Evaluation of Semi-Automated Image Processing

19 downloads 0 Views 238KB Size Report
May 13, 2005 - from the raster to vector conversion of the Sagar. Island shoreline raster file. Due to various diffi- culties involved in the format conversion of ...
International Journal of Geoinformatics, Vol. 1, No. 2, June 2005

Evaluation of Semi-Automated Image Processing Techniques for the Identification and Delineation of Coastal Edge using IRS, LISS-III Image – A Case Study on Sagar Island, East Coast of India Murali Krishna, G. 1*, Mitra, D. 2#, Mishra, A.K.2, Oyuntuya, Sh. 3 and Nageswra Rao, K.1 1 Department of Geo-Engineering, College of Engineering, Andhra University, Visakhapatnam-530003, India 2 Indian Institute of Remote Sensing (IIRS), 4 Kalidas Road, Dehra Dun-248001, India 3 Mongolian State University of Agriculture (MSUA), Ulaanbatar, Mongolia E-mail: *[email protected]; #[email protected] Abstract To determine the most reliable technique for identification of coastal edge in the IRS-IC LISS-III satellite imagery, various semi-automated methods were tested.†To define which pixels fall into land and water classes, methods like Level Slicing, PCA, NDVI, WI and ISODATA classification were tested. Due to the strong contrast between land and water in its SWIR band (1550-1700 nm), IRS-IC LISS-III imageries are well suited for generating the Sagar Island coastal edge, which is taken as a case study. After the classification of land and water pixels in the image, the raster boundary between land and water was identified and the shoreline was delineated. A shoreline vector was derived along the land/water raster boundary by raster contouring. The results were evaluated by comparing them with visually interpreted data of high-resolution PAN imagery acquired on the same day and time and GPS survey.

1. Introduction Precision plays a decisive role when the coastal zone managers wanted to identify and delineate the shoreline boundary using satellite imagery. Delineation of shoreline is a difficult task due to the presence of intertidal mudflats and marshy/ swampy areas along the coast most of which are usually misclassified as part of water. Pixels quite often represent the mixture of different spatial classes (Ghosh et al., 2002). Intermixing pixels of water-saturated land that represent shallow water bodies in the satellite imagery affects the discrimination of accurate shoreline boundary.

Various classification methods such as threshold level slicing, spectral enhancement, Principal Component Analysis, Normalized Difference Vegetation Indices, Water Index, spectral classification (ISODATA classification) and Tesseled Cap Transformation have been in use for the discrimination of land/water boundary, (Bagli and Soille, 2003) and (Frazier and Page, 2000). Though there are differences of opinion as to which one of the above is the best method for coastal edge identification, it is necessary to find out the best amongst them while engaged in coastal mapping in order to reduce the processing time and for accurate results, (Smith, 1997), (Frazier and Page, 1

International Journal of Geoinformatics, Vol.1, No. 2, June 2005 ISSN 1686-6576/ Geoinformatics International

IJG_1093

1

13/05/2005, 17:17

Evaluation of Semi-Automated Image Processing Techniques for the Identification and Delineation of Coastal Edge using IRS, LISS-III Image ñ A Case Study on Sagar Island, East Coast of India

2000) and (Ryu et al., 2002), and (Braud and Feng, 1998). Accurate demarcation of coastline from Synthetic Aperture Radar (SAR) data have been attempted by Lee and Jurkovich (1990), Shang et al., (1994), and Yu and Acton (2004). Computer cannot recognize the distinction among various classes in the image as humans can. However, if the process cannot be automated, the economic benefit of the imagery is most certainly lost (Allan, 2001). Consequently, if one desires modeling of erosion/accretion, measuring of shoreline length, or analyzing of the land/water boundary, a few interim steps are required (Frazier and Page, 2000). The image must be processed and classified according to the respective Digital Numbers (DN) values to specifically define land and water, which inturn highlights the divisible boundary between both the classes. Further, this boundary can be extracted as a physically defined feature, i.e., a digital vector line. Braud and Feng, (1998) had developed a novel method to choose the best semi-automated method for the identification of land/water interface using Landsat TM data based on the concept that larger the percentage of accuracy of detection and clustering of water pixels in the classification, higher the accuracy of land/water boundary distinction. Hence, the semi-automated method, which can classify the maximum number of water pixels precisely, can be considered as accurate for the land/water boundary delineation. Threshold level slicing of Landsat TM Band 4 (MIR) was found to be the best suitable amongst the other semiautomated methods (Braud and Feng, 1998). An attempt has been made here to see whether the same method is applicable to IRS, LISS-III data for the discrimination of land/water boundary using semi-automated methods. Semi-automated algorithms such as level slicing of Band 5, ISODATA, PCI, NDVI, WI techniques were taken for the present study. These algorithms are based solely on spectral analysis of individual pixels without taking into account the texture, shape, morphology, and context of regions in the images

(Whithe and El Asmar, 1999). And at the same time these techniques are simple, less time consuming and can be applied for large area mapping irrespective of the complexity of the region (Frazier and Page, 2000). In this connection, an example of Sagar Island along the east coast of India has been taken for the present study as its coast is composed of several sub-tidal and inter tidal mudflats and constantly experiencing severe erosion/accretion (Ghosh et al., 2000). 1.1. Data Used The Indian Remote Sensing satellite (IRS), LISS-III data (acquired on 16th December 2001) in all the 4 bands (2, 3, 4 and 5, respectively) have been utilized for the study. The Survey of India (SOI) topographic maps (No. 79 C/1 and 79 C/2), on 1:50,000 scale have been used as ancillary data. Classification accuracy (Arnoff, 1982) has been tested on a sample basis, assuming a binomial distribution for the probability of success and failure of sample tests. Control accuracies of the maps have been assessed as described by Patel et al. (1988).

2. Study Area The Sagar Island is situated between 21°37'21" and 21°52'28" N latitudes and 88°02'17" and 88°10'25" E longitudes (Figure 1) and is a part of the Sundarban mangrove forest at the mouth of Hooghly estuary on the east coast of India. Covering an area of about 250 km2, Sagar Island is the largest among others in the Sundarban deltaic complex.

3. Methodology IRS 1C, LISS III imagery pertaining to Sagar Island region was rectified and georeferenced with the SOI toposheets. The image was subset to the study area as the same was taken for the evaluation of all semi-automated shoreline extraction processing techniques.

2

IJG_1093

2

13/05/2005, 17:17

International Journal of Geoinformatics, Vol. 1, No. 2, June 2005

Figure 1: Location Map of the Sagar Island

Level slicing of edge enhanced Band 5, Principal Component Analysis, ISODATA or Unsupervised classification, NDVI and Water Index methods were tested and evaluated for the identification and delineation of coastal edge of the Sagar Island. All the land cover classes in the output images of these methods were clustered into three major classes namely water, land and transition zone. Evaluation and accuracy assessment of the results of all algorithms was done by visually comparing them with raw data, application of random sampling techniques as well as com-

parison among each algorithm output. The most suitable method for land/water discrimination is found out on the basis of the accurate segmentation of largest percentage of water pixels in the IRS-LISS III image. A shoreline raster was defined by the topographic analysis method depending upon the difference between the DN values of the land and water classes. The final vector shoreline was derived from the raster to vector conversion of the Sagar Island shoreline raster file. Due to various difficulties involved in the format conversion of raster 3

IJG_1093

3

13/05/2005, 17:17

Evaluation of Semi-Automated Image Processing Techniques for the Identification and Delineation of Coastal Edge using IRS, LISS-III Image ñ A Case Study on Sagar Island, East Coast of India

to vector, the resultant vector data suffers from the problems like discontinuities and zigzag lines at the places of highly complex and irregular coastline. Processing the data in GIS environment eliminated these errors.

4. Techniques for Coastal Edge Discrimination The entire portion of the Sagar Island at Hooghly estuary covered by IRS-1C LISS-III imagery dated 16th December 2001 was subset using ERDAS Imagine software. A single data set has been taken for the present study to get a better comparison and accuracy assessment between each methodology at the same given image conditions which varies with image to image. The scene was cropped so that areas outside the coastal region were excluded. The subset image was utilized to test and evaluate all the semiautomated image processing techniques for the classification accuracy of the same. A total number of 799,254 pixels have been evaluated for each method. 4.1. Level Slicing of Band 5 Ellis et al., (1989 &1991), Tao et al. (1993) and Tittley et al. (1994), studied the utility of infrared band in the identification of land/water interface using Landsat Thematic Mapper (TM) data. In the present study, a similar attempt is made using IRS/LISS-III Band 5, i.e. the Shortwave infrared (SWIR) band. A moderate 3 × 3 edge filter was applied to sharpen the image. Edge enhancement helps in distinctive identification of coastal edge by increasing the local variance of the feature classes and minimizing the ambiguity in classifying the transition zone. The coastal edge of the Sagar Island was determined in the edge enhanced Band 5 by identifying land, water and transition zone classes automatically using a simple density slice (threshold techniques). Land, water, and transition zones were segmented using the threshold values

obtained by empirically adjusting the interactive contrastive histogram. The threshold values derived for land, water and transition zone are 99-255, 1-16 and 17-98 respectively. Subsequently, based on their threshold values the three feature class pixels were segmented by assigning them with 1. DN values less than one (