Analysis of Water Clarity in Lake Superior Using MODIS Images
Analysis of Water Clarity in Lake Superior using MODIS Images
Yan Wang, Beth Peterson
Forest Resources 3262 / 5262
Remote Sensing of Natural Resources and Environment
12/10/12
Background
Satellite imagery has been used effectively to estimate water clarity levels. Improving the application of satellite data for analyzing water quality is an important issue because it would reduce the cost of gathering on site data in remote locations, and allow for the analysis of water quality from historical images when field data is not available.
Water clarity can be analyzed from satellite by measuring how much light in reflected in different bands. Water reflects a small range of light compared to most surfaces, but good data can still be acquired by comparing the return in different parts of the spectrum. The most reflection occurs near the blue end of the visible spectrum, while less light is reflected near the red portion of the spectrum.
Water clarity is an important measure of water quality. Measures of water clarity are often included in the standards for safe drinking water. There are several ways of measuring water clarity, one method is to measure turbidity. Turbidity can be measured in Nephelometric Turbidity Units (NTU), by using a nephelometer which measures the amount of light reaching a detector after passing through the water. In this way the amount of suspended particles scattering the light can be found. Clear water will have a low NTU, while murky water will have a high NTU.
Many human activities, mining in particular, have impacted the clarity of water in Lake Superior. Average water clarity for Lake Superior is 8m when measured using a secchi disk, which is very good relative to most lakes. Because it is such a large lake, clarity can vary in different places, and clarity may not be as good in all locations. Reduced clarity comes from suspended sediments or algal growth from eutrophication. In Lake Superior eutrophication is very low due to the relatively low amount of organic matter input.
Most of the land cover surrounding Lake Superior is forested. There are logging roads and other logging activities that could pollute the basin. There are several mining operations within the Mesabi Iron Range that could also be contributing contaminants to the lake, as well as other point and non-point sources in local communities.
Objectives
For this project, one objective was to analyzewater clarityin Lake Superior near the Duluth area from June to September in 2011 and 2012. Analysis was done in the summer and early fall because the lake was guaranteed to be free of ice during this period.
By taking data from different times we can evaluate how clarity changes over the season and from year to year. In 2012 there was a large flood in the area that resulted in a lot of water being washed into the lake, carrying with it debris from the surrounding area.
The other objective of this project is to simply evaluate the usefulness of the data and procedure for finding water clarity. We would like to evaluate whether MODIS data, particularly 8 day composite images with a resolution of 500m, can be used to accurately determine water clarity. We will also be evaluating how useful it is to compare the MODIS data to turbidity data collected in the field at streams emptying into Lake Superior.
Methods/Procedure
Data Acquisition
We decided to use MODIS 8 day composite images from the satellite Aqua. MODIS revisits each position once a day, so daily images are available. The 8 day composite images were used because contain the images with the least amount of cloud cover. To study water clarity over time we used the 8 day composite images collected from June to September 2011 and 2012. MODIS offers several different formats for data, we decided to use surface reflection data MYD09 which shows recorded surface reflectance, corrected for atmospheric effects from the satellite Aqua. We used the data product MDY09A1, which contains surface reflectance bands 1–7 at 500m resolution as an 8 day composite.
To establish the relationship between reflectance and turbidity we used turbidity data collected at several streams flowing into Lake Superior by the Lake Superior Streams organization. Turbidity is recorded in NTU. The data was recorded very frequently, approximately every 15 min every day at most sites. Data is available for many years at most sites.
Study Area
Figure 1 The study area was limited to the area of the lake near Duluth. This figure shows where the study area is in relation to the rest of the MODIS image.
Figure 1 shows the study area. The area of interest was limited to the area of the lake near where the field data was collected because we felt the accuracy would likely go down as we moved further from where the reference data was collected. The entire study area was contained in a single MODIS image, so we did not have to stitch together separate images. We clipped the MODIS images to a polygon containing the area of interest. We also used the polygon from the Lake Superior shape file provided by the MNDNR Data Deli to delineate water vs. land more precisely. We decided not to clip the image to this shape file because the collection points of the streams were not included in the lake area and we wanted to keep these since they would be important in understanding the results.
Data Extraction
The MODIS sattilite Aqua has a consistent passover time and passes over our study are at about 18:50 every day. The images that we used are a composite of the pictures taken each day at this time over an 8 day period. Using the latitude and longitude of each collection site we extracted the DN value of the pixel at each location and recorded this for each 8 day period.
For the field data, we extracted the turbidity value recorded at the time closest to the passover time as well as the values recorded immediately before and after. The values were averaged over the 8 day periods represented by each image to get one turbidity value for each site in each image.
Regression analysis
Once processed, a regression analysis was used to find the relationship between the turbidity data collected in the field and the reflectance data from bands 1 and 3 in the MODIS images. Ideally, a separate regression would be established for each image, but because the number of points was so limited it was decided that establishing a regression for each month would be better.
Image classification
The ArcGIS tool Raster Calculator was used to create a new raster image based on applying the regression equation to each pixel of the MODIS images.
Ranges were set manually for each class. The raster image could be classified using a supervised or unsupervised method. A supervised classification based on water quality would be ideal. However, because the results did not match the typical distribution of water clarity values, it was decided that manually defining the classes so that they could be interpreted visually was the best option. The same classification was used on all images for each year to make comparison easier.
Results
We experimented with several different ways of combining the data to produce the regression equation. We found the most reasonable trend lines and best R2 values when we set y as ln(field data) and x as (Band3-Band1)/(Band3+Band1). Figure 2 shows the regression equations for each month.
Figure 2
Figure 2 shows the regression equations to be used as an algorithm to find the turbidity of points from the MODIS images. Equations are unique to each month.
The range of resulting pixel values varied from image to image and month to month, but for the most part was 0-1. We expected clarity to be lowest near the shore and highest in the center of the lake. This relationship was found on some of the images but on some it was reversed. Figure 3 (attached) shows the resulting maps for 2011, and figure 4 (attached) shows the maps for 2012.
Accuracy assessment
Data for the accuracy assessment came from a set of lake water quality measurements taken in 2012. Water clarity was recorded as secchi disk depth measured in the field or as turbidity measured in the lab. Only records including turbidity in NTU were used since the field data used to establish the regression was in NTU, and secchi disk depth cannot be converted to NTU.
The DN value was recorded for each point on each data for which there was turbidity data available. The DN values were then used to predict the turbidity using the appropriate regression equation. Figure 5 shows this being done for July.
Figure 5. DN values from bands 1 and 3 were used to calculate the predicted turbidity at each site. This was recorded along with the field data from each site. Sites are repeated because they were surveyed twice during the month.
To assess the accuracy of the predicted values, the predicted values were graphed against the values collected from the field. The results for each month can be seen in figure 6.
Figure 6 Predicted vs field tested turbidity for each month in 2012.
If the turbidity was predicted correctly the regression line should be y = x since the predicted and field turbidity would be the same. The farther the linear regression equation is from y = x, the less acuratly the turbidity was predicted. Predictions were the least acurate for June, and the most acurate for September.
Discussion
The R2 for the regression equations we created were often very low. The range of results was very unexpected. Most of the results fell within a much smaller range than that of the reference data with outliers that were either much higher than what would be reasonable, or negative. We were unable to establish an accurate regression equation for several reasons.
The turbidity data collected at the various streams was likely flawed. Some of the recorded NTU values were negative, which should not be possible. There were many gaps in the data where NTU was not available. There were points were the data jumped between very high and very low values for no apparent reason.
The field data was not compatible with the MODIS data. The MODIS data used returns an image with a pixel size of 500m. The streams were much smaller than this, most only a few meters wide. Overhanging trees would also obscure the streams in satellite images. Each MODIS pixel contained the reflectance signature of many more surfaces than just the stream, the result being that the relationship between the pixel value and the turbidity was very small.
The number of reference points was too limited. Each image only had 5-7 points where the field turbidity was known. Even if the data was very accurate this would not be enough to establish a very reliable regression. Other studies referenced would not use an image if it had less than 20 points. By combining all the data for each month more points could be used for each regression but doing this introduced many new sources of error.
8 day composite images may contain too much variation. The 8 day composite images were used because they greatly limited the amount of interference by cloud cover. The drawback to this is that it’s impossible to be very precise about determining how the elements of the image relate to specific days.
Because the results were so far from the expected as a result of bad data, were unable to interpret much about the water clarity of the lake. Most of the images show a pattern where the water clarity in the center of the lake is different from the clarity near the shore which is expected. There was much more variation in the images from 2012 than 2011, which may have been due to the extra sediment in the lake from the 2012 flood. No clear pattern was observed for how the turbidity changed through the season.
Conclusions
While other studies have demonstrated the usefulness of MODIS data and other data of similar resolution for assessing water clarity, our study shows that this usefulness is limited by how well it relates to the reference data collected in the field. The large pixel size of MODIS images requires that the reference data also cover a large area.
The applicability of composite images is something worth looking into for further study. 8 day composite images are very nice since they reduce the need to sift through images for cloud free days, but it’s not clear how data taken in the field on specific days should be related to the images. This kind of analysis may work best for finding general patterns in large lakes but may not be sufficiently detailed to pick up on more subtle values in smaller lakes.
Works cited
"Amity Creek."Duluth Streams.Lake Superior Streams.Web. 9 Dec 2012.
<
Brezonik, Patrick L., Lief G. Olemanson, Marvin E. Bauer, and Steve M. Kloiber."Measuring Water Clarity and Quality in."Cura Reporter. Summer. (2007): 3-13. Web. 9 Dec. 2012. <
"Chester Creek."Duluth Streams.Lake Superior Streams.Web. 9 Dec 2012. <
Durskey, Jim, and Beth Kluthe."Source Water Assessment."City of Silver Bay, Minnesota. (2002): n. page. Web. 9 Dec. 2012.
Kloiber, S.M.; Brezonik, P.L.; Olmanson, L.G.; Bauer, M.E. 2002.“A procedure for regional lake water clarity assessment using Landsat multispectral data.” Remote Sens. Environ. 82, 38–47.
Knight, Joseph F.; Voth, Margaret L. 2012."Application of MODIS Imagery for Intra-Annual Water Clarity Assessment of Minnesota Lakes." Remote Sens. 4, no. 7: 2181-2198.
"Kingsbury Creek."Duluth Streams.Lake Superior Streams.Web. 9 Dec 2012.<
"Miller Creek."Duluth Streams.Lake Superior Streams.Web. 9 Dec 2012. <
Olmanson, L.G.; Kloiber, S.M.; Bauer, M.E.; Brezonik, P.L. 2001 “Image Processing Protocol for Regional Assessments of Lake Water Quality; Public Report Series #14”; University of Minnesota: Saint Paul, MN, USA.
Perlman, Howard, and Steve Lawrence."Understanding Turbidity."Chattahoochee Riverway Project.USGS, 11 2002.Web. 9 Dec 2012. <
Petus, Caroline, GuillemChust, et al. "Estimating turbidity and total suspended matter in the Adour River." Continental Shelf Research. 30.5 (2010): 379-392. Web. 9 Dec. 2012. <
"UnderMining Superior: A Report on Mining Activities and Impacts in the Lake Superior Basin."Northwatch. 2001: n. page. Web. 9 Dec. 2012. <
Vermote, E.F., and S.Y. Kotchenova."MOD09 (Surface Reflectance) User’s Guide."MODIS Land Surface Reflectance Science Computing Facility.MODIS Land Surface Reflectance Science Computing Facility, n.d. Web. 9 Dec 2012.
Vermote, E.F., and S.Y. Kotchenova. "MOD09 (Surface Reflectance) User’s Guide." MODIS Land Surface Reflectance Science Computing Facility. MODIS Land Surface Reflectance Science Computing Facility, n.d. Web. 9 Dec 2012.