3.2 Hydrologic Land Cover Classification

3.2.1 Conceptual Approach

The derivation of a land cover map specifically for the purpose of hydrologic studies in an urbanising area requires that the land cover representation be selected in a fashion that is specific to this application. The land cover classes must be selected on the basis of their representative hydrologic characteristics, rather than more commonly selected groups of classes in typical land use or land cover mapping exercises.

Deriving this application specific land cover mapping involved specifying the hydrologic land classes (HLCs) that would be used to represent the northeast Humber Watersheds's cover. The selection of these classes was based upon the dual criteria of what sorts of information the available source data supported, balanced against the sorts of information classes that are deemed to be meaningfully different in a hydrologic performance sense.

Through a field excursion to the study site, a useful set of HLCs were selected. This was further refined using airphoto interpretation. Two obvious land covers found in the less developed portions of the watershed are forested land and agricultural land. The portions of the watershed with greater impermeability were defined as the Developed Lands HLC, a land cover that was further subdivided, as described below. Two additional land cover classes with meaningfully different performance characteristics that could be reasonably located by our data sources are large bodies of open water and rural roads: lakes clearly respond differently in the runoff generation process, and rural roads represent a certain amount of impervious surface that are abundant enough that they should be distinguished from surrounding forests and fields.

The information sources available describing the northeast Humber Watershed were plentiful, providing many perspectives on the watershed. The different ways to describe the landscape are useful, since no singular description could adequately provide the information needed to arrive at a realistic hydrologic land cover classification.

The core descriptions of the watershed are derived from remote sensed products: Thematic Mapper multispectral imagery provided the basic means of describing the watershed. The TM spatial data model's 25 metre raster grid was used as the fundamental representation unit to which all other maps were matched. (Figure 9) The TM imagery was supplemented by an excellent supervised land cover classification of the imagery, provided by the MTRCA (Figure 10). This classification was extensively processed, providing a smoothed description of land cover, rather than the `salt-and-pepper' result associated with many multispectral classifications. Other remotely sensed data sources included two types of airphotos. Black and white 1:800 airphotos are used for more precise information on developed lands' imperviousness. High resolution digital colour airphotos were used to check the accuracy of classification results through visual comparison. A further remotely sensed data source was the MTRCA's forest polygons map (Figure 8), which was so accurate that it could only have been derived from some sort of imagery.

Other spatial descriptions of the watershed included the MTRCA's zoning derived land use map (Figure 11), and an extensive road network from Statistics Canada's Street Network Files (Figure 7). The land use map provided a less physically-based description of the landscape describes both existing and future land uses. The road network was quite useful in helping to locate certain positions based on intersections, as well as providing a thorough description of rural roads. This map layer was also used to locate areas of high road density, which are indicative of developed areas.

Integrating the several layers of information into the best possible description of the hydrologic land cover required a somewhat novel approach. Each information source was evaluated in terms of how well it could spatially define a given HLC. This evaluation was conducted by comparing source maps both visually and through coincidence table calculations (via GRASS's r.coin facility). The optimal combination of data sources to describe a given HLC are based upon an inspection of the study area, and the expectations about a data source, given the origins of its information.

Naturally, an approach that describes land cover classes using different combinations of source maps for different classes will result in some areas of overlap. Based upon the chosen map arithmetic, a pixel could be defined as being a member of two different classes. At this stage, decision rule based sorting is required. Areas of coincidence must be examined visually and through coincidence table calculations. Decisions upon which class should take precedence are made based upon the consequences of misclassification and the class most likely to be correct, given the source data's origins. By moving from the most accurately defined class to the least, this procedure was streamlined (Figure 12). The way in which this procedure actually operated is described in specific instances below.

The forested lands HLC was defined primarily based upon the MTRCA forest polygon data. Comparison of those forest polygons with both the Thematic Mapper imagery and airphotos showed these polygons to be an extremely accurate representation of forest locations. The forest polygons were compared with the locations of forested classes in the supervised land cover classification. These two portrayals of forest locations agreed to a great extent, with the exception of the supervised classification's sparse forests class. The sparse forests class included very small patches of forest, perhaps too small to be included in the polygons in many cases. The resulting union of the two was used for the forested HLC (Figure 13). The extreme accuracy of this class led to the choice that it take precedence over other classes in deciding the identity of multiply defined pixels (with the exception of open water).

Defining the developed class was less straight-forward. For the northeast Humber Watershed, developed area is primarily represented by residential development. This is supported by Table 2, planning documents and field observations. As a means of spatially defining development, the land cover's urban areas classes are clearly an incomplete description. By definition, the zoning based land use extended beyond existing development (based upon comparison with imagery and airphotos).

The map layers describing land use and land cover share similar categorical descriptions. A coincident table using r.coin in GRASS 4.1, is used to determine if one, or both of the data are necessary to include in the overlay analysis in determining land use classification. If there is a high mutual occurrence of categories for the two map layers, than only one may be used. However, if the categories shared by each map layer are not explained by the one another, than both layers need to be used and additional decision rules need to be applied to isolate the existing residential land uses.

A further description of developed areas is derived from the road network. Using a kernel to count the number of road pixels in a given vicinity, road density maps are created. A variety of kernel sizes are applied to try to ascertain the appropriate scale that best differentiates a low road density from the critical road density which delineates developed areas (Figure 14). A 27x27 kernel with a threshold value of 68 pixels per 729 (the number of pixels in the kernel) best describes the developed areas. From the coincidence of these three descriptions a strategy is selected. The three realizations agreed well for the most part, lending confidence to the choice to take the union of the three (Figure 15). This class has decision precedence only over rural roads and agriculture (see below).

The open water class is defined in a straightforward manner, simply taking the lakes to be located by the land cover deep water class. This class did not always produce seamless boundaries with adjacent classes, but some editing could remediate this problem. Water bodies are considered to be very well located by the spectral data source, and took precedence over all other classes (Figure 16).

The rural roads class is defined by assuming any part of the road network not within the developed HLC is a rural road. In this way, the impervious surfaces of these roads are accounted for in the classification, rather than being lumped in with other non-developed classes (Figure 17).

The final class, agricultural land, is thought of as the default land use for the watershed. That is, any pixel not assigned to the other four classes is assigned to agriculture. While agriculture and pasture is the most common land cover in the watershed, its boundaries are most poorly defined by the data sources. Attempts to use appropriate classes from the land cover and land use maps produced poor results, leading to the selection of this remainder approach. In this sense, the agricultural HLC is best defined in terms of the absence of the other well defined HLCs (Figure 18).