Twin Cities Urban Lakes Project
Ann Krogman
FR 5262 Section 001
Objectives
In 2002 the Sterner Lab was interested in investigating the biodiversity of lakes in the urban environment at a landscape level. Fortunately, the Twin Cities in Minnesota were an ideal area to study urban lakes as the 7-county metro area is dotted with lakes as small as storm water holding ponds to as large as Lake Minnetonka. It was hypothesized that land-use surrounding the lakes would impact biodiversity, specifically; increases in impervious surface would yield lower biodiversity. To conduct the study, 100 lakes within the 7-county Twin Cities Metro Area were selected via random stratified sampling. The metro area was broken into three concentric circles. Twenty-five lakes were selected from the inner circle or urban ring, twenty-five lakes were selected from the second, or suburban ring, and finally fifty lakes were selected from the outermost, or non-urban ring. The lakes were located by generating XY coordinates and then going into the field to find the water body nearest to the set of XY coordinates. Selected lakes ranged in size from .003 to 5,667 hectares. Along with sampling the phyto and zooplankton within each of the lakes, water chemistry parameters such as phosphorus, nitrogen, dissolved oxygen, dissolved organic carbon, and chlorophyll a were also sampled.
In order to determine the land-use surrounding each of the lakes,the Sterner lab worked with Dr. Fei Yuan who was a student of Dr. Marv Bauer in 2002. Dr. Yuan classified the area surrounding each of the lakes by finding the area of lake and determining the radius of a circle of equal area to the lake. She then classified land-use around each lake in 0.5, 1, 2, and 3 radius buffers. Land-use was classified into 12 categories: High intensity urban, low intensity urban, transportation, crops, grass, conifer, deciduous, water meadow, low shrubs, wetland, and impervious.
I came into the lab to work on the Twin Cities Urban Lakes Project as the biodiversity project was called in the fall of 2009. In the spring of 2009 my advisor and I worked together to analyze the water chemistry data. We used data mining in Statistica to create regression trees in Statistica in order to determine what land-use type had the most impact on different water quality parameters. Through this analysis we discovered that percent crops had the biggest impact on total phosphorus. High percent crop meant high phosphorus and low percent crop meant lower phosphorus. After percent crop, the next most important land-use on total phosphorus was lawn. Lawn greater than 49.9 percent meant higher phosphorus and lawn less than 49.9 percent meant lower phosphorus.
My advisor and I found these results to be very interesting and relevant. In 2004, the state of Minnesota banned the use of residential phosphorus based fertilizers in the 7-county Twin Cities Metro Area (TCMA). With this ban in place, we hypothesized that lawn should no longer be the most important land-use controlling total phosphorus concentrations in urban lakes with low crops. This hypothesis was generated because we anticipated that without excess phosphorus being applied to urban lawns, runoff from lawns into urban lakes would not be as significant a contributor of phosphorus.
Looking at phosphorus in the urban system is very important. In freshwater systems, phosphorus is not only an essential nutrient for plant growth, it is often the limiting nutrient of the system (Schindler, 1977; Lee et al, 1978). If phosphorus is the limiting nutrient, when more phosphorus is added to freshwater systems, increased algal and vegetative growth is expected. These excess blooms have been documented in numerous studies (Carpenter, 1998; Findlay and Kasian, 1987). This eutrophication can result in anoxic conditions, fish kills, and reduced recreational value. In the TCMA alone, 163 lakes are listed on the Clean Water Act 303d list for impairment due to excess nutrients/eutrophication (MPCA, 2010). The impacts of eutrophication are an estimated $2.2 billion annually in the United States (Dodds et al, 2009).
Urban systems are considered a major contributor of phosphorus to surface water bodies (Osborne and Wiley, 1986). The urban environment is very heterogeneous. The varying land cover classes transport large quantities of phosphorus to receiving water bodies via surface runoff and sediment transport to surface waters by storm water removal systems (Walsh et al, 2005). The natural flow pathways which are altered by impervious surfaces and man-made sewers increase the speed with which water and nutrientsare removed from the landscape and transported to receiving water bodies.
To test our hypothesis, during the summer of 2010, I re-sampled the 64 lakes which were classified as low crop in 2002. I followed the 2002 procedures and collected the same water chemistry parameters. In order to re-run our regression trees, the land-use surrounding the lakes needs to be updated for 2010 as well.
The goals of my project were to initially update the land-use in the .5, 1, 2, and 3r buffers surrounding the 100 lakes in our study. I intended to classify the percent land-use type and make a chart similar to the one produced by Dr. Yuan in the appendix. I also wanted to make the classification as accurate as the one done by Dr. Yuan in 2002. Her total percent accuracy was 93.3 and she had a kappa statistic of 91.6. After discussion with the lab TA, the goals of the study were refined to include classification of the 3r buffer only surrounding the fourteen lakes in our study located in Anoka County. The lakes in this subset ranged in size from .003 ha to 94.7 ha.
Materials
For this project I used four images. Two of the images were 30 meter resolution, 7 band, LandSat 5 TM images. These images were obtained from and these images were from July 2002 and July 2010. They are from path 27 and row 29. There was no cloud cover in the July 2002 image but there was 14% cloud cover present in the July 2010 image. This cloud cover did not block out any important portions of the image but it did make the image overall hazier by reducing the digital number values as compared to the 2002 image. The other two images were obtained from and the images were obtained by the National Agricultural Imagery Program (NAIP). They are 1 meter resolution with red, green and blue bands. The two images are from summer 2003 and summer 2010.
In addition to these images the guidance for this study was found in Yuan et al 2005 from Remote Sensing of the Environment. Data was made available from the Sterner lab existing data set for the Twin Cities Urban Lakes Project. Google Earth was also used to locate lakes on the images. All of the analysis for the project was done using Erdas Imagine 2010 and ArcMap in ArcGIS 9.3.1.
Procedures
When I started the project, I believed that the methods for the entire updated classification would be located in the Yuan et al, 2005 paper. After beginning the project, however, I realized that the paper was the methodology for a land-use classification of the entire 7-county Twin Cities Metro Area. There were no specific instructions or documentation on how the full TCMA classification became the percent land-use classifications found in the chart in the lab data. I had scheduled a meeting to talk with Dr. Yuan for October 29, 2010. Unfortunately, the death of my uncle on October 26, 2010 necessitated that I reschedule my meeting with her. I was not able to meet with her again until December 10, 2010. At the meeting she did not remember many specific details about the classification because it was six years ago. Therefore, I had to develop the protocol for classifying the areas in the 3r buffer around the lake.
My initial plan of action was to subset all images to include only the areas within the 3 radius buffer zone around each lake, classify all images using unsupervised classification, determine percent accuracy for all images using NAIP data for reference, compare the percent accuracies between the 30 m and 1 m resolution, and do change detection between the 30 m 2002 and 2010 images. I needed to classify all of the images prior to comparison between images in order to account for the 14% haze over the July 2010 LandSat image. I also wanted “to” and “from” changes so thematic change detection made sense. The goal of doing change detection and comparing accuracy among the 30 meter and 1 meter resolution images was to get at the differences in land-cover which were due to changes in land-use and the difference which were due to classification error.
I chose to make some changes to the protocol used by Yuan et al. Major differences included using primarily 1 meter data instead of 30 meter data because of the small size of many of the lakes. The minimum mapping unit of 30 meters was too big to accurately describe the land surrounding the lakes. I also decided to classify into only five classes. In Yuan et al, land-use classification is broken into 7 classes and in the lab data file 12 classes are present. I felt that based on ground referencing the sites and the needs of the analysis that lawn, impervious, trees, wetland, and water are the only classes necessary.
To begin with I added the XY coordinates of the center points of each lake in ArcMap. I then buffer each center point by 3 times the radius of the lake. The buffer distances used by Dr. Yuan were obtained from the lab excel file. Then I merged all of the buffered files. I added the merged buffer file in Erdas and copied them to an area of interest (AOI) layer and extracted them from the .sid (NAIP 1-meter) county file.
Unfortunately, this did not work for two primary reasons. The first was that the center points from the files were not at the center points of the lakes. Additionally and even more importantly, when the AOI file was extracted from the .sid none of the .sid file went with it so I just had empty circles and no raster after the extract.
To deal with these issues, I loaded all of the county rasters into Erdas in different viewers. I added an inquire cursor to the 2010 NAIP image and linked all views. I added an inquire box and used Google Earth to locate all of the lakes. For each lake I recorded the center XY coordinates in meters. I then used the subset feature to cut out a box about five times the size of each lake in each view. The coordinates of the inquire box were used for each subset. The subsets were then opened in ArcMap. In ArcMap, I used the extract by circle feature with the new center point coordinates and existing buffer radius to extract the area of interest for each lake. I did this for all the images of the same lake in order so that all subsetted images were loaded in the viewer simultaneously. This allowed me to overlay the images and check for geometric correctness. The images overlaid well. After creating circular subsets for each of the lakes in each of the images I attempted to use the MosaicPro from 2D feature in Erdas to put the images back together. This worked for the 30m resolution images but did not work for the 1m resolution images. The 1m resolution imagery was too large to save and took about four hours to process. I mapped a new network drive on the lab computer and reran the merge of 1 meter data. After spending the time doing the merge, it turned out the 1 meter merged images were too large to work with, the unclassified areas between lakes made it difficult to pan from one lake to another, and that I needed statistics on individual lakes so the classifications needed to be done on the subsetted images.
For classification, I originally tried a supervised classification on the 1 meter NAIP imagery but with no IR band to detect water, the classifier was confused and made major errors classifying water and wetlands. I then decided to do an unsupervised classification with many classes. I tried different numbers of classes ranging from 100 to 10. In the end it appeared that 60 classes were optimal for the 1m NAIP images. After completing the classification, however, I realized that the water within the lake and surrounding the lake of interest were classified both as water. The water in the lake should not be considered part of the classification so it needed to be classified as a distinct class that could be eliminated in the analysis. To exclude the water I opened the July 2010 1 mNAIP imagery in ArcMap. Then I digitized the exterior boundary of the lake using the polygon drawing tool. I then converted the drawing to a feature (shapefile) and extracted the shapefile from the 1 meter circular subset. I exported the extract as an image and opened it in Erdas. In a separate viewer, I loaded the digitized shapefile and a new AOI layer. I copied the shapefile into the AOI layer and saved the AOI. In order to give the lake a unique spectral signature, I then radiometrically corrected the extracted image by rescaling. I rescaled from digital number value 254 to 254 and rescaled by the just saved AOI. This produces a white water body. To add the newly rescaled water body to the 1m circular mosaic, I used the MosaicPro from 2D feature. For this process I also tried to recode the circular subset directly from an AOI file but without the extracted subset this process was not successful when mosaicked.
After I did this for each of the 2010 1m images, I once again did the 60 class unsupervised classification. Unfortunately, this time the thematic recode did not properly work. A possible source of the failure was a repeated message to close attribute editor prior to saving and before reopening for the recode. This message appeared both on the Saint Paul and Minneapolis campuses. I asked a lab tech to help me understand the error message but he could not understand the problem either. To get around the error message after the classification, I had to force save by closing the classification and then reopen for the recode. After I assigned new values to the classes and clicked apply and then close, when I reopened the image there was typically only the unclassified class and then one or two random old classes like trees 1 or wetland 3. Because this recode was now saved, all of my 60 previous classes were lost and the classification had to be done over again from scratch. By visual interpretation however it appeared that the 60 class unsupervised classification with the excluded water did a good job of distinguishing between the five classes. The main problem with this classification was that approximately two classes for each image were separate classes for the shadows of trees and buildings.
For the 30 m resolution images the 60 class unsupervised classification was not optimal. Due to the low pixel numbers in many of the lakes, when I ran the 60 class unsupervised on the merged image, some lakes would only have 1 pixel classified and would be very difficult to find and identify for the classification. I found that a 20 class classification worked better by increasing the pixel values per class. In the lakes with higher numbers of pixels especially, there were still mixed classes with the 20 class scheme. To reduce some of the confusion and eliminate the lakes from being part of the classification, I needed to digitize the 30m lakes and rescale them as well. Unfortunately, the lakes were less defined in the 30m resolution imagery so digitization was not possible. I tried to mosaic the rescaled 1 meter lakes with the 30 meter data and that did not work because they contained different numbers of layers.
Results and Discussion
Because of all of the issues with the classifications, I was not able to meet any of my goals but I did exert a lot of effort and I learned a lot. I also learned that I have a lot more to learn and do. I need to complete these classifications in order to complete my Masters project so I will continue to work on this project until it is complete.
Because the classifications were never properly recoded, I was not able to perform an accuracy assessment. When I do finish the classification process I will do an accuracy assessment using 50 randomly distributed points in each of the five classes. The fact that I digitized the water features of my lake will impact my accuracy and hopefully improve the accuracy. There may be confusion however because for many of the lakes the boundaries between wetland and lake were quite fuzzy and I used my memory from being on the lakes this summer and my best judgment to digitize the lake boundary while maintaining what I considered to be wetland. My accuracy assessment mechanism will differ from the assessment mechanism used by Yuan et al. For their assessment, they randomly placed approximately 363 polygons containing at least 100 pixels throughout the metro area and then assessed the accuracy of each pixel. This form of assessment is not possible for my areas due to the low pixel values in many of my subsets.