is not easily discernible. In contrast, the high resolution of advanced sensors increases the spectral within-field variability – and therefore may decrease the classification accuracy of per-pixel classification. In addition to the textural and contextual measures employed in the pixel-based classification methods, image objects also allow shape characteristics and neighborhood relationships. Their attributes are applicable to all the pixels inside the objects. This method basically includes the three steps. (a) Image segmentation to extract the regions from the pixel information based on homogeneity criteria. (b) Calculation of spectral parameters like mean vector, texture, NDVI and spatial/shape parameters like aspect ratio, convexity, solidity, roundness and orientation for each region. (c) Classification of image using the region feature vectors using suitable classifiers such as artificial neural network.

2. Image Segmentation
Given an image F = {Fi,j | i=1,…,M; j=1,…,N}, segmentation of F into subsets F(i) is defined as follows, according to the definition of Pavlidis (1982):

2.1 Region growing
Region growing algorithms take one or more pixels, called seeds, and grow the regions around them based upon a certain homogeneity criteria. If the adjoining pixels are similar to the seed, they are added to the region. The process continues until all the pixels in the image are assigned to different regions.

Chang and Li [1994] proposed a region-growing framework for image segmentation. This process is guided by regional feature analysis and no parameter tuning or a priori knowledge about the image is required. Mehnert and Jackway [1997] improved the above seeded region-growing algorithm by making it independent of the pixel order of processing and making it more parallel. Basu [1987] developed general sets of semantics for region detection to describe a number of image models using them. The semantic set is established empirically based on simple and intuitive properties of a region. Lu and Xu [1995] proposed a region growing technique for texture segmentation, in which a two dimensional autoregressive model is used for texture representation. In this technique an artificial neural-network is adopted to implement the parameter estimation process for the autoregressive model and to compute the local texture properties of regions during the segmentation process. He and Chen [2000] proposed a resonance algorithm for image segmentation that is not much different from seeded region growing algorithm. It is stressed that the resonance based segmentation algorithm is much more robust to illumination changes than conventional algorithms that work directly on grey scale images. Krishna Mohan et al. (2004) proposed a segmentation technique based on watershed transform to deal with 1-metre resolution images.