Land Cover Change Detection in the Twin Cities Metropolitan Area by Watershed, 1984 2009

FR 3262 - Final Project

Land cover change detection in the Twin Cities Metropolitan Area by Watershed, 1984 – 2009

Group members:

Alex Steele

Tobias Fimpel

Objectives

The aim of this project is to identify and quantify the land cover changes that have taken place over the past 25 years within the eight primary watersheds that the Twin Cities metropolitan area is part of. The rate of urban development in each watershed will be of particular interest. Urban development can have severe negative impacts on ecosystems by changing surface characteristics, which in turn affect the local water budget. When croplands or undeveloped areas are transformed into urban or suburban landscapes, the ability of fallen precipitation to permeate the surface and recharge soil moisture oftentimes decreases. Consequently the amount of surface runoff increases. When undeveloped lands are transformed into croplands, changes in the amounts of chemicals and sediments impact surface water conditions. Vegetative and hydrologic systems are thus directly affected by such land cover changes. A wealth of information about land cover conditions can be derived from remotely sensed data.

Study Area and Data

The study area is composed of seven adjoining counties located in south-eastern Minnesota: Anoka County (excluding a small portion in the northern part), Washington County, Hennepin County, Ramsey County, Carver County, Scott County, and Dakota County. In total, the study area covers approximately 3000 square miles and is home to a population of nearly three million people.

Eight primary watersheds are part of the study area. These range from approximately 46 square miles to 1011 square miles in size. To delineate watershed boundaries we obtained the shapefile “Watersheds in the Twin Cities Metropolitan Area - DNR Level 08 Catchments” from the Metropolitan Council.

Land cover information was derived from two Landsat images, one taken on 08/15/1984, and one taken on 07/25/2009. Both are Landsat scenes located at Path 37, Row 29, taken by a Landsat TM sensor. The images are made up of a total ofsix bands with a spatial resolution of 30 meters and one band with a spatial resolution of 120 meters.The spectral resolution is as follows: Band 1 0.45μm-0.52μm, band 2 0.52μm-0.60μm, band 3 0.63μm-0.69μm, band 4 0.76μm-0.90μm, band 5 1.55μm-1.75μm, band 6 10.40μm-12.50μm, band 7 2.08μm-2.35μm. The image dating from 2009 we obtained from the United States Geological Survey’s website, the image dating from 1984 was provided to us by the course instructor and its bands were already stacked.

Procedures

The software used to perform the following procedures was ERDAS Imagine 2010. Before beginning the image classification process, we created a layer stack for the later image and visually inspected the Landsat scenes. Using the “Layer Selection and Stacking” function we stacked the .TIF Image files we obtained from the United States Geological Survey’s website, thereby creating an ERDAS .img file. After this step, the image dating from 8/15/1984 as well as the image dating from 07/25/2009 were processed very similarly. Close visual inspection of the two scenes showed no apparent radiometric or geometric errors, so that we proceeded to the image classification process.

For the purpose of this study, we desired a classification scheme with the following four informational classes: water, undeveloped land, agriculture, and urban land. Water was to include wetlands and areas of open surface waters such as lakes and rivers. Undeveloped land was to include forests, grasslands, and brush. The agricultural class was to include areas used for growing crops. Areas occupied by streets and buildings made up the urban class.

After several failed attempts to devise an unsupervised pixel-based classification scheme that produces these four informational classes we decided to perform a supervised pixel-based classification on both images instead.

Using EradsImagine’s signature editor we created a signature file for each image. Because areas of barren fields showed fundamentally different radiometric properties than fields where crops were being grown, we created two spectral classes representing agriculture which we later merged. By creating two spectral classes, one for cultivated fields and one for barren fields, we avoided having one single spectral class representing agriculture with a very large variance which in turn may have led to the misclassification of grasslands as agricultural areas.

In order to minimize negative effects of differing atmospheric conditions possibly present in different parts of each Landsat scene, we delineated all training sites (AOIs) within the extent of our study area in the northeastern portion of the two scenes. For the image dating from 1984 we delineated20 AOIs for water, 27 AOIs for undeveloped lands, 15 AOIs for barren agricultural land, 30 AOIs for cultivated agricultural lands, and 27 AOIs for urban areas. We then merged signatures representing the same class. The resulting merged signature classes contained the following number of pixel: 38934 pixel for water, 14173 pixel for undeveloped lands, 2830 pixel for barren agricultural lands, 13349 pixel for cultivated agricultural lands, and 11651 pixel for urban areas. Next we ran the classification using the Maximum Likelihood classification algorithm and, following this, merged the two resulting informational classes representing agriculture using the thematic recode function. The image dating from 2009 was processed similarly.

Next we clipped the images to the study area’s extent via ErdasImagine’s “Subset and Chip” operation. To do this we converted theshapefile containing the watershed boundaries into an Erdas Imagine Area of Interest file (.aoi) and used this file to define the extent of the “Subset and Chip” operation’s output.

Before proceeding with the change detection step, we assessed the accuracy of each classified image subset. For the subset image dating from 1984 we useda stratified random sample of 120 points, weighted by area, with a minimum of 15 points per class. As reference data we used the same Landsat image that we used for the classifications. The accuracy assessment of the image dating from 1984 produced the following results (fig 1):

Fig. 1: Accuracy Totals for Subset Image 1984

The overall classification accuracy was 71.67%, the overall Kappa statistic 0.6014.

For the accuracy assessment of the subset image dating from 2009 our sample was made up of 20 randomly chosen points for each of the four classes. Also here, the same Landsat image that we used for the classification was used as reference data. The accuracy assessment of the image dating from 2009 produced the following results (fig 2):

Fig. 2: Accuracy Totals for Subset Image 2009

The overall classification accuracy was 83.75%, the overall Kappa statistic 0.7842.

When both subset images were overlaid in one Erdas Imagine viewer, it became apparent that although the pixel boundaries of both images were geometrically identical, the Landsat scene dating from 2009 extended slightly further towards northwest than the one dating from 1984. This difference was noticeable because both Landsat scenes did not include a small part in the northern portion of Anoka County, which was part of the Area of Interest file delineating the watershed boundaries. To remedy this we used the “Mask” utility to crop the extent of the 2009 subset image to the very same extent as that of the 1985 subset image.

To detect changes in the land cover between the two dates we used the “Matrix Union” utility to create a file containing a matrix of all “from-to” class change possibilities. This change-map provided land cover change information for the entire study area summarized by zone.

Overlaying the .aoi file containing watershed boundaries that was created in an earlier step allowed us to generate eight image files, one for each primary watershed, via the “Subset and Chip” utility. The attribute tables of these eight image files contained “from-to” land cover change information pertaining to the respective individual watersheds. The values contained in these attribute tables we then used in ArcGIS, Adobe Illustrator, and Microsoft Office Excel to generate maps, tables, and graphs presented in the following section.

Results

We were able to accomplish our objective, which was to identify and quantify the land cover changes that have taken place over the past 25 years within the eight primary watersheds of the Twin Cities metropolitan area. Tables about detailed landcover change information for each watershed can be found in the appendices. We found that the overall patternof land cover change has been similar for all eight primary watersheds (fig 3).

Fig. 3: Land Cover Change by Primary Watershed in Hectare, 1984-2009

Between 1984 and 2009 agricultural lands have largely been either converted to urban areas or have remained agricultural, all watersheds show very little conversion of agricultural lands into undeveloped areas. Undeveloped lands show a tendency of having been changed to agricultural and urban areas, with the exception of two watersheds, namely St. Croix-Stillwater and Rum River, where more than half of the undeveloped areas from 1984 remained undeveloped in 2009. Urban areas have mostly remained Urban, but a significant amount of change from Urban to Agriculture was detected, too. Areas occupied by Water in 1984 largely remained Water in 2009 but some change from Waterto Undeveloped and from Water to Urban was detected, too.

As illustrated by the following maps, the two watersheds “North Fork-Crow River” and “Mississippi River Rush-Vermillion” saw the highest rates of urbanization at the expense of bothundeveloped lands and agricultural lands. About 16% of the area contained within each of these watersheds has undergone a change from Agriculture to Urban, and about 20% of the area contained within each of these two watersheds has undergone a change from Undeveloped to Urban (fig 4, fig 5).

Fig. 4: Agriculture to Urban by Watershed, 1984 - 2009

Fig. 5: Undeveloped to Urban by Watershed, 1984 - 2009

Since the overall accuracies of both classified subset images were moderately high (73% and 84%) our results are most likely reflective of the general trends and their magnitudes within individual watersheds. The values could be used to inform decisions of public planners and natural resource managers but are not sufficiently accurate to be used as data for further scientific studies, such as for example statistical analyses of land cover change effects on water quality.

Discussion

To construe more definite conclusions from the results of this project, a more comprehensive accuracy assessment would be necessary. The accuracy assessment that was carried out has several limitations. Firstly, the number of points sampled was insufficient. Secondly, reference data that is better interpretable than Landsat scenes should be used instead. Thirdly, our samples used were biased in that not every pixel had the same chance of being selected.

The performance of the classification process could be greatly enhanced by the inclusion of additional data. Multi seasonal Landsat images would help to differentiate undeveloped lands from croplands that are tilled during spring or fall. Furthermore, it would accentuate differences in the radiometric responses of urban areas and agricultural lands. Including multispectral data from early spring or late fall would most likely have also been beneficial for discerning undeveloped lands from wetland areas. Considering the very simple classification system made up of only four classes, the acquisition of costly high-resolutionimagery would certainly not be justified.

Appendices

Fig. 6: Land Cover Change 1984-2009, Entire Study Area

Fig. 7: Land cover Change1984-2009, Cannon River Watershed

Fig. 8: Land cover Change1984-2009, MN River-Shakopee Watershed

Fig. 9: Land cover Change1984-2009, Mississippi River Rush-Vermillion

Watershed

Fig. 10: Land cover Change1984-2009, Mississippi River Twin Cities

Watershed

Fig. 11: Land cover Change1984-2009, Rum River Watershed

Fig. 12: Land cover Change1984-2009, St. Croix-Stillwater Watershed

Fig. 13: Land cover Change1984-2009, North Fork-Crow River Watershed

Fig. 14: Land cover Change1984-2009, South Fork-Crow River Watershed

1