The objective of this study was to identify the potential benefit and limitations of using very high resolution satellite images IKONOS in producing land cover/use maps at large scale (1:5,000). For that, a visual interpretation was performed on pan-sharpen IKONOS imageries with 1 m spatial resolution (October 2000) covering representative areas in Lebanon situated on the coast and at the mountains. A land cover/use legend was developed according to FAO classification including 72 classes divided into four hierarchical levels. Afterwards, spatial and statistical comparisons were conducted between the results obtained from the interpretation of IKONOS imageries and existing land use inventories. The later were produced from Indian satellite images (IRS-1C, 5.8 m) merged with multispectral Landsat TM images (30 m), both acquired in October 2000, using visual interpretation techniques. These comparisons indicate the enhancement identification capabilities offered by IKONOS images in most cases and the limitations that can be offset in some other cases. IKONOS images show an improvement of land use maps spatial accuracy in mountainous areas more than coastal areas. Moreover, they were particularly effective at delineating impervious surfaces prevalent in urban areas, which are problematic to map from other low resolution remotely sensed data. The visual interpretation of IKONOS imagery performed at 88.5 percent level of accuracy, whereas the processing of Landsat TM and IRS yielded 82 percent accuracy. The degree of spatial coincidence between land cover/use maps was equal to 87%.
Satellite remote sensing is widely accepted as a technique to study land use. The later is of extreme importance in protecting water quality and controlling erosion and associated sedimentation. Two main approaches to land use mapping have been reported. The first analysis involves spectral classification (textural, structural) of satellite imageries into land use categories while the second method, i.e. visual interpretation, depends on several image characteristics (e.g., tone or colour, texture, size, shadow, pattern, location and associations) in order to identify and deduce the significance of the components of the image. The majority of those characteristics are not used in conventional digital classifications. Although, digital classification techniques are considered much less subjective than visual interpretation; land use classes vary spectrally, especially when land covers present high spatial complexity.
Moreover, significant advantages of visual interpretation of image products over classified images can be distinguished as follows: 1) less time required with photo interpretation methods to create a usable product; 2) little expense beyond the acquisition of the image; 3) image illumination problems such as shadows and brightly illuminated surfaces can be used as an interpretation aid; and 4) minimal expertise required to interpret the image. Several attempts using different approaches (spectral image classification, visual interpretation) have been made to test the accuracy of each method (Conese and Maselli, 1991; Janssen and Vanderwel, 1994; Mas and Ramirez, 1996; Palacio-Prieto and Luna-Gonzalez, 1996; Bethel et al., 2001). These studies have demonstrated the performance of visual interpretation methods over automated ones in getting a broad-picture view of an area to understand land cover types and patterns.
Recent advances in technology have made a tremendous contribution to remote sensing through the launching of new digital sensors and improved algorithms to process imagery. Hence, the use of very high resolution satellites (IKONOS, Quickbird) had implemented new challenges to land use mapping and is becoming substitute for data derived from time consuming aerial photo interpretation. The objective of this study is to assess the capability and effectiveness of IKONOS data for such a purpose through comparing the existing land use inventories with the results that can be obtained via the interpretation of IKONOS imageries. This assessment can help researchers in identifying the best remote sensing data sources needed to create land use inventories.
2. Study area
To apply such methodology, the Lebanese coastal area and 6 others sites located in Mount Lebanon have been used (Figure 1).
Figure 1. Location of the selected areas within Lebanon.
The selection of the Lebanese coastal region was dependent on two main criteria, i.e. an altitude inferior to 100 m and slopes less than 10%. Its total area sums up to 335.8 km2 hence 3% of the total area of Lebanon. It extends 220 km from north to south and it spans 1.5 km from east to west. This area includes the main cities in the country and concentrates industrial, commercial and financial activities.
“Mtain” and “Marjhime” zones cover approximatively the same area 35.5 km2, with the biggest part (56%) on mountainous crests rising up to 1500 m. Qornayel area (8 km2) comprises the same physiographic features as Mtain area, but with the major part on upper slopes (1000-1500 m) (90%). Dmit area, occupying 3.5 km2, is the only pilot area extending from the coastal plain (< 100 m) to the medium slopes (500-1000 m). Ramlieh area (3.56 km2) and Ras El-Matn area (12.99 km2) occupy mainly the medium slopes of Mount Lebanon (1000-1500 m). The land-cover features are diverse in these selected areas, such as agriculture area, plantation and forest.
3.1. Data collection and description
IKONOS images are acquired as 1-meter resolution panchromatic and 4-meter resolution multispectral images. While the panchromatic images represent the visible range of the spectrum, the four bands of multi-spectral images represent the red, green, blue and near infra-red range of the spectrum. Pan-sharpened color imageries have been generated in this study by merging 1-meter resolution panchromatic image with the 4-meter resolution multispectral bands to generate true-color or a false-color images at 1-meter resolution. These 1-meter resolution color images provide exceptional depth in color and clarity of detailed for feature extraction. In addition, the color pan-sharpened images, compared with the panchromatic ones, required lesser supervision, provided more operator ease, reduced instances of misinterpretation and provided for higher speeds in interactive interpretation-cum-delineation.
Twenty two IKONOS scenes (each 121 km2) acquired in October 2000 were purchased for the study area. These images were initially registered and ortho-rectified with geographic lat-long coordinates using ground control points (GCPs) from a digital terrain model DTM (10 m), generated from toposheets at 1:20,000 scale.
3.2. Image processing
A visual interpretation has been performed on pan-sharpen IKONOS imageries. This interpretation consists of marking the boundaries of areas representing single cover units on the images using on-screen digitizing to assign nomenclature headings, as well as extrapolating the established delineation and identification of different parts showing similar characteristics.
The Land Cover Classification System (LCCS) software developed by FAO (2005) has been applied on these images for the formulation of the legend to comply with the CORINE (Coordination des informations sur l’environnement) multi-level classification system that enables the end-user to dynamically select the depicted types and scale. This system generates two main phases: an initial dichotomous phase, in which eight major land cover/use types are defined (artificial areas, agricultural areas, wooded lands, grasslands, wetlands, unproductive areas, water bodies and roads), followed by a subsequent modular-hierarchical phase, in which land cover/use classes are created by the combination of sets of predefined classifiers tailored to each major land cover/use type in order to use the most appropriate classifiers and to reduce the likelihood of impractical combinations of classifiers. Therefore, the structured legend meets three basic requirements namely mapping all territory leaving no heading for unclassified land, matching the headings with the needs of future users of the geographic database and avoiding vague or ambiguous heading terminology.
3.3. Ancillary data
Existing land cover/use maps at 1:20,000 scale produced through visual interpretation of Landsat TM (30 m) and PAN IRS-1C (5.8 m) images acquired in October 2000 based on CORINE Land Cover methodology (level 4) (LNCRS-LMOA, 2002) have been compared to the results obtained from the interpretation of IKONOS imageries.
The working scale (display of images, identification of features and digitizing) was set at the 1:5,000 in order to allow a finer presentation of the existing maps at 1:20,000. The minimum mappable unit (MMU) was then set at 500 m2 in relation with the 1:5,000 scale. Though this MMU facilitates the legibility of the printed map, allows an easy digitization from the interpretation manuscripts and is ideal in homogeneous areas like high mountains and large agricultural areas where spatial variation occurs at a minor rate, there was a need to refine it even further in urban or transitional areas, where changes in land use occur at a higher frequency.
3.4. Accuracy analysis
To determine the accuracy of each land cover/use map, 200 field sites were selected for comparing ground information with interpretation results. These sites were located using a GPS with a precision of 10 m. They have been chosen by a random stratified sampling to cover all obtained land cover/use classes and sub-categories. Two methods of accuracy assessment were used in this study. First, the overall (global) accuracy was estimated from the confusion matrix by calculating the total percentage of sites correctly classified. The average accuracy was then calculated as the average of the accuracies obtained for each class.
4. Results and discussion
4.1. Detection of land use classes
The identification (or on-screen digitizing) of different land cover/use classes ramified at level 4 on IKONOS images can be explained as follows giving some concrete examples: 1) The dense urban fabric class indicates 80% urban cover of a given area distributed non linearly (Figure 2a); 2) The medium density urban fabric refers to a mixing of urban (˜ 70% of the total area), vegetation and bare soil (Figure 2b); 3) The low density urban fabric can be extended on a linear or non linear surface (Figure 2c); 4) Diverse equipment regroups schools, universities campus, etc. (Figure 2d).
In addition, some classes can be misinterpreted such as railway station class that can be confused with industrial or commercial area class regarding its texture, necessitating a prior knowledge of the area or a topographic map for its identification (Figure 2e). Urban extension class is similar to urban vacant land class (Figure 2f), both sharing a future urbanization. The only distinction is based on the fact that urban extension occupies a more important and better infrastructure that urban vacant lands.
Figure 2. Examples of land use classes as shown on pan-sharpen IKONOS images (1 m).
Other classes are difficult to extract such as dumpsites representing variable spectral signature resembling to that characterizing other land cover classes (Figure 3a). While some are easily detected like mineral extraction sites appearing in white patches or land fill sites located always in proximity to the sea with a limited number (Figure 3b). The green colour appearance and the particular spatial identity allow an ease differentiation of olives from other types of vegetation (Figure 3c). Citrus appear in brighter colours than fruit trees and with a more compact density (Figure 3d). The texture of bananas is disorganized because of their big leaves that overlap (Figure 3e). Their colour varies between dark green and yellow. Greenhouses represent the easiest class to identify on IKONOS images because of their remarkable reflection (Figure 3f). The latter appears similar to a mirror once the plastic houses have small dimensions.
Figure 3. Difficulties of interpretation of some land use types on IKONOS imageries.
4.2. Spatial and statistical comparisons between LUC maps
The number of polygons equal to 4355 in land use/cover map (LUC) produced from Landsat and IRS imageries has increased to 5934 in the map resulting from the interpretation of IKONOS imageries, within a difference of 1500 entities. Moreover, the statistical comparison achieved at 2 land cover/use levels delineates closer difference percentage values between the 2 maps (1:20,000 and 1:5,000) in coastal areas (varying between -0.74 and +1.04) than in mountainous ones (extending from -2.39 to +2.49) (Table 1). This indicates the importance of using IKONOS imageries in land use mapping of mountainous remote areas, competing aggressively against topography.
Table 1. Statistical comparison of LUC classes produced from IKONOS (1 m) and merged Landsat TM (30 m) and IRS (5.8 m) imageries.
If we compare the concordance (in km2) of major land cover/use classes in the two corresponding maps through considering only the coastal area, we can evaluate the efficiency of using IKONOS images. Therefore, a confusion matrix was established indicating an overall coincidence of 87% (Table 2). Some LUC classes are less matched between the 2 maps than others, e.g., roads, unproductive lands and grasslands.
Table 2. Concordance (in km2) of LUC classes produced from IKONOS (1 m) and merged Landsat TM (30 m) and IRS (5.8 m) imageries.
4.3. Causes of spatial differences between LUC maps
The 2 LUC maps are showing several spatial differences that can be explained as follows: 1) a “5 times” better spatial resolution in the LUC maps produced in the current study (1 m IKONOS images against 5.8 m pan-sharpen Landsat TM-IRS images); 2) a bad positioning of the coastal line in LUC maps resulting from Landsat and IRS images with a shifting attaining in some cases hundred meters (Figure 4). For that, the line of high waters was adopted in the current study to define the maritime national limit since it is easily detected by the colour of the foreshore uncovered by the tide with a mean amplitude of 35 cm (Durand, 1998), but it is sensitive to seasonal evolutions of beaches; 3) photo-interpretation errors (digitalization) that can related to a bad georeferencing, a low accuracy of digital data and a high subjectivity of interpreters.
Figure 4. Coastal line detection in LUC maps.
The accuracy has been improved using very high resolution data. This ensures previous results done by Zhou and Li (2000) and Devriendt et al. (2005); and 4) an incompatibility of codification between land use maps produced from different satellite types.
4.4. Field verification of LUC maps
The use of IKONOS pan-sharpen imagery resulted in a more accurate LUC map (total precision 88.5%) as compared to previous LUC maps (total precision 82%) produced from other lower resolution remotely sensed data. The improved accuracy of the IKONOS imagery is most apparent in an urban environment where there is a large proportion of impervious surfaces.
Table 3. Field accuracy Assessment of land cover/use (LUC) maps
The user’s accuracy, which is the percentage of sites in a class derived from visual interpretation, correctly classified vis-à-vis the reference data (field), is ranging between 74 and 100%, with relatively low errors of commission (excesses), varying between 0 and 26% in the case of IKONOS images. This accuracy is much lower, once we consider merged Landsat TM with IRS images ranging from 70 to 92% with higher errors of commission (8-30%). The producer’s accuracy, corresponding to the percentage of sites of a reference class correctly classified by the images, is ranging between 74 and 95%, with similarly relatively low errors of omission (deficits) depending on the class into consideration (5-26%) for IKONOS images. Similarly to user’s accuracy, Landsat TM merged with IRS show higher errors of omission (12-33%) than IKONOS images. Therefore, IKONOS imagery adds additional quantifiable spatial and spectral advantages over lower resolution spatial data in land use mapping.
This study has shown that IKONOS imagery shows potential as a source of data within a national mapping agency, being more accurate than other pan-sharpen images. The created LUC map at 1:5,000 through the visual interpretation of IKONOS images serves the diverse needs and applications of the agricultural and urban development, policy making, risk assessment and planning. The applied methodology can be extrapolated to the whole country, being homogeneous, smooth, easy to obtain and updated. The time required for the visual interpretation of 2.3 km2 is equal to 1 hour and a half, therefore we need 6533 hours for the treatment of the whole country (studied areas excluded).
- Bethel, J.S., McGlone, J.Ch., Mikhail, E.M., 2001. Introduction to modern photogrammetry. John Wiley & Sons Inc., New York.
- Conese, C., Maselli, F., 1991. Use of multitemporal information to improve classification performance of TM scenes in complex terrain. ISPRS Journal of Photogrammetry and Remote Sensing, 46(4), 187-197.
- Devriendt, D., Goossens, R., Taillieu, K., Dewulf, A., 2005. Improving spatial information extraction for local and regional authorities using very high resolution data – geometric aspects. New strategies for European remote sensing. Millpress Science publishers, Rotterdam, ISBN 90 5966 003X.
- DURAND Paul, 1998. Cinématique d’un littoral sableux é partir de photographies aériennes et de carte topographies, Géomorphologie: relief, processus, environnement, n° 2, p. 155-166
- FAO, 2005. Land cover classification system, classification concepts and user manual software version (2). UNEP, FAO, Cooperazione Italiana, ISNN 1684 82 41, 208p.
- Janssen, L.L.F., Vanderwel, J.M., 1994. Accuracy assessment of satellite derived land-cover data: A review. Photogrammetric Engineering and Remote Sensing, 60(4), 419-426.
- LNCRS-LMOA, 2002. Land cover/use map of Lebanon at a scale of 1:20,000. Lebanese National Council for Scientific Research and Lebanese Ministry of Agriculture.
- Mas, J-F., Ramirez, I., 1996. Comparison of land use classifications obtained by visual interpretation and digital processing. ITC Journal, 3(4), 278-283.
- Palacio-Prieto, J.L., Luna-Gonzalez, L., 1996. Improving spectral results in a GIS context. International Journal of Remote Sensing, 17(11), 2201-2209.
- Zhou, G., Li, R., 2000. Accuracy evaluation of ground points from IKONOS high resolution satellite imagery. Photogrammetric Engineering and Remote Sensing, 66(9), 1103-1112.