Analysis of Landcover in the Wakarusa River Valley

Analysis of Landcover in the Wakarusa River Valley

Introduction

The Wakarusa River Valley is a sub-basin of the Kansas River Basin. Historically, the floodplain consisted of approximately 18,000 acres of wetlands which provided critical habitat supporting a rich diversity of flora and fauna, in addition to hydrological functions. Over the last century, this vast wetland complex has undergone drastic changes due to habitat reclamation. As areas were settled, wetlands were converted to agriculture resulting in a floodplain consisting of row crops and pasture. With the added pressure of population growth in the form of residential and commercial sprawl, the once self-functioning ecosystem lost the capacity to manage increased water levels and support unique species assemblages. In 1971 the Wakarusa River was impounded by Clinton Dam to prevent seasonal spring flooding which significantly altered the water table and the natural recharging of the wetlands. Today, approximately 600 acres of wetlands remain in the Wakarusa River Valley. This remnant tract is known as the Haskell-Baker Wetlands and it is located south of the city of Lawrence, Kansas.

Vegetation is consistent with poorly drained floodplain habitats. Some typical species associated with hydric soils (Wabash-Kennebec-Reading association) are prairie cord grass, aster, and sedges. The habitat types include wetland prairie, meadows, marshes, open water and mature riparian woodland.
This area is a critical refuge that houses a unique diversity of fish and wildlife Wetlands Photo Tour. It is of special concern because it provides nesting, migratory and wintering habitat for migratory bird species
The U.S. Fish and Wildlife Service estimate that up to 43% of the threatened and endangered species rely on wetlands for survival. The Northern Crawfish Frog, Rana areolata circulosa, is a Kansas species in need of conservation (SINC) that could possibly occur within the wetlands. The key habitat requirements for this secretive species are crawfish and small mammal burrows, groundwater near or at the surface, and pools that persist after the spring rains Center for North American Herpetology

The main objective of the project is to extract and classify a subscene of the Wakarusa River Valley from a Landsat 5 Thematic Mapper dataset in order to analyze land use , both visually and statistically. This project will address data obtained on May 23 , 1994 and eventually datasets from 1987 to 1997 will be incorporated to document landuse change. Consequently, the anlaysis will aid in the identification of threats to existing wetlands and find ways to monitor change to determine strategies for restoration efforts.

Methodology and Analysis

A Landsat 5 TM dataset (LT5027033009718310, path 27 row 33) that covers 32,400 sq km of northeast Kansas was acquired (purple outline) The dataset was already in IDRISI format, so bands 1,2,3,4,5 and 7 were imported. The study site includes the eastern, downstream portion of the Wakarusa basin to the mouth of the Kansas River near Eudora (red outline). Utilizing the WINDOW module the study site was extracted. The subset scene covers 304 sq km (75,084 acres) with a pixel resolution of 28.5 meters. The metadata for each band was analyzed and corrected for haze to minimize atmospheric effects.

Composite images

To enhance the visual display for feature extraction composite images were created with a 2.5% linear stretch. This involves combining three bands into one image by assigning each band to a color, either blue, green or red. This takes advantage of the different portions of the electromagnetic spectrum. The foundation of image processing relies on the fact that each band contributes differently to the discrimination of land features. Three composites images were created with standard TM bands.

Landsat Thematic Mapper Bands adapted from Campbell (1996, p.138)

Band Color Wavelength (µm) Appications
band 1 Blue 0.45-0.52 Separation of soil and vegetation
band 2 Green 0.52-0.60 Reflection of vegetation
band 3 Red 0.63-0.69 Chlorophyll absorption
band 4 Near Infrared 0.76-0.90 Delineation of water boundaries
band 5 Mid Infrared 1.55-1.75 Vegetative moisture
band 6 Far Infrared 10.4-12.5 Hydrothermal mapping
band 7 Thermal 2.08-2.35 Plant heat stress

Natural Color Composite. Combination of bands 1, 2, and 3 (B, G, R). This composite simulates the natural colors of the visible spectrum. Colors appear as they would on a normal color photograph. Active vegetation appears in different shades of green.
Standard False-color Composite. Combination of bands 2, 3, and 4. The active vegetation appears in shades of red, whereas emerging vegetation and fallow areas appear in shades of green
Special False-color Composite. Combination of bands 3,4, and 5. This composite does a good job separating water features from the surrounding landscape. It was able to differentiate the water skiing pool (royal blue rectangular feature south of K-10) from surrounding land features.

Principal Components Analysis

This enhancement method utilizes statistical procedures to reduce redundancy in multispectral image data. In this study 6 bands of true information were selected to create composites. It is inevitable that bands share similar characteristics and essentially describe the same information. The procedure will condense the data into new principal components that can be treated as single bands to be used in composite images

PCA Results

Component 1 Component 2 Component 3 Component 4 Component 5 Component 6
Eigen value 3133.37 560.51157.5423.624.192.11
% Variance 80.73 14.444.060.610.110.05

The eigenvalues express the amount of variation explained by each component. Looking at the table, principal components 1, 2, and 3 contain 99.23% of the variation. Thus, the six TM bands have been compressed to 3 new component images. It is now possible to maximize the full potential of composites. A PC3, PC 1, PC2 composite was created.

Unsupervised and Supervised Classification

The CLUSTER and ISOCLUSTER modules were used to create unsupervised classifications. This classification technique is recommended if the user does not have extensive knowledge of the region. It also provides useful information regarding the distinct clusters by identifying commonly occurring and distinctive reflectance patterns in the image. It assumes these classifications are distinct, so it is up to the analyst to determine the identity of classes by ground truthing techniques. In this study these methods were only used to obtain a general idea regarding the distribution and number of classes. ISOCLUSTER

Supervised classification relies on the analyst to define distinct areas with a unique spectral signature. These areas are referred to as training sites. Each area was digitized and assigned a unique identifier. The supervised classification is based on 14 training sites. These sites were selected by visually inspecting the composites and comparing each site to the high resolution DOQQ. The PCA composite was used as the base image for digitizing. The training sites were lumped into four categories: non-active vegetation /barren , active vegetation, water features, and urban areas

Non-active Vegetation -- in the broad sense, exposed soils to emerging vegetation, color tones ranging from white to green (PCA 3,1,2 composite). Assumed as inactive, these classes lack the pink to red tones of active vegetation (2,3,4 composite)

  • Fallow/barren -large white rectangular features and sand banks along river
  • Veg1 -yellow/white rectangular features
  • Veg2 -lime green rectangular features
  • Veg3 -light/dark green shades
  • Veg4 -dark green

Active Vegetation -- assumed to be active (pink to red tones in 2,3,4 composite)

  • Forest -bright red irregular clusters with texture
  • Agriculture -uniform pink to red rectangular features
  • Grass/pasture -open areas in urban and residential known to be grass
  • Wetland/prairie -brown/pink tones Baker wetlands and ground truth prairies

Water Features

  • Lake -Clinton Lake
  • Rivers -Kansas River

Urban

  • Urban 1 -blue urban cluster
  • Urban 2 -urban structures that appear white (with aid of DOQQ)
  • Urban 3 -light blue residential streets

After the classes were defined spectral signatures were created and classifications were conducted. The statistical techniques used were minimum distance to means classifier (MINDIST) and the maximum likelihood classifier (MAXLIKE). All of the classifications are based on the principal component bands 1, 2, and 3.

Boolean image creation

The final step in the project was the reclassification of a supervised image to create a boolean image depicting urban landuse.

Results and Discussion

The maximum likelihood classification resulted in misclassification of urban and fallow/ barren land cover categories. This is evident along the Kansas River where sandbars and shoreline were confused with white urban structures. It also misclassified pixels within the Baker wetlands. There should not be any urban areas within the four quadrants. A second classification using all of the TM bands resulted in similar misclassifications. It is quite clear that this technique will not be suitable for future change detection analysis.

The minimum distance to means classifier did an excellent job differentiating all fourteen landcover categories. Urban and barren/fallow areas were a problem in both the maximumlikelihood and unsupervised techniques. This classifier was able to differentiate urban and fallow/ veg categories north and south of the river. Only a small percentage of riverbank pixels were classified as urban and the sandbanks were correctly classified as barren. These areas are not critical for the analysis of the wakarusa valley that will be addressed in a separate project. The critical areas are the wetlands, agriculture, and urban areas south of highway K-10.

This is a subset of the remnant patch of wetlands, a DOQQ from October 1991 and the 1994 landsat subset scene. One must keep in mind that restoration efforts began in the early '80s with numerous projects to follow beginning in 1991. Only two small sections of virgin wetland meadow remain.

Looking at these two subsets it is evident, in this case, that the minimum distance to means classifier based on the PCA bands is superior. A classification using the 6 TM bands resulted with mixed pixels. The condensing of data from six bands to three bands was the factor that permitted a fairly accurate classification. The fallow areas west and east of the wetlands have been correctly classified. The forested riparian areas along the Wakarusa, the small pond in the northeast corner, residential streets and neighborhoods verify an accurate classification.

The final classification is based on the maximum distance to means classifier with combined urban classes to demonstrate the importance of selecting homogenous pixels. The first MDM classifier utilized homogenous training sites to identify urban areas. At training site A only blue pixels were classified, whereas the final MDM classifier incorporated a heterogenous residential training site. Light blue streets and pink vegetation (trees and grass) were lumped into one category. This was done to incorporate the active vegetation associated with residential neighborhoods. Older neighborhoods will typically have trees with large canopies which will skew urban pixel classification. C refers to "urban 1" pixels and D refers to "urban 2" pixels. It is evident that the new classification was able to incorporate residential vegetation throughout the entire scene. The only significant misclassifications occurred in the Baker Wetlands quadrant. It was not able to differentiate between the highly active aquatic vegetation in the northeast quadrant and residential vegetation.

This is a boolean image created to mask all classes except the three urban classes. It is based on the first MDM classifier. The red areas represent 32 sq km of residential, not including residential vegetation. Looking at the image it is evident that this figure underestimates total urban landcover. This will be considered during the second phase of the project. The goal is to correctly classify pixels within the Wakarusa river valley south of K-10, in particularly the Haskell-Baker wetlands

Conclusion

By carefully defining training classes and incorporating 3 principal component bands a suitable land cover classification was created. The next phase of the project will use similar techniques to analyze land use change from 1987 to 1997. Image diferencing techniques will be used to estimate suburban sprawl. The Baker Wetlands will also be analyzed by comparing TM datasets from multiple years to document restoration efforts.

References

Baker University Species Lists. Jayhawk Audubon Chapter and Environmental Protection Agency.http://www.skyways.org/orgs/jayhawkaudubon/

Campbell, James., 1996. Introduction to Remote Sensing. Gilford Press, New York, New York.

Center for North American Herpetology.http://www.cnah.org/index.asp

South Lawrence Trafficway K-10 (Kansas Department of Transportation). http://www.southlawrencetrafficway.org/1d2_wetlands.html

This webpage was created by Andy Schmidt to fullfill the requirements for the Spring 2002 Advanced Digital Image Processing (ES 775) at Emporia State University. For questions or comments contact Andy Schmidt at newt70@hotmail.com