2.1. Study Area
Physiographic boundaries formed by the complex topography play an important role in allocating development resources. Therefore, the valley as a study area in this research is delineated based on the watershed boundaries, which were derived from 20-meter digital elevation point data (
Figure 3).
Figure 3.
Study area–Kathmandu Valley.
Figure 3.
Study area–Kathmandu Valley.
The elevation in the valley ranges from 1,100 to 2,700 meters above the sea level, and forms complex topography within a small geographic area. Half of the study area has slopes of less than 5 degrees, while more than 20% of the land has slopes greater than 20 degrees. Geographically, the valley is situated between 27°31′55″ to 27°48′56″ North latitude and 85°11′11″ to 85°31′52″ East longitude. The valley is drained by the Bagmati river system. The river system is the main source of water for drinking and irrigation in the valley [
3,
23]. Politically, the valley is composed of five municipal urban centers (Kathmandu, Lalitpur, Bhaktapur, Kirtipur, and Madhyapur Thimi), in addition to 97 surrounding villages. The study area covers 684 km
2, and the urban centers make up only 14% of the land.
2.3. Mapping of Spatial Patterns
The satellite data are in image form and contain many details, but not in an objective thematic setting. Image analysis techniques are evolving rapidly, but many operational and applied remote sensing analyses still require extracting discrete thematic land surface information from satellite imagery using classification-based techniques [
7,
25]. Prenzel and Treitz [
26] and Thapa and Murayama [
15] argued that the heterogeneity and complexity of the landscape in urban regions, for example, suburban residential areas forming a complex mosaic of trees, lawns, roofs, concrete, and asphalt roadways, require land use and land cover classification techniques that combine more than one classification procedure to improve remote sensing-based mapping accuracies. Therefore, a series of processing steps (
Figure 4) is followed to transform those data into meaningful thematic information.
Figure 4.
Land use mapping scenario for remote sensing images.
Figure 4.
Land use mapping scenario for remote sensing images.
The geometric rectification process was carried out for all satellite images using a road network map in the local projection system (i.e., UTM WGS 1984). Image enhancement, contrast stretching, and false color composites were created to improve the visual interpretability of the image by increasing the apparent distinctions between the features. Knowledge-based visual interpretation, texture, and association analysis were performed at the preliminary stage. Furthermore, field survey data, aerial photographs, high resolution satellite images, and city planning documents were carefully analyzed while preparing the land use classes. After analyzing all the information collected so far, only twelve types of land uses were considered for mapping, i.e., agricultural areas, forest, shrubs, open space, water, built-up areas, industrial areas, roads, airport, institutional areas, government secretariat area, and royal palace. The last six land uses cover small spaces in the valley; therefore, these land uses were merged into an urban/built-up area category for detailed quantitative assessment purposes. However, all twelve legend units were listed in the land use maps.
CORONA image was resampled using a nearest neighbor resampling technique to match the data resolution of LANDSAT (30-meter) for maintaining spatial resolution consistency in the data sources. Resampling the CORONA image (1-meter) into 30-meter may introduce modifiable areal unit problem (MAUP) due to generalization of continuous geographical phenomenon as discussed by Openshaw and Alvanides [
27]. When values are averaged over the process of aggregation, variability in the CORONA image is lost which may result the variation in spatial patterns arising from the statistics computed at coarser resolution. This problem is also known as scale effect in MAUP. However, after following the steps shown in
Figure 4, such problem will be removed in resulting map.
An unsupervised approach with the ISODATA clustering technique [
28] available in Erdas Imagine 9.0 was applied to obtain different land use clusters of similar spectral pixels in the Corona, MSS and TM images. This preliminary interpretation reduced the artificial errors and selected the most appropriate clusters for further processing. Then, the supervised approach with the maximum likelihood parameter was run to improve the accuracy of the land use classification for the images for all three dates (1967, 1976, 1989, and 1999). Aggregating the detailed remotely sensed surface characteristics into thematic information always contains some degree of errors, hence, an accuracy assessment should be performed [
14,
15]. The image classification accuracy was performed by evaluating the overall classification accuracy using geographically referenced vector datasets. The classification accuracies of 83.66, 80.66%, 84.44%, and 83.33% were achieved for the years 1967, 1976, 1989, and 1999, respectively.
Because of the complex topography in the valley, land use types are closely related to altitude and slope and have some specific distribution rules. Confusing areas were detected mostly between the water areas and shadows of mountain areas; bare lands, brick factories, and construction sites; and golf courses and shrub lands. The areas of confusion were further verified with DEMs and slope data, road data, the 1978 land cover map, high resolution imagery, including CORONA (1967), SPIN (1991), IKONOS (2000), aerial photographs acquired at different time periods, and fieldwork information to determine the appropriate land use type. Editing and digitizing were carried out to resolve all the confusion and conflicts that occurred in each map. This process helped to improve the accuracy of the mapping. After updating the maps, the reference year of each map was fixed to 1967, 1978, 1991, and 2000, respectively.
2.4. Analysis of Spatial Patterns
The land use transition matrix is a useful tool that has been widely accepted in land use change analysis [
11,
16]. Three land use transition map layers for the years 1967–1978, 1978–1991, and 1991–2000 were prepared for the detailed land use change pattern analysis in the valley.
Empirical studies have substantiated the use of both spatial metrics and remote sensing in urban modeling [
10,
11]. The use of spatial metrics has provided a new platform for describing the spatial land use and land cover heterogeneity and morphological characteristics within the urban environment. Spatial metrics are already commonly used to quantify the shape and pattern of landscapes [
11,
29,
30]. Recently, there has been an increasing interest in applying spatial metric techniques in an urban environment to link land use heterogeneity to structures and dynamic changes in urban land uses [
10].
Table 2.
Description of spatial metrics used in this study (Compiled from [
30]).
Table 2.
Description of spatial metrics used in this study (Compiled from [30]).
Metrics | Description | Units | Measure of |
---|
PD | PD equals the number of patches of a specific land cover class divided by total landscape area. | No./ 100 ha. | fragmentation |
ED | The sum of the lengths of all edge segments involving a specific class, divided by the total landscape area multiplied by 10000. | Meters/ ha | fragmentation |
LPI | The area of largest patch of the corresponding class divided by total area covered by that class, multiplied by 100. | Percent | dominance |
ENNMN | The distance mean value of all patches of a land use to the nearest neighbor patch of the land use based on shortest edge-to-edge distance from cell centre to cell centre. | Meters | isolation/ proximity |
AWMPFD | It describes the complexity and fragmentation of a patch by a perimeter-area ratio. Lower values indicate compact form of a patch. If the patches are more complex and fragmented, the perimeter increases representing higher values. | None, range: 1-2 | fragmentation and complexity |
COHESION | Approaches 0 as the portion of the landscape comprised of the focal class decreases and becomes increasingly subdivided and less physically connected. | None, range: 0-100 | physical connectedness |
CONTAG* | Contagion index describes the fragmentation of a landscape by the random and conditional probabilities that a pixel of patch class is adjacent to another patch class. It measures to what extent landscapes are aggregated or clumped. | None, range: 1-100 | fragmentation and the degree of aggregation |
SHDI* | Shannon’s diversity index quantifies the diversity of the landscape based on two components: the number of different patch types and the proportional area distribution among patch types. | Information | patch diversity |
A set of spatial metrics were selected to measure and monitor the landscape fragmentation, land use complexity, proximity, dominancy, and diversity (
Table 2). The selected metrics are patch density (PD), largest patch index (LPI), edge density (ED), area weighted mean patch fractal dimension (AWMPFD), Euclidian nearest neighbor distance mean (ENNMN), cohesion (COHESION), contagion (CONTAG), and Shannon’s diversity index (SHDI). These metrics describe the composition and configuration of landscape pattern changes in the valley.
The metrics were computed for each land use map at the class and landscape levels. Metrics at the class level are helpful for understanding landscape development, while those at the landscape level provide relatively general information on the assessment [
31]. All these metrics were calculated using the FRAGSTAT software [
30] while Erdas Imagine and ArcGIS software were used for image analysis and GIS data processing.