This preview shows page 163 - 164 out of 355 pages.
Image classification methodsPhoto-interpretation is a method by which the user visu-ally inspects an image and identifies habitats. Multi-spectral classification is a computer-based method bywhich a number of wavebands are input to a statisticalclustering algorithm which organises individual pixelsinto distinctive groups (classes) which are assumed torepresent specific habitat types. Each method will be dis-cussed in turn.Photo-interpretation (visual interpretation)Visual interpretation is the process whereby an image isstudied and habitats identified by eye. The image may bea photographic print or colour hard copy of a satelliteimage bought from a data supplier or digital data viewedon a computer monitor. Digital imagery is advantageous inthat it may be processed (e.g. contrast stretched) to empha-sise the habitats of interest. For example, land can bemasked out of an image and the full dynamic range of thedisplay device applied only to the aquatic habitats. Imageinterpretation should be carried out using either a range ofindividual wavebands or an optimally-combined colourcomposite. For the latter case it will be necessary to choosethe most appropriate wavebands for the habitats in ques-tion, e.g. Landsat Thematic Mapper bands 3, 2, and 1 as anRGB colour composite for submerged habitats.Habitats are identified visually by an analyst based onexperience, then delineated and labelled manually. Deci-sions on the boundaries between different habitats aremade with reference to local knowledge, common sense,field data, aerial photographs or maps and a sound under-standing of remote sensing theory. In essence the analystedits and classifies the image in context (see Box 10.1).These decisions are reached without any computationaleffort or statistical processing. Delineation of boundariesis achieved using a tracing overlay and pencil or a mouse-driven cursor on a computer screen. Whether using a pencil or mouse-driven cursor, photo-interpretation requires a skilled and experienced operatorto make sound decisions. Figure 10.1 (Plate 9) illustratesthe challenge that visual interpretation presents: it is aLandsat TM image (bands 1, 2 and 3) of a 12 x 10 km areaof uniform depth with patch reefs to the west, and areas ofdense algae to the east. Some of the boundaries betweenhabitats are clear, because the contrast with adjacent habi-tats is high, and should be relatively straight-forward to delineate (e.g. boundaries around A and B).Area C is a seagrass bed – its eastern boundary (whiteline) is clear but to the west the seagrass grades into sand.There is no clear boundary and the analyst must decidewhere the boundary lies (red line). Ideally, supplementarydata will be available to guide this decision but typicallythis is not the case and the analyst is obliged to rely oncommon sense and local knowledge. Similarly it wouldbe difficult to draw a boundary around areas D and Ealthough their colour and texture would seem to indicatedifferent habitats. The south-eastern boundary of F is