Morphological Automatic Extraction of Coastline from Pan-European Landsat TM images

Stefano Bagli, Pierre Soille

EC-Joint Research Centre - Institute for Environment and Sustainability, Ispra (IT)

The delineation and extraction of coastline and water bodies, e.g. rivers and lakes, is an important task useful in different fields such as coastline erosion monitoring, coastal zone management, watershed definition, flood prediction, and evaluation of water resources. This task is difficult, time consuming, and sometimes impossible for a huge region, such as an entire continent when using traditional ground survey techniques because water bodies can be fast moving as in floods, tides, and storm surges or may be inaccessible. In addition, automatic and replicable techniques are required to update coastline maps, evaluate the spatial and temporal evolution of alterations due to natural and anthropic events, and extract the waterline for vast regions. Following the increase in the availability of satellite images, the development of tools for geographic data analysis (GIS platforms) and image processing techniques, numerous research studies have been carried out to extract and delineate water bodies from these images. The extraction of features, such as coastlines and water bodies directly from satellite images overcomes the problem of matching available coastline data sets with the studied image data set. In fact, owing to projection system biases, the matching of a coastline coming from a different data set together with the available images may turn out as a tedious if not impossible task. Beyond manual digitalization, several techniques have been reported in the literature for the derivation of the coastline position from satellite images. The most common are density slice using single or multiple bands and multispectral classification both supervised and unsupervised (e.g. ISO-DATA, PCA, Tasseled Cap, NDWI). 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.
Motivated by the need to generate a pan-European coastline database from Landsat 7 TM images, which were collected for updating the European Corine Land Cover database (Image 2000 - Project), we present a new methodology for extracting automatically the coastline and lakes and its application to the entire European continent. Our approach consists of the combination of spectral and spatial information for the images using morphological image segmentation techniques. Mathematical Morphology (MM) or simply morphology is a theory for the analysis of spatial structures and a powerful image technique known as morphological image analysis. In MM, 2-D grey tone images are seen as 3-D sets by associating each image pixel with an elevation proportional to its intensity level. Another set of known shape and size, called the structuring element (SE), is then used to investigate the morphology of the image. MM transformation and operators are therefore looking for objects defined as a specific spatial arrangement (shape, texture) of image pixels, rather than a single or cluster of pixels with a specific spectral signature (classification techniques). The delineation and extraction of objects from images through MM consists of the application of a sequence or chain of morphological operators aimed at iteratively suppressing all undesirable spatial structures occurring in the image. For our scope, we applied a Seeded Region Growing (SRG) algorithm initially proposed by Adams and Bishof. It is based on the postulate of region growing algorithms, where a criteria of similarity of pixels is applied, but the mechanism of region growing is similar to the morphological watershed algorithm. Instead of controlling region growing by tuning homogeneity parameters, SRG is controlled by choosing a small number of pixels, called seeds. It starts with the selection (manual or automatic) of a number of seed regions to which a region growing technique is applied. At each step, the algorithm proceeds by adding one unassigned pixels to one of the above sets until there are no more unassigned pixel in the image. This results in a tessellation of the image into the same number of regions, as those given by the seed regions. The boundaries of each homogeneous region are extracted by determining the boundary of the corresponding region using a gradient function. In our case, each seed region belongs to either sea/water or land/soil. Seed regions were identified automatically on multi spectral bands and panchromatic Landsat images using a simple density slice (threshold techniques) on single or multiple (NDWI) bands. The threshold values for extracting land and sea/water seeds were derived empirically on test images and they remain valid for the entire area of interest. The main requirement for identifying seeds, consists of the delineation of representative pixels for each class without including noisy pixels that can invalidate the seeded region growing procedure.
The methodology was applied using different multi spectral bands for identifying seeds for region growing, but better results were obtained on band 5 (Short-Wawelenght Infrared SWIR) with a resolution of 25 m. Panchromatic images with a resolution of 12.5 m produce a more detailed coastline. The paper will presents a discussion on the quality of the coastline results obtained for different European regions.