Mapping coral reef benthic substrates using hyperspectral space-borne images and spectral libraries

Abstract

The suitability of Hyperion, the first civilian hyperspectral sensor in space, for mapping coral reef benthic substrates has been investigated in this study. An image of Cairns Reef, in the northern section of the Australian Great Barrier Reef (GBR), was acquired during Hyperion Calibration and Validation activities. A field experiment was carried out on Cairns Reef to collect information about the optical properties of the water in the area and to map benthic cover by means of video transects. An approach was used to classify the Hyperion image that allows convenient mapping of benthic substrate type and water depth simultaneously. A hyperspectral library of radiance at Hyperion altitude was simulated using a spectral library of GBR benthic substrates, a Hydrolight 4.1 radiative transfer model, and an in-house atmospheric model similar to Modtran-3.7. The image was then classified using the Hyperion at-sensor radiance data and the Spectral Angle Mapper metric using the simulated at-sensor spectral library. The results suggest that using spectral libraries created with forward modelling from the sea bottom to top of the atmosphere are useful tools for interpretation of reefs and can give better results in image classification than classifying the image after removing atmospheric and water column effects. The results also suggest that bottom type and water depth can be separated and mapped simultaneously provided hyperspectral data is available.

Introduction

Multispectral satellite sensors such as Landsat (both the lower resolution Multi-Spectral Scanner (MSS) and higher resolution Thematic Mapper™), as well as SPOT, have been used successfully since the mid-1980s map reef extent, major geomorphologic zones and broad substrate types (Jupp et al., 1985, Bour et al., 1986, Luczkovich et al., 1993, Ahmad and Neil, 1994, Maritorena, 1996, Morel, 1996, Knight et al., 1997, Mumby et al., 1997). Remote sensing sensors (known generally by their acronyms) have improved over the last three decades, culminating in the currently available airborne hyperspectral sensors (e.g. CASI, AISA, AVIRIS, HyMap), high-resolution space-borne multispectral sensors (QuickBird, IKONOS) and the first civilian hyperspectral satellite sensor Hyperion, which was launched in November 2000. The improved spatial and spectral resolution of these sensors and then development of complementary software to process the sensor data has vastly increased the number of geomorphologic sub-zones detectable by remote sensing, especially when using airborne data (Borstad et al., 1997, Clark et al., 1997, Pratt et al., 1997, Holasek et al., 1998, Mumby et al., 1998a, Mumby et al., 1998b, Green et al., 2000). However, the suitability of space-borne hyperspectral data for worldwide mapping of coral reef benthos has yet to be fully tested and established. It is the aim of the present paper to advance this discussion.

There are number of ways to analyse the remote sensing images. An approach that fits well with reef mapping as carried out by reef scientists is classification of the benthic areas into reef zones, substrate types and cover by coral and algal associations. Commonly used methods of image spectral classification that have been applied to interpret the remote sensing images proceeds (e.g. Ahmad and Neil, 1994, Mumby et al., 2004) by dividing the image into as many different classes as are statistically meaningful using unsupervised or partly supervised classification methods and then naming each class on the basis of available field survey data, or the knowledge of a skilled interpreter augmented by image statistics. This approach can provide useful reef maps but it has the disadvantage that the results are not easily comparable between images, because the number and types of classes in images acquired with the same instrument at the same location may vary greatly depending on illumination and/or atmospheric conditions. It can also involve variable and subjective decisions by the interpreters.

It is also quite difficult in general to compare classification results and interpretations from different sites because there is currently no accepted and standard classification of coral reef substrates in use. Each author tends to use a classification most suitable for a particular region and the sensor used. Some attempts have been made to elaborate standardized classification systems for remote sensing (Kuchler, 1986a, Kuchler, 1986b, Mumby and Harborne, 1999, Green et al., 2000, Andréfouët et al., 2003). However, the classification systems have been based primarily on geomorphology, physiognomy, ecology or other criteria and their applicability to all available sensors (with different spatial and spectral resolution) is limited. Unlike geomorphological classifications, ecological assemblages do not lend themselves easily to standard classifications (Green et al., 2000). Assemblages of species and/or substrata exhibit considerable variability and several assemblages may inhabit each geomorphological zone (Fagerstrom, 1987). It must also be noted that the benthic assemblages tend to have less distinct boundaries than geomorphological zones making their mapping by remote sensing even more difficult.

Several coral reef monitoring programs use different habitat classes. For example, the Reef Check classification scheme (see http://www.reefcheck.org) is widely used. A benthic classification technique compatible with, and derived from, the classification system adopted for coral reef benthic surveys by the GCRMN (Global Coral reef Monitoring Network) is used in the Great Barrier Reef (GBR) to interpret video transect data (Page et al., 2001). A similar classification scheme has been developed by the U.S. National Oceanographic and Atmospheric Administration, NOAA (Coyone et al., 2003). The Reef Check classification scheme has been used for interpretation of remote sensing data in a single location (Joyce et al., 2004) where a lot of field survey data is available. Attempts have been made (Andréfouët et al., 2003) to use the Reef Check classification scheme and consistent image processing techniques in the case of multiple sites in different parts of the world. This experiment showed that while it was possible to provide valuable maps at each site it was not possible to use a single classification scheme everywhere in the world if multispectral data is used and image processing is carried out in the “classical way”. This method, in which reef information is extended from site information to surrounding areas also requires a large amount of field data, which is difficult to acquire for all coral reefs in the world.

One way to improve this situation is to develop a consistent and standardized optical classification with which to compare classification results for different sites—especially if different remote sensing sensors are used. This fits well with remote sensing, as the only information that optical remote sensing instruments can measure from different bottom types is their optical properties. In fact, any benthic classification scheme used in the interpretation of remote sensing data has to be based on the variations in optical signatures of different bottom types. Of course, the optical classification must also be meaningful from a coral biology and ecology point of view to be able to give results that can be used by other coral reef scientists and managers. Defining optical classifications that are also informative for reef health and ecology is a major goal of great interest and value.

Optical signatures of corals and other coral reef benthic types collected in different parts of the world (Maritorena et al., 1994, Holden and LeDrew, 1999, Kutser et al., 2000a, Schalles et al., 2000, Hochberg et al., 2003, Hochberg et al., 2004, Kutser and Jupp, 2006; Kutser et al., 2006) show that the shapes of reflectance spectra of live and dead corals, sand, seagrasses, green, brown and red algae are consistent in different parts of the world. This suggests that a classification of remote sensing imagery based on the optical signatures of different benthic types should be applicable in different parts of the world even if no field data is available.

A particular issue involved in interpreting coral reef images is the need to separate the bottom signal from the effects caused by the water column with variable depth overlying the reefs. Some methods have been proposed (Lyzenga, 1978, Lyzenga, 1981) for extracting water depth and bottom type from passive remote sensing data. Lyzenga, however, assumed that the water column is uniform over the scene and the signatures of optically deep-water pixels are needed to perform the transformations. The original Lyzenga method has been improved for waters with non-uniform composition (Tassan, 1996) and several algorithms have been proposed for determination of the bottom depth and composition in waters with variable content of particulate and dissolved material (Spitzer and Dirks, 1987). Those algorithms, however, need tuning for particular conditions. One of the aims of this study was to investigate whether it is possible to map water depth and bottom type simultaneously if hyperspectral remote sensing data is available?

Because of these needs, the amount of work that has had to be done by qualified personnel in the past to obtain substrate classification maps from a raw image has been very high (since atmospheric, water column and air-water interface effects need to be removed by complex software processing) and is limiting or preventing the operational use of the image data by managers or non-remote-sensing specialists. This is one of the main factors limiting wider use of remote sensing in monitoring and managing coral reefs. Automating and simplifying the procedures used in pre-processing and interpretation is necessary to widen the use of remote sensing data by coral reef ecologists and managers.

In this paper we study the possibility of using an alternative method for the classification of coral reef imagery that assumes some standard pre-processing has been applied to the data. It could potentially simplify interpretation and the general use of remote sensing data by users who are not remote-sensing specialists. This is possible since the basic spectral library collection and necessary modelling required by the method can be done in advance by optics/remote sensing specialists, and the following activities of image interpretation can be learned quickly and with minimal specialist remote sensing knowledge.