The rationale underlying the traditional approaches to computer-assisted land cover classification using digital remote sensing data is that pixels from within the same land cover class tend to group together or cluster in multispectral feature space, and that groups of pixels from different cover classes tend to be separate from one another in multispectral feature space. The tendency for pixels from within the same land cover class to form spectrally distinct clusters is the foundation of the algorithm employed in this project for thematic feature extraction and classification. In unsupervised approach, which is a part of the computer-assisted classification of digital multispectral remote-sensing data, and often referred to simply as cluster analysis, a computer algorithm first partitions a multispectral image into self-defining spectral clusters. After the classification is completed, the analyst then employs a posteriori knowledge in labeling the spectral classes into information classes. An unsupervised approach was utilized in this project, specifically the Iterative Self-Organizing Data Analysis (ISODATA) algorithm. This algorithm was used as implemented in the Leica Geosystems ERDAS Imagine 8.7 image processing and pattern recognition software suite.
For 120 cities theLandsat Thematic Mapper (TM) circa 1990 (T1) and Enhanced Thematic Mapper-Plus (ETM) circa 2000 (T2) data, both cloud-free, especially within the area of interest surrounding the cities, and on a date within two years of the respective country’s population census were selected as the basis for image analysis and land cover classification. All Landsat data were orthographically corrected to remove geometric distortions and displacements. Each scene was geo-referenced to the Universal Transverse Mercator (UTM) projection and the WGS-84 datum. Image pixels were re-sampled to 28.5 meters. Each full-scene Landsat was subset to just the area required to cover each city.
The ISODATA clustering algorithm was used to partition the T1 subset scenes into 50 spectrally separable classes. Using the Landsat data themselves, along with independent reference data when available, each of the 50 clusters was placed into one of four pre-defined cover classes: water, urban, vegetation, barren (including bare soil agriculture). Because per-pixel, spectral data-alone classification methods often encounter difficulty in discriminating between urban and barren cover types, the classification maps were carefully scrutinized to detect obvious misclassifications by comparing results with the source image, through a careful, section-by-section examination of the Landsat imagery. On-screen editing of regions of pixels obviously misclassified was performed through heads-up digitizing. The resulting land cover classifications were recoded into two classes: non-urban and urban. All the pixels, classified as urban in the T1 classification, were then extracted from the T2 Landsat image as it was assumed that urban development only increases with time. Then the same ISODATA algorithm along with latter class labeling and on-screen editing was performed to the portions of the T2 Landsat image with results recoded into non-urban and urban classes.