CLASSIFICATION OF HISTORIC LAKES AND WETLANDS Golden Valley, Minnesota Image Analysis Heather Hegi & Kerry Ritterbusch 12/13/2010 Bassett Creek and Theodore Wirth Golf Course, 1947 FR 5262 Remote Sensing of Natural Resources and Environment
Objectives Heather, who works for the City of Golden Valley, Minnesota as a Geographic Information Systems (GIS) intern, inquired of her supervisor if there were any projects that would benefit from the use of remote sensing. Her supervisor expressed an interest for new shapefiles of historic wetlands in the city (a lake and stream shapefile). Golden Valley currently does not have very accurate shapefiles of their historic water features. Instead these shapefiles are a conglomeration of overlapping and inconsistent data obtained from multiple sources spanning several years (figure 1). Some of the data is derived from county plats which contain errors; for example, wetlands end at the edge of a section without continuing into the adjacent section on the following page. It is important for the city to be aware of the location of past wetlands for structure and road construction projects; the composition of the soil and terrain can greatly impact any building efforts. When infrastructure is built on top of these historic wetlands, the land on which it is constructed is likely to be prone to shifting and subsiding; this can necessitate excessive maintenance just to keep the infrastructure in working order. Golden Valley provided us with aerials taken from 1937 and 1945 for the retrieval of historic wetlands. Since these are simply black and white, one-band, panchromatic images, we were faced with the predicament of deriving water feature classifications without the usual multispectral imagery that is available today and which is used in this course. Although this presents complications, we welcomed the challenge and the opportunity to examine how others have tackled their studies and worked around this dilemma. 2 P age
Materials We obtained two sets of aerial photographs from the City of Golden Valley. These include a mosaic of six panchromatic images from 1937 (figure 2) and one panchromatic image from 1945 (figure 3). The cell size of the 1937 image is unknown, but appears to be about 3 by 3 meters. The 1945 image has a cell size of 1.161185 by 1.1644701 meters. We also obtained a May, 2009 multispectral, three band (RGB) aerial photograph (figure 4) with a 0.5 by 0.5 meter cell size from Hennepin County. In addition to the aerial imagery, we gathered a shapefile of the city boundary from the Minnesota Department of Transportation, a high resolution elevation model from the Minnesota DNR, and a shapefile of the current lakes from the City of Golden Valley. Procedures Preparing the data We have six raster images from 1937 of different locations covering the area of Golden Valley. The original tiles contain black boarders, but we clipped them using the Clip Raster tool so they would not affect the analysis. Then, using the Mosaic To New Raster tool in ArcMap s ArcToolbox, we were able to merge the images to make one continuous raster layer. The 1945 raster image that we had obtained is already one continuous image, so we did not have to use the mosaic tool. We obtained Digital Elevation Model (DEM) data (figure 5) from the Data Deli as well as municipal boundaries. The DEMs were merged and clipped to fit the city boundary more concisely. This data will help in error checking lower elevations are more likely to hold lakes and wetlands. 3 P age
Extraction of Water Features We first conducted an unsupervised classification of the 1937 image (figure 6), but as we thought, classification of black and white photos does not work. Our alternative was to digitize the water features to create a new layer. We are more familiar with ArcMap, so we used the digitization tools from this program, as opposed to working with Erdas. We created a new shapefile called Water Features and with the editor tool, we drew boundaries around the wetlands and lakes. Because 1945 was a wetter year, it was much easier to see distinct boundaries, as the features in the 1937 image were a bit more muddled. We are still using the 1937 image to confirm the location of water bodies. If there is standing water in 1945 and not in 1937, then we are designating these areas as wetlands. If features in both 1937 and 1945 appear to be water, we are designating them as lakes. Classification To illustrate the current locations of the lakes, unsupervised classification on the 2009 aerial photo was performed with a mix of successes and failures. We had many options to choose from for the source of the imagery as well as the date for this classification, but settled on a 2009 aerial (figure 4) from Hennepin County since the lakes in this aerial are very consistent in their color and placidity. We kept our file in its original file format, MrSID, performing seven iterations for seven classes. After the software ran the iterations, we converted the file from a.sid to a.img file and tried the classification again. We left the classes at seven, but increased the iterations to twelve as increasing the iterations and classes can increase your chances of delineating minute differences between the classes. The resultant image (figure 7) was speckled with wrongly classified pixels due to shadows. We preformed one more unsupervised classification using twenty-two classes (figure 8) hoping it would separate the shadows from the 4 P age
lakes, but it only divided up the lakes more. We decided to proceed with using the classification with seven classes. We brought this image into ArcScan to clean up the excess pixels caused by the shadows. Our results were not usable because we had trouble converting the raster to vector polygons in ArcMap. The best option we had was to use the current lake shapefiles from Golden Valley. Change Detection We wanted to analyze the changes in lake surface area from 1945 to 2009. Using the Intersect tool to overlap the current and historic wetland shapefiles, we were able to establish a polygon layer that showed the lakes that remained throughout the years. The Erase tool was utilized to present two polygon layers showing where there have been changes to the lake distribution. The tool was utilized twice. First, we used the 1945 lakes shapefile and erased locations in which the 2009 overlapped with the 1945 shapefile; this left us with a polygon shapefile displaying locations in which lakes were present in 1945 but are no longer present in 2009. Secondly, we used the 2009 lakes shapefile and erased locations in which the 1945 overlapped with the 2009 shapefile; this left us with a polygon shapefile where lakes are present in 2009 but were not present in 1945. Results Once the change detection was accomplished, we obtained the lake coverage in square feet and square miles from 1945 and 2009 using the statistics feature in ArcMap. This allowed us to mathematically illustrate surface area changes. The statistics derived from this information can be found in table 1 in the appendix. We also created two maps to illustrate the changes in lake surface area distribution. The first map (figure 9) displays the distribution of the lakes in 5 P age
1945 and 2009, while the second map (figure 10) focuses more in-depth on the exact changes that have occurred throughout the years. Discussion Classification of multispectral images can be a fairly easy way of delineating the electromagnetic radiation feedback to determine various land types. However, this becomes complicated when the only images available are historical black and white, or panchromatic, photographs. We ran an unsupervised classification process on the black and white images, but as we assumed, the results were not be usable: Objects such as trees would register with the same tonal qualities as lakes and crops. There was not enough distinction between the tones in the black and white images. In the article, Quantifying historical changes in habitat availability for endangered species: use of pixel- and object-based remote sensing, Pringle, Syfert, Webb, and Shine explain: One difficulty associated with semi-automated analysis of historical photographs, however, is that these images contain limited information typically a single, panchromatic spectral band. Traditional methods of analysing such images assume that pixels in the same land-cover class are spectrally similar. This method is sub-optimal for several reasons. Even in relatively simple landscapes, individual land-cover classes (e.g. forest ) may comprise a broad range of pixel spectral values, which may overlap with the ranges of other land-cover classes. (Pringle et al., 2009, p. 545) There are several factors that can alter how an image is classified: the classification of water features depends on the turbidity, color, and placidity or roughness of the surface due to factors such as disturbed sediment, pollution, aquatic flora, wind and water speed. If one lake is clear while another is murky, the tonal qualities will read much differently. Furthermore, if a river is flowing quickly and roughly over rocky riverbeds or large gusts of wind are producing white caps, the reflectance will be much more diffuse rather than the specular response that would be 6 P age
received by a river or lake that is calm and tranquil. To further complicate things, the angle of the sun affects the specular reading, an example of which can be seen in the 1937 aerial (figure 11). At certain angles, lakes can appear to be silver/white or mirror-like, or the darkness of the water can absorb the radiation and the water appears to be a dark blue/black. With rough flowing waters, the water has a mix of dark and light values. Our alternative was to perform heads-up digitization of the lakes and wetlands and turn them into a new shapefile; the boundaries of the water features were traced using the 1945 image, due to its clarity and fine resolution, while the 1937 image was used for elucidation. We felt vindicated in our decision to digitize, as a study conducted by Labrecque, Lacelle, Duguay, Lauriol, and Hawkings also employed the digitization of lake features. Labrecque et al. quantified lake distribution changes in the Old Crow Basin of the Northern Yukon in Canada from 1951 to 2001. They used historical black and white air photographs for analysis of the lakes in 1951 and 1972, and panchromatic Landsat ETM+ images for the 2001 analysis. Since there are over 2,000 lakes within their area of study, they selected 300 lakes from a random set of geographic coordinates to represent the entire area. With their digitized polygon features, they were able to determine surface area changes from each year of study. Digitization can become a fairly time consuming process, but our area of interest was of a large enough scale (the entire city fits a 1:30,000 ratio; 4 x 3.5 miles) and the number of water features was manageable, unlike the over 2,000 lakes in the Old Crow Basin lake distribution study. We wanted to have the opportunity to work with multispectral data, so we decided to conduct change detection analysis by classifying the image from 2009, extracting the water features, and comparing the current lake surface area to the historical lake distribution. As mentioned previously, we were not successful in classifying and extracting the lakes from the 7 P age
2009 image. We gathered several images, but each presented problems. Landsat 5 data from November 11, 2010 was obtained so we could have a satellite image that included an IR band. However, for the scale we needed, the resolution was too course and had too many mixed pixels (figure 12). We also downloaded a 2010 National Agriculture Imagery Program (NAIP) aerial (figure 13), but this was only used with RGB bands (no IR band), and each lake was a distinctly different color. We considered conducting a supervised classification, but this process would have been extensive, time consuming, and out of the scope of our project, as we would not be able to simply train for just the lakes, but instead, we would have had to train for every land type. We also discussed using object oriented classification, but decided this would not work due to the variances in lake shapes and sizes. We assumed at the beginning of this project that there would be a loss in lake area totals due to development, so it was surprising when we discovered that total lake area has increased. Even though development has forced the drainage of many lakes (figure 14), the city has created several more lakes in areas where lakes did not exist previously. We believe this may be due to the presence of more wetlands and marshes in 1945 which are not as visible on aerial images as lakes. These wetlands were drained and pooled into the lakes we see in the 2009 imagery to make way for development. This resulted in fewer wetlands and an increase in lakes. Conclusion Although we were not successful in deriving a polygon shapefile from the classification of the multispectral image, we have gained a greater understanding of classification methods, and have learned how to work around problems that arise when working with historical data. We researched various ways in which other analysts have dealt with the difficulties and obstacles we encountered, and used their findings to guide our work. There are several approaches that we 8 P age
could have taken in retrieving and creating the water feature polygons, but we ultimately found digitization to be our best option in dealing with historic, black and white imagery. 9 P age
References City of Golden Valley. (Publisher). (1937). 1937 aerial. [Photograph]. City of Golden Valley. (Publisher). (1945). 1945 aerial. [Photograph]. Hennepin County. (Publisher). (2009). 2009 aerial. [Photograph]. Labrecque, S., Lacelle, D., Duguay, C. R., Lauriol, B., & Hawkings, J. (2009). Contemporary (1951-2001) evolution of lakes in the Old Crow Basin, northern Yukon, Canada: Remote sensing, numerical modeling, and stable isotope analysis. Arctic, 62(2), 225-238. Minnesota DNR MIS Bureau. (Distributor). (2006). High Resolution Elevation Model (DEM). [Data file]. Retrieved from http://deli.dnr.state.mn.us/metadata.html?id=l390005620606 Mn/DOT. (Distributor). (2001). City Boundary [Data file]. Retrieved from http://deli.dnr.state. mn.us/metadata.html?id=l390001310201 Pringle, R. M., Syfert, M., Webb, J. K., & Shine, R. (2009). Quantifying historical changes in habitat availability for endangered species: Use of pixel- and object-based remote sensing. Journal of Applied Ecology, 46(3), 544-553. 10 P age
Appendix: Figure 1 Golden Valley Lake Distribution Golden Valley Old Shapefile Lagoon Lowland Historic Wetlands Description Drainage Wetland Meadow Dry Meadow Pool Dry Marsh Unknown Water Creeks & Ditches Sources: Historic Wetlands - The City of Golden Valley City Boundary - MnDNR Data Deli I 0 0.125 0.25 0.5 0.75 1 Mile
Figure 2: 1937 Aerial Figure 3: 1945 Aerial
Figure 4: 2009 Aerial Figure 5: DEM & Digitized Lakes and Streams
Figure 6: Unsupervised Classification on the 1937 Aerial with a few classes colored Figure 7: Unsupervised Classification on the 2009 Aerial with 7 classes
Figure 8: Unsupervised Classification on the 2009 Aerial with 22 classes
Figure 9 Golden Valley Lake Distribution 1945-2009 Lake Distribution 1945 Lakes 2009 Lakes No Change; 1945-2009 Golden Valley I 0 0.125 0.25 0.5 0.75 1 Mile Sources: 1945 Lakes - Derived from Aerial provided by the City of Golden Valley 2009 Lakes - City of Golden Valley City Boundary - MnDNR Data Deli
Figure 10 Lake Change Distribution Golden Valley Lakes Present in 1945; no longer present in 2009 Change in Lake Distribution 1945-2009 Lakes Present in 2009; not present in 1945 No Change; 1945-2009 Golden Valley I 0 0.125 0.25 0.5 0.75 1 Mile Sources: 2009 Aerial Photo - Hennepin County 1945 Lakes - Derived from Aerial provided by the City of Golden Valley 2009 Lakes - City of Golden Valley City Boundary - MnDNR Data Deli
Figure 11: Lakes in 1937 Aerial Figure 12: 2010 Landsat Imagery
Figure 13: 2010 NAIP Imagery
Figure 14: Neighborhood were lakes used to be
Table 1: Area Square Feet Square Miles 1945 Lake Area 13078361 0.469 2009 Lake Area 16357942 0.587 Area of Lakes Present in 1945; no longer present in 2009 4396164 0.158 Area of Lakes Present in 2009; not present in 1945 7761675 0.278 Area of Change (sum of the 2 preceding categories) 12157839 0.436 Area of No Change; 1945 2009 8682197 0.311 Total Lake Area Existing & Historic (sum of the 2 preceding categories) 20840036 0.748 Statistics 1945 Specific Statistics Percentage (Area of Lakes Present in 1945; no longer present in 2009) / (1945) 33.61% (Area of No Change; 1945 2009) / (1945) 66.39% Totals: 100.00% 2009 Specific Statistics (Area of Lakes Present in 2009; not present in 1945) / (2009) 47.45% (Area of No Change; 1945 2009) / (2009) 53.08% Totals: 100.53% Description Percentage of Lakes Present in 1945 that are no longer present in 2009 Percentage of the Lakes in 1945 that did not change Percentage of Area of Lakes Present in 2009 that were not present in 1945 Percentage of the Lakes in 2009 that did not change (Some error must have affected this result) Overall Statistics (2009 1945) / (2009) 20.05% Increase in Lake area from 1945 to 2009 (Area of No Change; 1945 2009) / (Total Lake Area Existing & Historic) 41.66% Percentage of the Area with No Change (Area of Change) / (Total Lake Area Existing & Historic) 58.34% Total Change in Lake Distribution Totals: 100.00%