Semi-automatic multi-level approach for extraction of tidal channels from Aerial Photographs

Abstract
Tidal channels play a critical role in morphodynamics of tidal environment. In order to incorporate tidal channels for understanding this dynamics and for management of tidal zones it is essential to map these channels accurately at high speed. Mapping of tidal channels using land surveying is a cumbersome, time-consuming, and error-prone task. The remotely sensed data provide a suitable alternative for channel mapping. In a number of studies, employing remotely sensed data, the channels are mapped using manual interpretation and tracing. However, the manual approach suffers with several limitations including subjectivity in the outcome and inaccuracy. Furthermore, the manual approach is not suitable if the channel mapping is to be carried out for a large area or involving several images spread in time. In view of the above, development of an automatic procedure for tidal channel extraction from remotely sensed data is much warranted.

The aim of this paper is to generate a method which will work for aerial photographs, MSS and hyperspectral data. The results in this paper, however, are shown using an aerial photograph. The technique developed is a semi-automatic multi-level approach, which seeks to classify the image in a feature constrained space. The low level processing employs edge detection followed by hysteresis thresholding to identify edge locations. The centrelines of all features (bounded by edges) are identified using distance with destination transform. The user is prompted to identify a few channel pixels, using which channel spectral clusters are formed by ISODATA classification. The Mahalanobis distance is used to classify entire image using these clusters. At this stage, beginning from the seed pixels the channel centreline segments are located and connected to other similar components. The final step includes fattening the located channel centrelines to generate full channels. This paper presents the results obtained till this stage and will outline the further work proposed to make the procedure complete.

1. Introduction
The study of tidal channels plays a critical role in understanding the tidal morphodynamics of intertidal zone. The tidal channel morphology is still not well understood and differing schools of thought exit to explain their development (French and Stoddart, 1992; Pye, 1992; Steel, 1996; Pethick, 1992; Rodriguez-Iturbe and Rinaldo, 1997; Rinaldo et al., 1999). It is important to understand tidal channels and their evolution for various applications involving coastal management, salt marsh reclamation, artificial salt marsh development, and for managing lagoon cities like Venice (Fagherazzi et al., 1999).

Physical measurement of tidal channels, i.e. mapping of their planimetric extent and cross-section, is essential for their study. Conventionally, these measurements are obtained using land surveying. Despite availability of advanced equipments, land surveying fails to meet the requirement when accurate and dense measurements are needed within a short time. The satellite and airborne data have potential to be used for this purpose. These data are also capable of providing the historical measurements for channels. In most of the geomorphological studies employing these data the channel networks are extracted using manual tracing. This approach is accurate and appropriate if the area of study includes only a few images spread in time and space. However, with the availability of a large number of these images, as are being produced by various air- and space-borne sensors, and also the need to include these in morphological studies spanning wider in space and time, the manual methods of channel mapping become too cumbersome. Furthermore, as at every step of manual tracing the input of an operator is involved, the mapped channels also suffer from subjectivity. Considering the above, there have been attempts to develop computer based methods to extract channel information from remotely sensed and other spatial data. Lohani (1999) provides a detailed review of algorithms for extracting channels (mostly terrestrial) from low resolution Digital Elevation Models (DEMs). The tidal channels, in view of their different morphology than their terrestrial counterparts, require different treatment and the algorithms designed for terrestrial channels fail to extract tidal channels (Lohani, 1999).

Furthermore, in view of small sizes of tidal channels (up to a width of 30 cm) there is a need to use high resolution data. The recently emerged technique of airborne altimetric LiDAR provides high resolution DEM and have been employed by Lohani and Mason (2001) and Mason et al. (2003) for tidal channel network extraction. With the geometric information contained in them, LiDAR data should be better for channel extraction (Mason et al, 2003). However, there are circumstances in which only spectral data are available, e.g. mapping channels using historical images and cases when only spectral sensor is flown. Also, it is common now that with most of the LiDAR flights a spectral sensor is also flown for strengthening the data captured which increases the accuracy of information extracted. In view of the significant role that spectral data can play and their availability this papers attempts to use these for extraction of tidal channels.

2. Data used
The data used in this paper is an aerial photograph (Figure 1 (a)) of the tidal zone of Wash in the UK. The image consists of channels which can be interpreted manually owing to their 1) spectral signatures, 2) typical shape, and 3) connectivity. The aerial photograph is separated in red, green and blue bands. It can be observed from the image that despite the manual interpretation being easy, computer processing for extraction of channels is not a straightforward process. This is basically due to the high variability in characteristics (spectral signature, shape, connectivity etc.) of channels which makes it difficult to extract channels using a direct low level image processing technique.

Figure 1: (a) Original image; (b) Non-maxima suppressed edge image; (c) Distance transform image

3. Methodology
The methodology consists of several steps run in parallel and sequence. These steps are discussed below and shown in the flow diagram (Figure 2). These steps are coded in C under Sun Solaris environment.