Regional Land Cover Mapping

Sushil Pradhan
GIS Analyst
International Centre for Integrated Mountain Development
Kathmandu
[email protected]

A GIS and remote sensing (RS) based methodology has been developed and tested for the land cover mapping of Bhutan and Nepal using IRS-WiFs data. The study mainly focused on generating good and reliable training samples, for the accurate classification of the image.

The Hindu Kush-Himalayan (HKH) is the mandated area of International Centre for Integrated Mountain Development (ICIMOD) covering eight regional member countries: Afghanistan, Bangladesh, Bhutan, China, India, Myanmar, Nepal, and Pakistan. The HKH region contains world’s highest, largest, and most populated mountain systems. To manage these scarce resources effectively it is important to study and understand the dynamics of land use and land cover change of the region and to determine which factor contributes significantly in a specific area before proper land use planning can be done. Land use/land cover change study is a diagnostic tool for determining sustainability and is therefore important that this tool can be done carefully and properly for the sustainable development of the HKH region.

Objectives
The primary goal of the project is to understand the dynamics of land use and land cover change in mountain ecosystems of the HKH region. To meet this primary goal, a pilot study has been carried out, which is the main focus of this paper, with the main objective as follow:

• Develop a standard methodology to understand and explain the pattern of land cover characterisation of the region using satellite images.

The specific objectives of the study are:

• To investigate the spectral information content of different land cover types to generate spectrally homogenous and spatially significant training samples;
• To study to improve the land cover classification using knowledge-based rules, topographic factors, and higher resolution remotely sensed data;
• To study to produce detailed forest types classification; and
• To assess the accuracy to validate the proposed methodology.

Sources of Data
The study used the IRS WiFs (Wide Field Sensor) satellite data. The WiFs satellite data has two bands – Red and Near Infrared (NIR), with spatial resolution 188.3m. The project acquired the WiFs satellite data between the periods of 1996-1999 covering the whole HKH region. The study also used other ancillary data, DEM (Digital Elevation Model), rainfall and temperature data to meet the main objective.

Methodology
The study acquired 12 scenes of the WiFs satellite data covering the whole HKH region. Each of these scenes were rectified and geometrically corrected using ground control points (GCPs) from Defense Mapping Agency Aerospace Centre (DMAAC), Missouri, USA. All the GCPs were verified in the Operational Navigation Chart (ONC) of scale 1:1000,000, and the same location were identified on the images and registered using Erdas Imagine 8.4 software. Overall root mean square error (RMS) was limited within a pixel. Then it was resampled to pixel size 180m x 180m. Resampling was done using nearest neighborhood method, which maintains the original DN of pixels.

The images were projected into Albers Conical Equal Area with WGS84 spheroid and datum. After geo-referencing, all the images were combined (mosaic) into a single image as shown in Fig. 1. The conceptual model applied in the study to meet the overall objective is depicted in Fig. 2. After geo-referencing and mosaicing of the image, the image were subset to individual band. From each band, the training samples were generated. Remote sensing technology sounds very interesting and is useful, but its quality highly depends on the quality of training samples. It requires good training samples for the accurate image classification. Good training sample means it should be spectrally homogenous and spatially large enough. The main theme of this study to generate good training samples, i.e. spectrally homogenous and spatially significant, for the image classification. Due to limitation of the human eye, we can’t distinguish the spectral homogeneity of the pixels. Therefore, an algorithm was used to extract the spectrally homogenous pixel groups (Fig. 3). To generate spectrally significant and spatially homogenous training samples, following steps were carried out:

Figure 1 : Mosaic of 12 scenes of WiFs data of the HKH region

Figure 2 : Overall methodology of the study

Figure 3 : A process for extracting spectrally homogenous and spatially defined sample areas

• Derived the possible spectral groups of pixels of Red and near infrared (NIR) bands that are within the specific range of values, e.g. 5. The pixels having 0 histograms were omitted while selecting. Pixels within the defined ranges were recoded to unique values. The pixels belonging to same contiguous groups were grouped and given unique identifier (Fig. 4).

  
  Figure 4 : Neighborhood analysis to form a group of pixels and assignment of new ids to each group

• The appeared salt and pepper pixels were eliminated by defining area criteria (Fig 5a & 5b). A threshold of 50 pixels, i.e. 1.62 km2, was defined to obtain spatially homogenous training samples. The group of pixel was again recoded to an integer identifier of particular spectral class. These samples were used as the basic units for regional reviews and legends for land use land cover mapping which indicate the basic spatial unit.

  
  Figure 5 : Before (a) and after (b) applying area criteria

• Number of Area of Interest (AOI) was created from each spatial unit in each band (Fig 6). While creating AOI, standard deviation was maintained within the limit of less than or equal to 2.0 in each class.

  
  Figure 6 : Creation of AOIs from each sample

• After delineating AOI from Red and NIR band, they were merged together as a single file.

After merging the AOIs, statistical report was derived. The homogeneity of training samples was measured by means of standard deviation (d), which can be viewed as providing the a measure of the uncertainty. So, higher the d, there will be more uncertainty. Low d of a set of data values indicates how similar enough they are. The classes having highest frequency numbers of pixels in order and low d are selected as the final training samples for the land cover types identification.

Image Classification
Thus generated spectrally homogenous and spatially significant training samples are the basic spatial units for ground truth and primary data collection for the image classification.

Case of Nepal
The WiFs satellite data was first subset based on the international boundary of Nepal (Fig 8). Using the methodology (refer to Fig.3) as described above, spectrally homogenous and spatially significant training samples were generated. The field visit was carried out to verify the land use and land cover types in each of these training samples. In this process, the existing land use and land cover data (LRMP 1978/79) of Nepal was used as reference. Then the image was classified using these ground truth data (Fig. 9) and the standard land use/land cover classification scheme was maintained while classification the image. For the detailed forest types classification, on forest types by altitude, and temperature and rainfall data at different altitudes of Nepal (Table 1) were collected. These information were integrated with a DEM (Digital Elevation Model) and the newly classified land use and land cover data for the detailed forest type classification (Fig. 9). The percentage of different land use and land cover types are given in Table 2. About 10% of the total sample areas were randomly selected for the accuracy assessment. The overall accuracy assessment of the result was 88%.

Table 1 : Climate belts of Nepal

Table 2 : Percentage of land cover types obtained from the result

Figure 7 : WiFs satellite data of Nepal

Figure 8 : Land use and land cover classification of Nepal based on WiFs satellite data, 1996

Figure 9 : Detailed forest type classification for Nepal

Case of Bhutan
Similarly, the image for Bhutan was subset (Fig. 10) and the spectrally homogenous and spatially defined training samples were generated for the ground truth data collection.

Figure 10 : WiFs satellite data of Bhutan
For various reasons, no field visit was carried out. The Ministry of Agriculture of Royal Government of Bhutan provided the digital version of land use and land cover data of 1994, which was produced using Landsat-TM data with thorough field verification. Assuming land cover change within two years in mountainous country like Bhutan would not be much, the provided digital land cover data of 1994 was used as the reference for image classification. For this,

• The existing land cover data of 1994, was rasterized using 180m x 180m pixel size (same as of WiFs). Unique cell value was assigned to different land cover types.
• The above generated training samples were rasterized using the same pixel size. The cell value 1 was given to all.
• The rasterized training samples (2) was overlaid with 1994 land cover data (1). It resulted the training samples with certain land cover codes as of 1994. Different types of land cover within an AOI are identified (which is the main purpose of field visit) (Figure 11a & 11b).

  
  Figure 11 : (a) Training samples (b) samples overlaid with 1994 data

• The resulted samples (3) with different land cover codes was converted to Imagine format and used for classifying the image. The maximum likelihood method was applied for classification.

Classified data often manifest a salt-and-pepper appearance due to the inherent spectral variability encountered by a classifier when applied on a pixel-by-pixel basis (Lillesand and Kiefer, 1994). The 3-by-3-majority filter was applied to remove salt-and-pepper appearance.

Post classificationwas mainly required due to existence of cloud cover. Due to the lack of mid infra-red and thermal infra-red channels in WiFs data, there was mis-classification between cloud and snow. The area covered by clouds were masked with null values by screen digitization, and later, neighbourhood classification of the masked area (null values) was done. The whole classified image and the classes under clouds was improved by applying knowledge-based rules, taking the land cover data of 1994 as reference. The final classified map is presented in Fig. 12.

Figure 12 : Land cover classification of Bhutan based on WiFs satellite data, 1996
Accuracy assessment of the classified image was done with 143 randomly selected points. For this purpose, the land cover data of 1994 was taken as the reference. The overall accuracy of the classified image is 83.10%.

Result and Discussion
From the analysis, broadleaf forest and coniferous forest has been found as the dominant land cover classes, which is 43.3 and 26.6 percent, respectively. The coverage area of different land cover classes and their percentage are given in Table 3. LandSat-TM (30m), for the identification of different agricultural land use types and forest types. There is no doubt that it would produce more accurate result, but the question is about the cost, which is about 20 times expensive than using the IRS-WiFs satellite data. Therefore, WiFs data is good enough to be incorporated with the proposed methodology for a regional or national level study. The methodology will work well and is recommended to use at a watershed level using medium or higher resolution satellite data.

Table 3 : Area Coverage of Land Cover Types

Reference

• The Global Environmental Change and Human Securities (GECHS), https://www.gechs.org
• Global change SysTem for Analysis, Research, and Training – START, https://www.start.org
• Land Use and Land Cover Change – LUCC, https://www/geo.ucl.ac.be/LUCC/lucc.html
• International Geosphere-Biosphere Programme (IGBP),
• International Human Dimensions Programme on Global Environmental Change (IHDP),
• People and Pixels – linking remote sensing and social science, National Academy Press, Washington, DC, USA, 1998.
• ICIMOD,
• Pradhan, S.: Progress report on land cover classification of Bhutan using IRS-WiFs satellite data, MENRIS / ICIMOD,2000.