Minneapolis Environmental Study:
Mississippi River Water Quality and
Land Cover Changes
Team members: Thomas Bales & Andrea Claassen
Course: FR5226
Date: 12/14/2011
Instructor: Joe Knight
- Introduction.
The 72-mile Mississippi River Corridor in the Twin Cities was designated as a State Critical Area in 1976, because of its unique natural, scenic, cultural, historical and recreational resources – shared community assets that improve Minnesota's economy and quality of life.
The Mississippi River Corridor Critical Area is a resource of regional, statewide and national significance that requires special management to retain its health and vitality. Designated a part of our National Park System in 1988, the Mississippi National River and Recreation Area (MNRRA) relies upon the State Critical Area framework to ensure protection of park resources. The Mississippi River is a drinking water source for more than 20 million Americans. Unfortunately, every mile of the river in the MRCCA fails to meet State standards for water quality. New standards are needed to reduce runoff pollution to the river.
– Friends of the Mississippi, 2011
Problem Statement
The goal of this project is to use remote sensing and GIS techniques to assess trends in environmental characteristics of the Mississippi River, thereby giving various organizations the ability to make policy and/or dedicate resources to specific areas to promote local river health.
The Mississippi River is an important economic, cultural, and ecological resource. In recent decades, riverfront development policies have shiftedas communities increasingly view the Mississippi River as a community asset. In the Twin Cities, the Mississippi River is an important economic resource for the cities, is a source of municipal drinking water, offers recreational opportunities for residents, and provides important habitat for wildlife.
Although federal, state, and local government agencies have passed amendments to improve water quality, water quality is still impaired in the Mississippi River in the Twin Cities metro area and fails to meet state and federal water quality standards. Local government agencies, non-profits, and communities are working to enhance water quality and watershed protection through education, monitoring and stewardship programs through the Twin Cities metro area. Additionally, these groups have become engaged in various river protection activities such as native plant restoration.
In light of the economic, cultural, and ecological importance of the Mississippi River it is important to document current environmental characteristics, as well as to identify and assess changes over time. This project will use remote sensing to examine and assess changes in general river and bank health of the Mississippi River in the Twin Cities region, specifically focusing on a few key variables related to river health such as water quality, sedimentation patterns, and vegetation patterns. This information will better enable various local agencies and organizations to make policy decisions and dedicate resources to specific areas to promote local river health.
- Methodology.
Preprocessing phase
To conduct a water clarity/quality analysis, imagery from The National Agriculture ImageryProgram (NAIP) was obtained. This is aerial imagery that is capturedevery year since 2003 during the agricultural growing seasons in the continental United States. This imagery is high resolution (between 1 and 2 meters) and contains the red, green, and blue color bands. The exception is 2008 which contains an additional infrared band.
In order to do a time comparison, images from 2004 and 2009 were downloaded to cover a five year period. These photographs were then clipped by using polygons representing the river, obtained from MetroGIS. Another polygon, which included the river channel and a 0.25 mi buffer,was created to represent the surrounding riverbank and floodplain characteristics.In addition, the shapefile representing the different mosaic tiles from the NAIP imagery had to be used to analyze where overlaps occurred to differentiate between the different photograph exposure levelsin the tiled mosaic (Appendix II, Figure 1).
In order to pre-test differences in water clarity, a principal component analysis was performed using2011 Terralook imagery. The principal component analysis was done in ArcInfo 10 which had settings modified to isolate and highlight urban, vegetation, and water areas of the image. This process transformed the multiband raster from the input multivariate attribute,and a new multivariate attribute whose axes are rotated with respect to the original was created. The new attributes in the new space are uncorrelated, and data in a principal component analysis compressed the data by eliminating redundancy. The result was that different water qualities were able to be differentiated over large areas (Appendix II, Figure 2).
Once the process was complete and the remaining raster bands were dropped, it was clear that differences in water using RGB imagery could be extracted. For example, figure 2 (in Appendix II) clearly shows red as urban, green as vegetation and mixed classes, but more importantly that two types of water are present in the image. Purple represents the less clear or cloudy water and blue represents clear water.
After conducting this preliminary analysis, three areas along the Mississippi River were determined to be of concern for water quality assessment. They were the sections of the Mississippi River in North Minneapolis, in downtown Minneapolis, and belowthe confluence of the Minnesota River in South Minneapolis.For repeatability, aniterative self-organizing data analysis (ISO) clustered unsupervised classification was conducted on the series of input raster bands using Esri’s maximum likelihood classification tools. Three areas were determined to have different water signatures. However, for the purposes of this study only the northern (North/Northeast Minneapolis) and southern (confluence of the Minnesota and Mississippi River) areas were targeted for further analysis (Appendix II, Figure 3). Lastly, the shapefiles created in ArcInfo were imported to ERDAS to create Areas of Interest (AOIs).
ERDAS: Sites for further analyses
For the following analyses we focused on two different sites on the Mississippi River within the Twin Cities metropolitan area. The north site contains a stretch of the Mississippi River situated in North/Northeast Minneapolis and the northern suburb of Columbia Heights. The north site is primarily industrial use, with some residential areas and parkland. The south site lies at the confluence of the Mississippi and Minnesota Rivers in south Minneapolis and is primarily residential use with some parkland. The following analyses (classifications, change detection, and accuracy assessment) were all performed using ERDAS (ERDAS Imagine 2010 and 2011 software).
Changes in river clarity:
Unsupervised classifications of the river (clipped to the river channel) were conducted for both the north and south sites for both 2004 and 2009. The unsupervised classifications were performed several times for the 2009 south site imagery using different numbers of clusters to test for the most appropriate number of clusters to use for further analyses. Using ten clusters was the most appropriate for distinguishing various water clarity levels in the river and for identifying areas to be excluded from water clarity analyses (e.g. bridges over the river, emergent sandbars, shadows). A convergence threshold of 25 iterations and the principal axis intializing option were set for conducting the unsupervised classifications in ERDAS. The defaults were used for all other parameters. Areas of non-interest (bridges, sandbars, and shadows) included those with the two lowest pixel values (i.e. pixels pertaining to the two darkest cluster levels) and the highest pixel value (i.e. pixels pertaining to the lightest cluster value). After these areas of non-interest were excluded, seven clusters pertaining to different levels of water clarity remained for inclusion in water clarity analyses.
Change detections were conducted to compare land use changes between years for both the north and south river sites. Although the classified images with seven water clarity levels were useful for visual interpretation (Appendix II, Figures 4 and 6), this was deemed to be too many water classes to use for conducting change detections. Therefore, thematic recode was used to combine clusters into a simpler subset of three classes, combining the two darkest clusters, the three medium clusters, and the two lightest clusters (Appendix II, Figures 5 and 7). These recoded clusters represented low, medium, and high levels of river turbidity. Change detections were conducted for the recoded classifications containing the three water clarity classes.
Changes in floodplain and channel land cover:
Supervised classifications of the buffered area (river area plus 0.25 mi buffer) of the southern site (confluence of Mississippi and Minnesota Rivers) for 2004 and 2009 NAIP imagery were conducted. We identified the following classes corresponding to land cover types of interest: clear water, silty water, river sandbar, trees/forest, grass, wetland vegetation, bare soil (dirt), and road/impervious surface. We established at least 5 training areas for each class. We used the maximum likelihood “parameter rule” for assigning pixels to classes based on the training data that were input into ERDAS.
A change detection was performed to assess changes in land cover in the river buffer area (i.e. river floodplain and channel areas) between 2004 and 2009.
Accuracy assessment:
An accuracy assessment was performed for the supervised classifications for the south site for both 2004 and 2009. The same data (NAIP 2004 and 2009) that were used for classifying the images were also used as reference data for the accuracy assessment because other high resolution reference data were not available. A stratified random sampling distribution was used for each accuracy assessment, with 100 sample points used and a minimum of 10 points per class. Although the number of sample points used was lower than the recommended 50 points per class, time limitations prevented the use of a larger sample size.
- Results
Based on results of the change detection of the unsupervised classifications, water clarity did not appear to change in the north area of the Mississippi River from 2004 to 2009. Although clear water declined by 5.5%, silty water also declined by an equivalent amount (5.4%), and medium clear water increased by 9.6% (Appendix II, Table 1). These changes likely reflect difficulties in separating the classes rather than any true differences in water clarity in the north area of the Mississippi River. In the south area (confluence of the Mississippi and Minnesota Rivers), however, water quality declined during the same period, as evidenced by a negative change (-3.9%) in the amount of clear water, and a positive change (+3.8%) in the amount of silty water (Appendix II, Table 3). In the south area, water clarity changes reflect sedimentation patterns of the Minnesota River, a highly polluted and turbid river (Appendix II, Figures 6 and 7).
Although results of the change detections of the unsupervised classifications of the river channel provide estimates of net change to river clarity, inspections of the classified images themselves also reveal changes in sedimentation patterns (Appendix II, Figures 4-8). Of note, the classified images of the south area (Appendix II, Figures 6 and 7) reveal that in 2009 sedimentation was lower on the Minnesota River and higher on the Mississippi River below the confluence of the Minnesota River than in 2004.
The results of the change detection of the buffered south area (Mississippi and Minnesota River channels, plus a 0.25m buffer to include the floodplain), seem to somewhat contradict the results of the unsupervised classification of the river channel. The change detection of the buffered south area shows clear water increasing by 16.7% and silty water increasing by a smaller 4.2% (Appendix II, Table 5). These differences may be a result of the unsupervised classification having three water clarity classes, whereas the supervised classification of the buffered south area only used two water clarity classes. The use of two rather than three water clarity classes for the supervised classification was necessary due to difficulty in visually selecting training areas for use in the supervised classification. The discrepancy in change detection results for water clarity from the two types of classifications performed may also reflect differences in classification accuracy of the two methods.
Overall accuracy assessment of the classifications of the buffered south area were 65.4% and 66.0% for 2004 and 2009, respectively (Appendix II, Tables 7 and 8). In 2004, bare soil (dirt) had the lowest user’s and producer’s accuracy, trees/forest had the highest user’s accuracy, and wetland vegetation had the highest producer’s accuracy (Appendix II, Table 7). In 2009, river sandbar’s had the lowest user’s accuracy and one of the lowest producer’s accuracy (tied with grass), and trees/forest and wetland vegetation both had 100% user’s and producer’s accuracy (Appendix II, Table 8). The results of the accuracy assessment suggest that for this analysis the darker land cover types had higher accuracy than the lighter colored land cover types. Of the river channel landcover types of interest, clear water was most confused with trees/forest, silty water was most confused with clear water, and river sandbars were most confused with roads/impervious surfaces (Appendix II, Tables 7 and 8).
- Discussion.
The results of this study are useful to provide some initial insight into environmental characteristics of the Mississippi and Minnesota Rivers, including how river clarity and floodplain landcover may change over time. These preliminary results indicate that water clarity in the area of the confluence of the Mississippi and Minnesota Rivers in south Minneapolis has declined somewhat from 2004 to 2009, and that the pattern of sedimentation has changed, with lower sedimentation in the Minnesota River above the confluence and greater sedimentation in the Mississippi River below the confluence with the Minnesota River. However, this study should be viewed as phase 1 of a more thorough investigation into changes in river landcover classes over time, with an emphasis in future project phases on improving the accuracy of land cover classifications used to detect environmental changes. Further investigation should be conducted over a longer time period, further trial supervised classifications should be performed to better identify appropriate and relevant land cover classes for analysis, more extensive accuracy assessments with a greater sampling frequency should be conducted, and relevant supplemental data (e.g. rainfall patterns, agricultural runoff data) should be included in the analysis to better understand apparent changes in river clarity and sedimentation patterns.
Data Limitations
The image classifications do not account for the possibility of different river water levels, river discharge rates, or water velocities between years. Also, slightly different time of year for the 2004 (August) and 2009 (September) images may affect the color intensities between the year; however, the image months are as close as possible.
Additional data limitations are the lack of a longer term data set for the analysis. The 5-year period from 2004 to 2009 was likely too short to discern large changes over time among land cover classes. Also, the difference in spatial resolution between 2004 (2m resolution) and 2009 (1m resolution) may have affected the accuracy of the river classifications, thus affecting the results of the change detection for water clarity between years. Furthermore, shadows and other areas of non-interest (river bridges, emergent river sandbars) in the images used for water clarity analyses may have affected the accuracy of land cover classifications, as well as the interpretation of the imagery and results of change detections. However, we tried to account for this by excluding areas of non-interest (shadows, bridges, sandbars) from the imagery used in the water clarity analyses, first by clipping the areas of the river channel to exclude as many of these areas as possible, and second by coding these features as areas on non-interest in the classifications that were performed. Lastly, unaccounted for radiometric image differences between images may have affected the results and interpretation.
- References.
APPENDIX 1 (Materials and Software)
NAIP Imagery (2004 and 2009)
Terralook imagery (ASTER 2011)
ERDAS Imagine
Esri ArcInfo
APPENDIX II (Tables and Figures)
Tables
Table 1: Water clarity changes for Mississippi River in North Minneapolis, 2004 to 2009.