as photo clarity, it is likely that biological factors which determine the sharpness of the habitat boundaries, such as plant density and morphology, are factors which influence the minimum size that an interpretator can detect. During ground-truthing on shorelines at low tide, the area of habitats with well defined boundaries, such as seagrass beds, can be assessed with conventional surveying techniques but as mentioned above horizontal control can be a problem in remote areas. However DGPS can be used to delineate the perimeter of habitats, and algorithms are available to compute area. The position of the habitat can be georeferenced and placed on a base map. Accuracy can be improved by taking a large number of readings at single points around the perimeter of the habitat, but tracing the perimeter by walking around the boundaries was almost as accurate.
G. Error
There are several error issues that we have encountered in trying to quantify habitat boundaries and areas. Clearly defined boundaries between habitat types occur rarely in nature, and their occurrence is scale dependent. A boundary between water and adjacent land may be easily mapped at 1:250000 scale. But at 1:5000 scale, the visibility of details in the shoreline may make location of the boundary less certain, or may result in a more complex boundary being mapped. The problem is even more apparent for boundaries between vegetation types, where two communities of vegetation grade into each other. Identification and mapping of such a boundary will depend both on scale and on the classification system being used.
Identification of boundaries from remotely sensed images, such as air photos or casi multispectral data, is a complex and somewhat subjective process, as mentioned above. In the case of air photo interpretation, identification of features based on the tone, texture, and colour of the photo is a skill that takes some time to develop. Classification of multispectral imagery involves identification of training sites (areas on the image for which the habitat type is known, and that the computer can use as a basis for classifying the rest of the image) and selection of the classification algorithm.
Determination of the area of habitat from remotely sensed images depends partly on image resolution. The area cannot be calculated to a precision of one square metre if the resolution of the image is only 10 m2. Furthermore, habitat boundary generalization that results at smaller scales will decrease the precision of area calculations. The resolution of the image depends partly on scale, which depends on flying height. For example, 1:6000 scale air photos are obtained at a flying height of 6000 ft (1829 m), and can give a resolution of less than a metre. However, photos at this scale may not cover enough area to include recognizable features, such as roads and piers, which makes georeferencing more difficult and less accurate. casi data collected for us at Baynes Sound at a flying height of 6700 ft (2042 m) have pixels of four by four metres. Decreasing the pixel size to obtain greater resolution may increase the data set to an unmanageable size, and will also mean that a longer flying window (time when tides, light angles, and weather conditions are suitable for flying) is required to cover the study area.
Georeferencing is required both for the imagery itself and for the field data used to help classify the imagery. GPS data and aircraft roll, pitch, and yaw data can be collected at the same time the imagery is collected, and used in processing, along with existing basemaps, to georeference the imagery. Even after this processing, there may be some inaccuracy in the georeferencing (e.g. a few metres). Corrections for aircraft motion cannot completely remove this source of error, and DGPS accuracy is limited. Our equipment gives us an accuracy of 2 to 5 m (at best) for DGPS data collected to accompany field data; GPS receivers used to georeference casi imagery are more accurate (e.g. < 1 m).
H. Conclusions and Research
Recommendations
The use of technological advances such as those described in this paper will improve the effectiveness of coastal zone managers but there are still several areas where applied research and development are needed before they can be fully implemented. An urgent need is for the development of standards for field checking of photo/multispectral data classification and for collection of training-site data for multispectral data classification, especially
