Methodologically, our scientific goals mean that:
- Our products are one-time products. If they intend to reflect geological-scale processes, there is no need to recompute them every 5 years to observe changes, because changes will not be detectable. However, they can obviously be used to stratify change detection studies that seek to highlight different smaller-scale processes within one or more geomorphological classes, such as phase-shifts, strategy shifts, mortalities, hurricane impacts, etc. However our product would be only one component of a more complex set of image processing techniques aimed at reaching a very specific mapping goal (Andréfouët et al. 2001 a). The methods applied here are neither adequate for nor intended for change detection analysis.
- The properties/functions that our maps should highlight are not captured at the pixel level but at a higher level of organization (i.e. reef zones). Thus using pixel-level algorithms is possible but not immediately effective. A Landsat pixel may provide quantitative properties (e.g. in term of coral-algal-sand cover), like a SeaWiFS pixel provides chl a concentrations or AVHRR provides Sea Surface Temperature. But a pixel alone does not provide the qualitative information we would like to be inherent to the map, such as direction of reef growth or possible antecedent Pleistocene control of the reef platform. These are non-numerical (i.e. symbolic) information that can be inferred only from groups of pixels and groups of classes. These clusters of pixels can be computed using a supervised, expert-driven, segment-based approach, automatic or manual. Indeed, classification algorithms based only on spectral (color) pixel-level information generally performs poorly (65-70% accuracy using Landsat TM or ETM+ for a 6-8 class classification scheme) (Andréfouët et al.). Such automatic processing typically provides noisy products because pixel-based approaches are sensitive to changes in water depth, to artifacts such as breaking wave patterns, and of course they are unable to discriminate classes with the same spectral properties (Bouvet et al. (2003)). They require many corrections that can be avoided by working directly at the segment-level. Therefore, classification algorithms alone are not able to provide the map products identified as objectives for this project.
- Expert-driven analysis is the most cost effective way to extract the mapping information we need. Since color/spectral information is not the primary factor for differentiating end-members, the fastest and most accurate way to proceed towards final maps consists of expert-driven segmentation by photo-interpretation, followed in some cases, by intra-segment spectral classification, and then contextual editing, until all the shallow lagoonal areas and reefs are assigned to the most appropriate class. With this scheme, a trained operator can process one scene containing a small atoll (<100 sq km) in about an hour, a large atoll or a barrier reef complex in 2-5 hours, and more complex, heterogeneous areas like some of the Caribbean reefs in one or two days.
- Expert-driven processing techniques are quite different from those currently recommended for other scales where spectral information is the primary factor of differentiation between classes. For our goals, at our scale of interest, shapes and topologies are more important. Therefore atmospheric correction (to retrieve accurate water leaving radiances, see Hu et al., 2001) or bathymetric correction (to enhance spectral discrimination) is generally not required. However, we intend to build the metadata-base required for image calibration to allow other users working on different projects to conduct specific local-scale analytical (e.g. using radiative transfer theory) or statistical (e.g classification) processing using remote sensing reflectance.
- In most cases, there is no suitable ground-truth data, or a poor description of reefs throughout entire reef systems, along hundreds of kilometers. Typically the literature provides detailed local description at a scale of a few tens of meters. Systematic ground-truthing during the course of this project is unfortunately not doable. Thus, most of the time, map accuracy needs to be assessed only using image-derived criteria since ancillary in situ data are scarce. The only reliable information to fix misclassifications is contextual. It means that the classes are primarily differentiated by their topology, not their color. Thus, it means that segment-level topological criteria (such as "outside", "along", "between", "in contact", etc.) are mandatory to both define the classes (see above) and assess the accuracy of the products. Topology is also of interest because topological errors are easier to identify than spectral errors.
- Automatic description and reconnaissance of intra-reef and inter-reef topologies would be very desirable. To date, algorithms derived from Artificial Intelligence techniques are suitable for this goal as demonstrated by pilot studies (Suzuki et al. 2001), but they still represent a substantial effort in research, tests and implementation that are not compatible with the deadlines of this project.