Methods

Delineation of Study Region - The Minnesota portion of Lake Superior's shoreline is about 240 km in length excluding bays and points (Figure 1). Since most of the nearshore region has never even been sounded in detail, the exact size of the area adjacent to shore less than 30 meters in depth was unknown prior to surveys. However, cursory estimates based on available maps indicated the width of the depth zone between five meters (shallowest depths we could record with the survey vessel) and 30 meters varied between 0.3 kilometers and 2 kilometers. Using an average width of 0.75 kilometers we estimated an approximate 180 kilometers² of lake bottom at the appropriate depth. Due to the logistics of boat speed, access for boat launching or harboring, and weather complications, we established a series of priority regions for new surveys based upon historical lake trout spawning areas and other management concerns. In 1995 we ran data collection transects parallel to shore. The spacing of the transects was about 50 meters. Starting in 1996, transects were run perpendicular to the shore with a spacing of 60 meters. Figure 2 shows where data were collected. Approximately 65 kilometers² of the nearshore area were mapped.

Instrumentation - In order to map a large portion of the potential study area we needed an efficient method that would work well in shallow water near shore. We decided to use a sonar-based system manufactured by Marine Micro Systems, named RoxAnn^TM, which uses the return signal of an echo sounder to characterize the lake bottom. Along with the RoxAnn^TM unit an Innerspace (model 448) echo sounder was used and a real-time differential GPS receiver system allowed accurate navigation and positioning on the order of one meter. Depths were adjusted to the International Great Lakes Datum of 1985 using water level gauges in Duluth and Grand Marais. Coastal Oceanographic's Hypack software was used for data collection and navigation. With this system, data were collected at a rate of two samples per second and at speeds of up to 20 km/hr, although navigation safety concerns along the shore and rough weather generally resulted in slower speeds.

During the first year of data collection in 1995, the equipment was mounted on a 57 foot boat (Figure 3). Since the U.S. Coast Guard GPS beacons were not operating in the survey area, we established a real-time differential GPS system. The system consisted of two Motorola LGT1000 GPS receivers and a UHF radio link which provided real-time accuracy of about 2 meters for navigation. In addition, carrier phase GPS data were collected and post processed to allow sub decimeter position accuracy to be obtained. While this level of accuracy was not required for horizontal positioning, it did allow us to verify that water level gauge data from Duluth and Grand Marais could be used with sufficient accuracy for the whole shoreline.

The 57 foot boat was large enough for the crew to sleep on board and it was able to handle the rough weather, but the disadvantages were that it was difficult to maneuver in and out of bays, required a larger crew, was slower than a smaller boat, and more expensive to operate. Consequently, in 1996 we mounted all the equipment on a 25 foot boat (Figure 4). The Coast Guard also began operating beacons in Duluth and on the Keweenaw Peninsula in Upper Michigan. Using a NovAtel 12 channel GPS receiver and a Magellan beacon receiver, the two U.S. Coast Guard beacons provided approximately one meter positioning accuracy for all of our area of interest.

Substrate Classification - The RoxAnn^TM system is wired in parallel to an echo sounder and it processes the first and second echo returns to create two values, called E1 and E2, which correspond to the bottom's roughness and hardness respectively (Figure 5). These values, which range from 0 to 4.1, can be plotted to form a RoxAnn^TM square and from this information the bottom type can be inferred (Figure 6). However, a RoxAnn^TM square did not exist for substrates and conditions encountered in the Great Lakes and this technique has no statistical properties that allow the user to understand the confidence with which substrate classifications are made. Consequently, we created a customized RoxAnn^TM square for our project and developed a technique for describing the variation inherent in the substrate classifications (Bonde et al., 1998, Yin et al., 1998).

RoxAnn^TM data were processed and interpreted using similar methods to any remotely sensed data. Samples of RoxAnn^TM readings were collected over areas of known bottom type. Multiple samples of each substrate type of interest were combined to create a RoxAnn^TM signature for that bottom type. As an example, Figure 7 shows several hundred actual RoxAnn^TM values collected over a sand and cobble (video) area.

An underwater video camera was used to collect ground truth information. The camera was lowered overboard and as long as the bottom remained fairly uniform, RoxAnn^TM data were collected for several minutes as we drifted over the area. Each data set then consisted of several hundred E1/E2 pairs. More than 100 of these data sets were acquired, but because of variability in the bottom only 49 were used in the final analysis.

From our underwater video work we identified 8 categories of substrate that RoxAnn^TM could distinguish. These categories are sand(video), sand and cobble (video), sand over bedrock (video), smooth bedrock (video), rough bedrock (video), conglomerate (video), cobble (video), and boulders (video). This terminology follows that of Edsall et al. (1992). While we did find sand and gravel (video) areas with the video camera, RoxAnn^TM could not reliably distinguish this category from the others. Also, while RoxAnn^TM seemed like it might be able to distinguish among different sized fine materials such as sand, silt, and clay, we combined them all into our sand category since we were most interested in harder bottom types.

The RoxAnn^TM signatures of the eight categories defined above are shown in Figure 8. Each ellipse center is defined by the category's mean E1 and E2 values and the axes by the standard deviation of E1 and E2. The signatures of several categories overlap which isn't surprising as these categories often blend into one another. The final RoxAnn^TM square we used is shown in Figure 9. Areas of overlap were assigned to the most likely category using a maximum likelihood classifier. Also, an ellipse region based on two standard deviations from the mean was added. The unknown area was further divided into several sections. The resulting square can be used to create bottom maps with various levels of confidence in the classification. When higher confidence levels are used in map creation, greater amounts of area are classified as unknown. For this map series we minimized the amount of unclassified area.

Map Creation - We illustrate the method of map creation from the RoxAnn^TM data with a three kilometer stretch of shoreline about 17 kilometers northeast of Two Harbors, Minnesota, known as Castle Danger Figure 10. Transects were run perpendicular to shore with an approximate spacing of 60 meters. With a sample rate of two per second and a maximum speed of 20 km/hr, readings were taken at least every three meters along a transect, usually more often. The position of individual RoxAnn^TM readings taken in the Castle Danger area is shown in Figure 11. The readings were dense enough to show the boat track well. When zoomed in, as in Figure 12, the point defining each reading can be seen.

The RoxAnn^TM data, represented as points, were first classified using the RoxAnn^TM square shown in Figure 9. The point data were then converted into polygon data by creating Thiessen polygons (Gold 1991). These are created by growing a polygon around a point. The polygon continues to grow until it runs into another growing polygon. The borders between polygons with the same classification are removed and the resulting polygon map is shown in Figure 13. The individual categories of this map are shown in Figure 14. The map is then smoothed and filtered to create Figure 15. Depth information from the echo sounder was used to create 10 foot contours for all areas surveyed. To show depth without cluttering the image, twenty foot depth contours are overlain as shown in Figure 16. Ten foot contours are included in the GIS data files.

HOME